graph theory recent research papers

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 08 September 2022

Explainable artificial intelligence through graph theory by generalized social network analysis-based classifier

• Serkan Ucer 1 ,
• Tansel Ozyer 2 &
• Reda Alhajj 3 , 4 , 5  

Scientific Reports volume  12 , Article number:  15210 ( 2022 ) Cite this article

7335 Accesses

5 Citations

11 Altmetric

Metrics details

• Computational biology and bioinformatics
• Mathematics and computing

We propose a new type of supervised visual machine learning classifier, GSNAc, based on graph theory and social network analysis techniques. In a previous study, we employed social network analysis techniques and introduced a novel classification model (called Social Network Analysis-based Classifier—SNAc) which efficiently works with time-series numerical datasets. In this study, we have extended SNAc to work with any type of tabular data by showing its classification efficiency on a broader collection of datasets that may contain numerical and categorical features. This version of GSNAc simply works by transforming traditional tabular data into a network where samples of the tabular dataset are represented as nodes and similarities between the samples are reflected as edges connecting the corresponding nodes. The raw network graph is further simplified and enriched by its edge space to extract a visualizable ‘graph classifier model—GCM’. The concept of the GSNAc classification model relies on the study of node similarities over network graphs. In the prediction step, the GSNAc model maps test nodes into GCM, and evaluates their average similarity to classes by employing vectorial and topological metrics. The novel side of this research lies in transforming multidimensional data into a 2D visualizable domain. This is realized by converting a conventional dataset into a network of ‘samples’ and predicting classes after a careful and detailed network analysis. We exhibit the classification performance of GSNAc as an effective classifier by comparing it with several well-established machine learning classifiers using some popular benchmark datasets. GSNAc has demonstrated superior or comparable performance compared to other classifiers. Additionally, it introduces a visually comprehensible process for the benefit of end-users. As a result, the spin-off contribution of GSNAc lies in the interpretability of the prediction task since the process is human-comprehensible; and it is highly visual.

Similar content being viewed by others

Song lyrics have become simpler and more repetitive over the last five decades

Emilia Parada-Cabaleiro, Maximilian Mayerl, … Eva Zangerle

Highly accurate protein structure prediction with AlphaFold

John Jumper, Richard Evans, … Demis Hassabis

Reconstructing the evolution history of networked complex systems

Junya Wang, Yi-Jiao Zhang, … Yanqing Hu

Introduction

Machine learning (ML) has dominated our daily life with many real-world applications, such as healthcare decision support systems, search engine recommendation systems, and autonomous driving, among others. One of the most common tasks of ML is ‘classification’ which broadly involves two steps, training and testing. While training inspires a model from one part of the data, testing uses the remaining data to check the accuracy of the model which reflects its capability in correctly realizing the classes of new samples. The data types processed with ML are not limited to tabular form, but may exist in a wide range of forms, including multimedia. Still, tabular data may be considered the most common form of data types 1 , dominating most of the applications of ML. The main struggle of the current ML paradigm is explainability of existing ML model to naive users. While the performance of the classification accuracy increases as new models emerge, surprisingly, classification decisions become complex to be justified by humans 2 . This ‘explainability’ phenomenon limits the usage of ML models in critical real-world applications (e.g., law or traffic management) since the context of a decision is hard to be justified and explained to the end-users. Our proposed social network analysis-based visual classifier, GSNAc, aims to solve this problematic aspect of ML by introducing visual cues, both at the model and at the prediction steps.

Social networks form a specific type of network graphs used for depicting real-life paradigms in the form of ‘actors’ and their ‘interactions’ with each other in a closed ecosystem 3 . In its most primitive form, a social network consists of a finite number of nodes (called actors) connected to each other via weighted edges (a weight of one for all edges is also possible) which represent a sort of ‘social interaction’ based on some kind of social quality, e.g., ‘frequency of their coffee meetings’. A pervasive form of a social network is ‘the friendship network’, e.g., Facebook.

Social Network Analysis (SNA) is a study conducted to understand various aspects of a network in general as well as its nodes individually and in groups. For example, in a friendship network, it is possible to find the most influential node (i.e., a person) in the analyzed domain. SNA techniques are widely used in various applications from link prediction for recommending useful connections in a ‘friendship network’ to detecting fraud in a ‘financial transactions network’, among others. SNA techniques mainly include, but are not limited to, detecting node or actor centrality, community (structure) detection, edge capacity, flow, and representative selection.

An overview of our past works

For a social network, it is not always necessary for the interaction to be linked to a human activity. Indeed, various types of entities other than humans may also socialize, or the relationship between them may be interpreted as a social activity analogous to human activities. Accordingly, interactions can be defined for a broader spectrum of entities such as among genomes, mice varieties, animals, insects, fruit types, etc. In this respect, increasing interest in the SNA domain has allowed many diverse real-life phenomena to be studied. In our former works 3 , 4 , we tried to address two machine learning tasks using graph theory fundamentals: “feature reduction via SNA” and “the SNAc—classification via SNA of time-genomic datasets”, respectively. For our work described in 5 , we employed SNA techniques and studied specific data from the life sciences domain, which in general do not represent any real ‘social’ value. The techniques themselves proved to be useful in the classification of a tabular dataset. We represented genomic datasets as social network graphs, where genomes are represented as ‘actors’ (nodes) and ‘their vectorial similarity’ to other genomes together with their interaction levels form ‘edges’.

Current model and the key contributions

Following the success of the proof-of-concept of this approach, we now continue searching for an extended generalized framework that is again based on graph theory and SNA. The target involves addressing the classification of tabular datasets containing either or both numerical and categorical features. Thus, the proposed GSNAc Model is a new type of visually explainable machine learning classifier that works on tabular datasets.

Our approach (see Fig.  1 ) moderately increases the explainability of the whole classification process, since it presents a traceable visual output for each predicted sample. Besides, the ‘graph classifier model—GCM’, which sits at the heart of the classification process, is again in a social network form that makes it visualizable. Those aspects merely address the current problem posed by the most advanced current set of state-of-the-art classifiers, which is the explainability of the classification task (i.e., XAI—explainability of artificial intelligence applications). To summarize, the key aspects and contributions of the GSNAc model could be enumerated as follows:

Clarity of the method, explainable process, deterministic classifier,

Graphical interface, explainability, visual engagement,

Versatility (i.e., works well for the classification of both numerical and categorical features),

Superior or on par performance (in terms of prediction accuracy) with the well-known state-of-the-art tabular data classifiers.

The brief pipeline of the GSNAc model.

Structure of the research presented

We will present our current work by introducing the problem and the techniques that will be used to address the problem posed in “ Related work and key concepts ” section. We will then discuss details of the GSNAc Model starting with an overview in “ GSNAc: a visual supervised learner ” section. We will further detail the model in “ The data: converting tabular data to raw network graph ”, “ The model: graph enrichment and extraction of the graph classifier model ”, “ The prediction process: using visual graph classifier model (GCM) ” sections. Experiments and results of the comparison with conventional machine learning classifiers will be present in “ Experiments and results ” section. We dedicated “ Explainability of the GSNAc model ” section to elaborate on the explainability aspect of the proposed GSNAc Model. Discussion and further study are covered in “ Discussion and further work ” section.

In Appendix I , we provide the code, the detailed parameters of the classifiers & the datasets, and the runtime parameters of the GSNAc Model, including the raw performance results, and the URL to the directory of outputs GSNAc produced throughout the experiments, all for the sake of the reproducibility of the conducted research. Finally, in Appendix II , we provide the user interface and the output of GSNAc.

Related work and key concepts

Graphs, social networks, and sna.

Social networks 6 form a subtype of network graphs that originated from graph theory. Graph theory is briefly the study of network graphs, which are mathematical representations used to model pairwise relations between objects. A graph is normally characterized by a set of nodes and edges which connect nodes. Typically, edges may have a default weight of 1; alternatively, varying weights may be assigned to various edges to reflect the value of the relationship between the connected nodes. Indeed, many real-world problems and systems can be reduced and represented as a network graph. For instance, the traffic system connecting roads in an area may be reorganized as a graph where intersections and roundabouts act as nodes and edges represent roads connecting them. This way, a navigation map application may identify a possible shortest path between two locations by analyzing the graph to determine the most appropriate route based on the combination of the costs of road segments connecting some consecutive junctions.

A network graph is simply the combination of nodes and edges connecting nodes. Each edge has a source node, a target node, and a weight value showing its effect based on the criteria applied to construct the network. Assume an edge e1 between nodes a and b has weight 4, and another edge e2 between a and c has weight 3, then we conclude that a is more similar to b than c. In some cases, edges of a network graph can be directed, i.e., the direction of connection between the source node and the target node matters, e.g., in a directed graph, two edges e1 from a to b and e3 from b to a are two different edges. If the direction doesn’t matter in the domain of study, it is safer to use undirected networks.

Having originated from the social sciences domain, Social Network Analysis (SNA) is the sum of measures inspired from graph theory to analyze the relationship among social entities. Social Network analysis encompasses many measures that can be broadly grouped as:

Centrality studies for nodes

Edge analysis

Node distance-similarity methods

Community structure studies

SNA measures used in this study are node similarity analysis, edge analysis and centrality studies.

Machine learning: the classification task

It is very common to come across the results of scientific experiments involving measurements of various aspects (in other words, features) of objects in a specific domain, and at the end objects are often categorized into a limited number of cases (called classes).

In the machine learning domain, the term ‘classification’ implies the task of first learning the characteristics of a group of objects from specific classes and then predicting the correct class for some previously unseen objects forming the test data.

Any learner model follows almost the same black-box model shown in Fig.  2 . According to this generalization, it starts with the ‘Data’ box, where the dataset under study is examined in terms of quality, completeness, and complementarity of its feature space and sample data points. Tasks such as decomposition, transformation, elimination, and improvement of the feature space are some of the common methods used in this step.

Illustrative box diagram of a general machine learning classification model.

The aim of the ‘classifier model’ box is to find and describe the connection between the data and its class, by employing mathematical and statistical techniques. Finally, in the ‘Prediction’ box, the proposed model implements its technique to process data with the model presented to produce appropriate predictions. Crucially, in this step ‘benchmarking’ ensures that we assess the merit of the classifier. Here, internal and external benchmarking can be employed, i.e., ‘statistical hypothesis testing techniques’ as the internal evaluation, and comparing the performance of the proposed classifier with the state-of-the-art established models as the external evaluation.

There is no silver bullet technique for classification tasks 7 . For instance, some classification techniques may perform efficiently on biological data while may fail to produce meaningful predictions on financial data. State-of-the-art techniques for tabular data classification consider decision tree-based classifiers. Among those, notably XGBoost 8 depends on tree learning, and dominates most of the recent 9 data science competitions. Another recent and popular research is deep learning, where its main contribution is on feature space reorganization and selection. Despite the fact that deep learning could be used for classification, their effective usage is still mainly on text, image, and video data types 10 .

Machine learning and network graphs

Even though SNA techniques are widely used to solve various problems from engineering to media; machine learning classification using network graphs is a domain that has not received enough attention from the research community. In our former work described in 5 , we employed various SNA techniques to convert time-sequential genomic data into a social network and developed a classification model to predict classes. The method proved to be useful when compared to conventional algorithms, but was limited to being applied only on time-sequential data. Within this work, we aim to generalize our classification model (abbreviated as GSNAc) to an extent that the method can be applicable to numerical/categorical datasets which include class information.

The machine learning applications for the social network domain are generally centered around two topics 11 : (i) the similarity between two graphs (or subgraph matching), and (ii) the similarity of the nodes/edges in a graph. The first one aims to find a motif between possibly two different-sized graphs (in terms of nodes and edges) to deduct a result, e.g., a deduction that both graphs have similar network dynamics. Here, a recent work in this area is worth mentioning 12 . It employed graph theory techniques for text classification. In 12 , NLP researchers first convert sentences into adjacency matrices based on the co-occurrence of words, and then to graphs. The process then efficiently predicts word similarities by employing graph similarity techniques. At this point, we state that the proposed GSNAc model has no relation with this ‘inter-graph similarity’ aspect of the mentioned research domain.

On the other hand, finding possible similarities ‘within’ a graph, in terms of the similarity between nodes or edges, is a problem partially addressed by GSNAc. The ‘node classification’ problem refers to the process of propagating labels to the unlabelled nodes on a partially labeled graph by evaluating the similarity of nodes 13 , 14 . A recent research in this area (refer to 15 ) handles node classification problems by introducing graph neural networks over graphs (particularly Graph Convolutional Networks—GCNs). This approach basically learns hidden layer representations that encode both the local graph structure and features of nodes. It has been proved useful in predicting labels on some types of networks including citation networks where it predicts the type of the document. Within this context, the node classification problem partially matches the definition of the problem tackled by GSNAc; since (after converting data to a graph and after its enrichment) we also predict the class (i.e., label) of an unseen node by evaluation against labels of the training nodes. However, in comparison to the node classification problem setting, there are major challenges that GSNAc needs to deal with. First, in the node classification problem, the complete (weighted) graph (including all nodes—i.e., training and test parts) is already in place initially, making it easier to assign labels based on their topological properties based on the complete connection structure within all nodes (labeled or unlabelled). Second, node classification can predict labels iteratively by depending on known positions, possibly starting with a frequently labeled portion of the graph. We believe that these dynamics differentiate the two problems well enough. Here it is worth mentioning that the GSNAc model is actually an end-to-end machine learning classifier on its own.

To sum up, we would like to present a recent trend on converting tabular data into networks. This approach becomes pretty useful when the aim is to visualize data to get an intuition-based idea. A recent research described in 16 efficiently converts tabular data into images (rather than networks as GSNAc does), and predicts drug responses of gene expression profiles using deep learning techniques on the generated images. While this approach does not coincide with GSNAc, it shows the popularization of expressing tabular data as 2D visible graphic items, a reasonable step towards explainability of tabular learning.

GSNAc: a visual supervised learner

In this section, following the formal problem description, we will present an overview of the GSNAc model both in diagram and pseudocode formats. A detailed description of the methodology and parameter selection strategies will be covered in the following three subsections.

Problem description and naming conventions

The formal problem description of GSNAc is exactly the same as the problem definition of machine learning for the categorical classification task: given a multi-class tabular dataset with numerical and categorical features, the process involves learning from the seen data and predicting the class of unseen data. Multi-class classification refers to predicting one or more classes for each sample. Imbalanced classification refers to a classification task where the distribution of samples across the classes is skewed. In other words, most of the existing objects represent specific classes. Fortunately, GSNAc is not merely a bi-class classifier. It does not aims to work only on balanced data, it also works on imbalanced data which makes it useful on a broader spectrum of problems. A final note on the use of GSNAc is that it currently does not support regression prediction which is predicting continuous numerical values.

Before we delve into detailed descriptions of the GSNAc methodology, we would like to present (see Fig.  3 ) the ‘naming’ convention for this research. Throughout this work, for the dataset part, our preference is to call individual observations (of an experiment) as the ‘samples’, and their defining characteristics (i.e., predictors) are called ‘features’ of the samples. Categories of the results of the experiment are called ‘classes’.

(i) Tabular data converted to adjacency (similarity) matrix (ii), and graph (iii) constructed and enriched with node degrees, edge weights, and colored by communities (classes). Layout graph intuitively shows the importance of some nodes (i.e., H. Sapiens and T-Rex have the highest number of weighted connections, therefore, larger in size) and visually depicts how similar nodes are (e.g., Shark has the smallest similarity to other nodes so it is distant from the core).

After converting the dataset to a network graph, basically dataset ‘samples’ become ‘nodes’ in the graph, and dataset ‘features’ are blended into ‘edges’ of the graph. A node may have a ‘neighbor’ if there is an edge between them. Finally, the classes are represented in the network graph as the ‘communities’.

The pipeline of the GSNAc methodology is diagrammatically described in Fig.  4 . It starts with the preprocessing of raw tabular data. After data is separated into training and test parts, a raw graph is generated using only the training part of the data. This raw graph is decorated and further processed to achieve a leaner version called graph classifier model (GCM) with fewer edges and repositioned nodes.

General overview of the GSNAc model illustrated on Iris dataset.

This GCM was then employed for predicting the class of test samples using two different similarity scores. Next, we present the master pseudocode summarizing the GSNAc methodology.

In Fig.  1 we described the GSNAc Models’ process pipeline: this type of framing comes useful when proposing a novel model. In this sense, we organized the following three sections focused on the boxes of the model, namely ‘Data: Graph generation’, ‘Model: GCM Generation’ and ‘Class Prediction’.

The data: converting tabular data to raw network graph

In this subsection, we present details on how we process the dataset, turn it into a network graph and finally how we produce, and process features that belong to the graph. Topics to be covered are:

splitting the data,

preprocessing,

feature importance and selection,

computation of similarity between samples, and

generating of the raw graph.

Splitting the tabular data

After preprocessing the data, the next step is to split the dataset into training and test samples for validation purposes. We selected cross-validation (CV) as the validation method since it is the de facto standard in ML research. For CV, the full dataset is split into k folds; and the classifier model is trained using data from (k-1) folds then tested on the remaining k’th fold. Eventually, after k iterations,, average performance scores (like F1 measure or ROC) of all folds are used to benchmark the classifier model.

A crucial step of CV is selecting the right proportion between the training and test subsamples, i.e., number of folds. Determining the most appropriate number of folds k for a given dataset is still an open research question 17 , besides de facto standard for selecting k is accumulated around k = 2, k = 5, or k = 10. To address the selection of the right fold size, we have identified two priorities:

Priority 1—Class Balance: We need to consider every split of the dataset needs to be class-balanced. Since the number of class types has a restrictive effect on selecting enough similar samples, detecting the effective number of folds depends heavily on this parameter. As a result, whenever we deal with a problem which has low represented class(es), we selected k = 2.

Priority 2—High Representation: In our model, briefly, we build a network from the training subsamples. Efficient network analysis depends on the size (i.e., number of nodes) of the network. Thus, maximize training subsamples with enough representatives from each class (diversity) is our priority as much as we can when splitting the dataset. This way we can have more nodes. In brief, whenever we do not cross priority 1, we selected k = 5.

By balancing these two priorities, we select efficient CV fold size by evaluating the characteristics of each datasets in terms of their sample size and the number of different classes. The selected fold value for each dataset will be specified in the “ Experiments and results ” section. To fulfill the class balancing priority, we employed stratified sampling. In this model, each CV fold contains approximately the same percentage of samples of each target class as the complete set.

Feature space organization and preprocessing

Preprocessing starts with the handling of missing data. For this part, we preferred to omit all samples which have one or more missing feature(s). By doing this, we have focused merely on developing the model, skipping trivial concerns.

As stated earlier, GSNAc can work on datasets that may have both numerical and categorical values. To ensure proper processing of those data types, as a first step, we separate numerical and categorical features 18 . First, in order to process them mathematically, categorical (string) features are transformed into unique integers for each unique category by a technique called labelization. It is worth noting that, against the general approach, we do not use the one-hot-encoding technique for transforming categorical features, which is the method of creating dummy binary-valued features. Labelization does not generate extra features, whereas one-hot-encoding extend the number of features.

For the numerical part, as a very important stage of preprocessing, scaling 19 of the features follows. Scaling is beneficial since the features may have a very different range and this might affect scale-dependent processes like distance computation. We have two generally accepted scaling techniques which are normalization and standardization. Normalization transforms features linearly into a closed range like [0, 1], which does not affect the variation of values among features. On the other hand, standardization transforms the feature space into a distribution of values that are centered around the mean with a unit standard deviation. This way, the mean of the attribute becomes zero and the resultant distribution has a unit standard deviation. Since GSNAc is heavily dependent on vectorial distances, we do not prefer to lose the structure of the variation within features and this way our choice for scaling the features becomes normalization. Here, it is worth mentioning that all the preprocessing is applied on the training part of the data and transformed on the test data, ensuring no data leakage occurs.

Feature importance and selection

Feature Importance (FI) broadly refers to the scoring of features based on their usefulness in prediction. It is obvious that in any problem some features might be more definitive in terms of their predictive capability of the class. Moreover, a combination of some features may have a higher effect than others in total than the sum of their capacity in this sense. FI models, in general, address this type of concern. Indeed, almost all ML classification algorithms use a FI algorithm under the hood; since this is required for the proper weighting of features before feeding data into the model. It is part of any ML classifier and GSNAc. As a scale-sensitive model, vectorial similarity needs to benefit much from more distinctive features.

For computing feature importance, we preferred to use an off-the-shelf algorithm, which is a supervised ‘k-best feature selection’ 18 method. The K-best feature selection algorithm simply ranks all features by evaluating features’ ANOVA analysis against class labels. ANOVA F-value analyzes the variance between each feature and its respective class and computes F-value which is the ratio of the variation between sample means, over the variation within the samples. This way, it assigns F values as features’ importance. Our general strategy is to keep all features for all the datasets, with an exception for genomic datasets, that contain thousands of features, we practiced omitting. For this reason, instead of selecting some features, we prefer to keep all and use the importance learned at this step as the weight vector in similarity calculation.

Computation of similarity between samples and construction of the raw graph

In this step, we generate an undirected network graph G, its nodes will be the samples and its edges will be constructed using the distance metrics 20 between feature values of the samples. Distances will be converted to similarity scores to generate an adjacency matrix from the raw graph. As a crucial note, we state that since we aim to predict test samples by using G, in each batch, we only process the training samples.

In our study for constructing a graph from a dataset we defined edge weights as the inverse of the Euclidean distance between the sample vectors. Simply, Euclidean distance (also known as L2-norm) gives the unitless straight line (shortest) distance between two vectors in space. In formal terms, for f-dimensional vectors u and v , Euclidean distance is defined as:

A slightly modified use of the Euclidean distance is introducing the weights for dimensions. Recall from the discussion of the feature importance in the former sections, some features may carry more information than others. So, we addressed this factor by computing ‘a weighted’ form of L2 norm based on distance which is presented as:

where w is the n-dimensional feature importance vector and i iterates over numerical dimensions.

The use of the Euclidean distance is not proper for the categorical variables, i.e. it is ambiguous and not easy to find how much a canary’s habitat ‘sky’ is distant from a sharks’ habitat ‘sea’. Accordingly, whenever the data contains categorical features, we have changed the distance metric accordingly to L0 norm. L0 norm is 0 if categories are the same; it is 1 whenever the categories are different, i.e., between the ‘sky’ and the ‘sea’ L0 norm is 1, which is the maximum value. Following the discussion of weights for features, the L0 norm is also computed in a weighted form as \({dist\_L0}_{w}\left(u,v\right)=\sum_{f}{w}_{j}(({u}_{j}\ne {v}_{j})\to 1)\) , where j iterates over categorical dimensions.

After computing the weighted pairwise distance between all the training samples, we combine numerical and categorical parts as: \({{dist}_{w}\left(u,v\right)}^{2}={{dist\_L2}_{w}\left(u,v\right)}^{2}+ {{dist\_L0}_{w}\left(u,v\right)}^{2}\) . With pairwise distances for each pair of samples, we get a n x n square and symmetric distance matrix D, where n is the number of training samples. In matrix D, each element shows the distance between corresponding vectors.

We aim to get a weighted network, where edge weights represent the ‘closeness’ of its connected nodes. We need to first convert distance scores to similarity scores. We simply convert distances to similarities by subtracting the maximum distance on distances’ series from each element.

Finally, after removing self-loops (i.e. setting diagonal elements of A to zero), we use adjacency matrix A to generate an undirected network graph G. In this step, we delete the lower triangular part (which is symmetric to the upper triangular part) to avoid redundancy. Note that, in transition from the adjacency matrix to a graph, the existence of a (positive) similarity score between two samples u and v creates an edge between them, and of course, the similarity score will serve as the ‘vectorial weight’ of this particular edge in graph G.

The raw graph generated in this step is a ‘complete graph’: that is, all nodes are connected to all other nodes via an edge having some weight. Complete graphs are very complex and sometimes impossible to analyze. For instance, it is impossible to produce some SNA metrics such as betweenness centrality in this kind of a graph.

The model: graph enrichment and extraction of the graph classifier model

In this section, we first compute basic SNA metrics over the network graph, then decorate the raw graph with certain properties of the tabular data, and finally process edge analysis of the raw graph. To reduce the complexity and denseness of the graph, we will trim unimportant edges by considering their structure to get a subtle form of the graph which will serve as our visual classifier model.

To increase the understandability of the procedures presented, we visualized the whole process in Fig.  5 , which belongs to a biological dataset. This dataset inspired from C. Elegans Connectome ( https://commons.wikimedia.org/wiki/File:Caenorhabditis_elegans_hermaphrodite_adult-en.svg ) incorporates 279 samples over 12 features (2 categorical, 10 numerical). It has 10 (imbalanced) classes. C. Elegans is indeed a simple worm, and its whole neuron system has been extracted, filed, and widely researched in recent years 21 , 22 , 23 . This worm has exactly 302 neurons (having unique names and positions) connected to 158 muscles and organs on its body. Out of those 302 neurons, following the Connectome research 24 , non-pharynx neurons and neurons with no synaptic connections are omitted, leaving 279 somatic neurons out of the initial 302. This dataset is useful in showing the capabilities of the proposed model. We choose the ‘AY Ganglion’ feature as the class to be predicted. In this specific case, we intend to find a predictive relationship between the neurons and their connections to specific regions throughout the worm’s body.

From raw data to GCM demonstrated on C. elegans Dataset.

Graph decoration and preprocessing

Node naming.

In general, the dataset associated with classification has dedicated sample names, e.g., names of the specific neurons in Connectome dataset or CustomerID’s in Caravan Insurance dataset 25 . Usually, this potentially beneficial information is ignored during the classification. In contrast to this, for the sake of explainability, we save and assign sample names throughout our raw graph as node names.

Community partition assignment

Community detection in a network graph refers to finding partitions of the graphs (subgraphs of connected nodes) displaying similar characteristics. It is beneficial since it reveals hidden relations among nodes in a network. Inevitably, we need to incorporate community detection within the raw graph. While many algorithms have been developed to detect communities, e.g., 26 , 27 , 28 , we do skip employing them intentionally since our network is already partitioned based on the given set of classes. That is, since the raw graph is only composed of training samples, we already do know the classes they have and the communities they belong to. In this fashion, we assign classes as communities in the raw graph.

Centrality SNA metrics

Within a graph, finding how central a node is (compared to other nodes topologically) in terms of the capacity of its entropy is an important branch of SNA metrics as mentioned in “ Related work and key concepts ” section. Among other centrality metrics 3 , the most primitive is ‘degree centrality’ which refers to the number of edges connected to a given node. ‘Weighted degree’ is the improved form of degree centrality, which also considers edge weights when aggregating node connections. Despite not using centrality metrics in the GSNAc model in a predictive capacity, we still find them useful in the model to enrich the displays produced. For this purpose, we use ‘weighted degree centrality’ for scaling the size of the nodes. The latter is a visually engaging concept and gives comparative clues about the entropy capacity of a node within a graph.

Edge pruning

We already highlighted the need for removing self-loops in the former section to simplify the complexity of the raw graph. But, still, this graph is a complete graph, i.e., every node is connected to every other node leading to a graph density of 100%. Now we need to identify and further remove unimportant edges of this complete graph to ensure that our model includes only relevant and important edges; this way, its entropy increases.

Some research 29 , 30 already addressed the raffination of the so-called ‘hairball graph’ which is overly dominated by unimportant edges. We prefer to use a custom raffination at this step, keeping only at most k strongest edges for each node and dropping the remaining edges. This way, by keeping the k value small, we can achieve a raffinated, less dense version of the raw graph. Also, by only keeping the most information-bearing edges, entropy of the graph will increase.

Instead of setting a constant k value for each dataset, we have set an upper limit, and dynamically identified the value of k by analyzing the frequency of the least represented class within the population.

Edge fortification

So far, we have only described how the ‘class’ information of tabular data is only used in graph community assignments. GSNAc might incorporate class information in the model generation step, since it is a supervised model. We prefer to use this very important class information as a secondary step in a reward/loss scheme for edges. This time, we reconsider weights of the edges by analyzing classes of their endpoints. The weight of an edge is improved by a constant factor when it connects two nodes from the same class. Likewise, the same factor is used to decrease the weight of an edge when its endpoints are from different classes. After these steps, the graph may become disconnected since some nodes can lose all of their edges and become isolated, or some group of nodes can be only connected to themselves, creating an isolated connected component.

The prediction process: using visual graph classifier model (GCM)

As described in the previous sections, we have carefully constructed a vectorial similarity-based network graph, pruned its edges by their importance, and finally generated a leaner graph which will act as our classifier model named GCM. At this point, we will feed test nodes one-by-one into GCM and conduct the prediction by analyzing most similar nodes. For this similarity analysis, we have employed two complementary methods: vectorial and topological.

Adding test node to GCM

Vectorial similarities of test node to gcm nodes.

Vectorial similarity is the weighted (L2 and/or L0) distance-based similarity score already computed during the network generation phase. To calculate this vectorial similarity score for each test data to be predicted, upon normalization of test nodes vector, L2 norm-based distance for the numerical part, and L0 norm-based distance for categorical part of the test data are calculated using the same weights that were learned from the training data at step FI. These distance scores are calculated pairwise between the test node and all nodes of GCM. Following the conversion from distances to similarity scores, the test node becomes connected to all nodes of GCM via different edge weights.

Topological similarities of test node to GCM nodes

As we already have a topological structure in the form of a graph, we can also analyze the neighborhood structure of nodes. A topological similarity, in its simple form, can be calculated by comparing the neighborhood structure of two nodes: if they have the same neighborhood with similar weights, they can be considered topologically similar. Pairwise cosine similarity 31 has been used to compute the topological similarity between the test node and nodes of GCM; this measure is useful when one needs to compare two bags of words (with frequencies). Indeed, it is widely used in natural language processing for comparing text similarity.

Our usage of cosine similarity is to define a bag of words as neighborhood lists of nodes and frequencies as the weights of edges (i.e., vectorial similarity) connecting them. As can be seen in the formula, it is again a pairwise calculation as the vector product of the test node (test) and a node from GCM (v) divided by the product of L2 norms of the same vectors. Note that, the value k in the formula refers to the number of neighbors of v; it is expected to be smaller than the number of neighbors of the test node which is already connected to all nodes of GCM.

To conclude, in topological similarity, we do not only consider the direct neighbors of a node from GCM. We also compare two nodes’ neighborhood structure, i.e., how their indirect neighbors match.

Prediction strategy: assigning class to the test node

Topological similarity is somewhat dependent on vectorial similarity since the former uses edge weights (i.e., vectorial similarity) as input. However, topological similarity does not produce the same graph as vectorial similarity, and GCM built by topological similarity gives some clue about nodes’ position within the graph. In this respect, it can be used as a secondary similarity metric, complementing vectorial similarity. We have decided to blend these two similarity metrics to achieve a higher prediction performance and it turned out that using topological similarities as secondary ‘sage’ (whenever decision margin is small with vectorial similarity) has been proven beneficial for prediction as described next.

For actual prediction, we primarily use vectorial similarity and keep the maximum number of edges for a given test node from each represented class in GCM based on their weights. Our prediction is then to group and aggregate those edges class-wise to get an average similarity of the test node to each class, respectively (see Fig.  6 a). To assign a class to a test node, we only consider the top two high-ranked classes and compare their aggregated similarity scores.

( a ) Prediction Phase 1 sample diagram on C. Elegans dataset. İllustrating the prediction attempt of the test node ADFL (not displayed in the diagram). Raw list of neighbors of ADFL listed, sorted by vectorial similarity, grouped by class, and finally similarities are aggregated by class. A prediction is not made since the top two highest averaged classes have very close scores (under margin). ( b ) Prediction Phase 2 sample diagram on C. Elegans dataset. Illustrating the prediction of the test node ADFL (not displayed in the diagram). This time, the prediction is made based on topological similarity in the same fashion with the first round. This time, without seeking a distinctive margin, the top (highest topological similarity average) class is assigned as the prediction.

When the top most classes differ by a magnitude, we assign the top-ranked class to the test node. However, in case the vectorial similarity-based model fails to find two distinctively separated classes (i.e., setting a margin threshold of 1% means the difference of the scores of the top two classes must be higher than 1%) we leave the prediction completely to the secondary sage, i.e., the topological similarity (see Fig.  6 b). This time, in the same fashion, we aggregate the average similarities of the classes over topological similarities, and (without further check) we assign the top-ranked class to the test node.

Experiments and results

The datasets.

GSNAc was initially designed to work efficiently with balanced and bi-class datasets (i.e., binary classification). However, during the development, we discovered that it can effectively classifies some multiclass and imbalanced datasets, as well. As a proof of concept for GSNAc, we carefully selected as many real-world diverse datasets from various domains with various feature and class structures. We also considered including synthetic datasets as they give the chance to compare similar work by other researchers. Finally, we considered some other custom datasets compiled for this research. Table 1 summarizes the datasets used in the experiments.

In the preprocessing steps, we standardized the datasets before feeding them to the ML classifiers since some classifiers (especially those based on scale sensitive distance metrics like SVM and kNN) heavily depend on the data to be scaled. To select the standardization over other competing techniques, normalization is also needed especially for SVM RBF classifier (GSNAc’s main competitors are reported in Table 3 ) since they assume that all the features are centered around zero and variance of the data is of the same order 32 .

Experimental setup

We implemented the GSNAc Model using Python. We selected the platform of GSNAc to be compatible with python’s sklearn library 33 since it is the de facto standard platform in the ML domain. Also, it allows parameter optimization by grid or random search methods, making it useful to find the best performer set of parameters for a given task. Within this context, we did not practice parameter search or optimization for this research in order to present a fair comparison with other ML classifiers. We also used other python libraries such as networkx, bokeh, pandas, and NumPy.

Classifiers used for benchmarking

To be comprehensive and fair in the testing, we decided to include as diverse ML models as possible in terms of their approach for the classification. All the classifiers run with their default parameter set (except for ANN where the maximum number of iterations was upgraded from 200 to 2000 since it does not usually converge with 200 iterations), and likewise, we used default parameters for GSNAc, i.e., we did not search for the best parameters. We used scikit-learn 33 implementation of the utilized classifiers except xgboost, where we used the native xgboost python library 8 . We present in Table 2 the list of classifiers and their parameters used for the comparison.

Performance comparison of GSNAc with traditional tabular learners

In order to test classification efficiency of the GSNAc on the given datasets, we have investigated the comparison of SNAc with traditional machine learning classification algorithms. We selected ten traditional machine learning classification methods to compare their performance with the GSNAc Model over the twenty datasets listed in Table 1 .

Machine learning classification algorithms usually fall into 3 categories, namely statistical, function-based and tree-based (hierarchical) models. We selected at least one algorithm from each category in order to present a balanced comparison.

The reported results, as summarized in Table 3 , show that GSNAc beats or on par with other classifiers on 10 of the datasets. Most notable success of these is on the PBC dataset 34 , where the problem is to predict one of 4 heavily imbalanced classes. On the remaining 10 other datasets, where GSNAc doesn’t score at top, we have noted comparable performance with other classifiers.

Detailed performance results of GSNAc in comparison with other classifiers are presented in Appendix I .

Explainability of the GSNAc model

Considering the rise of deep learners, ML models are increasingly getting complex in terms of their prediction model generation. The so-called black-box classifier models are not comprehendible for end-users on how the actual prediction is decided. For instance, a person may be interested in learning the reason for declining his/her loan application which was evaluated by an AI system.

Explainable AI (XAI) is a recent trend in machine learning research 2 , 35 which aims to identify how predictions are made in an explainable manner (see Fig.  7 ). Explainability and/ or interpretability is essential for end-users to effectively trust, and manage artificial intelligence applications 36 .

Explainable AI approach versus todays’ classifier models in a nutshell.

There are various (recent) research efforts conducted to address the explainability of the already existing machine learning classifier models. In 37 , researchers efficiently extended the use of a convolutional neural network classifier by developing an explainability module. With this extra module, they displayed comprehensionable prediction results of microseismic wave-form signal data while keeping excellent classification performance. Similarly, in 38 , researchers put significant effort to translate deep neural networks-based prediction results over traffic analysis data into visually comprehensible map data; thus providing an explainable (learned) trajectory segmentation to the end user of the deep learner model. Though relevant, we argue that these efforts do not actually coincide with the work developed in this research since GSNAc’s procedures and predictions are already designed with explainability in mind, and hence need no further translation procedure for explainability purposes.

SNA concentrates on generating, and visualizing interactions and information flows among network actors. The power of social networks is stemming mostly due to their capacity in revealing and visualizing interactions in a connected structure. It is no doubt that one can comprehend data more efficiently in a visual way, and network graphs allow us to better understand how nodes are similar to each other, which nodes bear most information, and most importantly, how information flows within a network. By converting and visualizing tabular data, we can reveal and learn several aspects, such as: which samples can distribute information to the largest audience, who is connected to the most influential nodes, and who creates fragility to graph, i.e., the single point of failure by being a single connector between two separate communities. As powered up by the idea of network graphs, GSNAc surely presents some advantages of social networks to its users as briefly summarized in the following three stages:

The first advantage is the visual classifier model (i.e., GCM) which makes sense of the data by keeping sample names as node names and keeping class information as a colored community. This is illustrated in Figs.  3 , 4 , and 5 .

The second advantage towards explainability lies in the actual visual prediction step (see Fig.  8 ): end-users can receive visual clues on how the test node is assigned into one class but not others by analyzing the structure of GCM nodes and weights of the edges.

The final advantage is based on the comparison of prediction results of GSNAc across different classifiers (see Fig.  9 ). GSNAc produces a set of graphs and displays the comparison on the overall prediction process. This way, an end-user can perceive how classifiers (in contrast to GSNAc) predicted a certain sample, whether true or false.

Sample visualization of the prediction step. Test node ADAR from the C. elegans dataset has been classified as CL-E (orange).

Overall classification result on C. Elegans dataset. Blue color implies that the specific sample has been predicted true by the respective classifier (such as AdaBoost); red color indicates false prediction. Boxes in black indicate cases where the same sample has been predicted true (as in blue colors) by XGBoost and AdaBoost classifiers, but false (as in red colors) by GSNAc, and boxes in orange indicate opposite cases (i.e., false by XGBoost and AdaBoost while true by GSNAc).

Discussion and further work

We have proposed GSNAc as a novel graph-based machine learning classifier which works on both numerical and categorical tabular datasets. It outperforms the previous version named SNAc both in performance and capability of handling a wide range of datasets. The main contribution of GSNAc is its capability of visualizing the decision-making process behind the model, while maintaining a fine prediction performance when compared to other classifiers.

GSNAc uses easy-to-grasp similarity-based metrics in graph generation and prediction steps; so might be a choice of some group of end-users who would like to diagrammatically perceive the whole classification process. Since GSNAc also provides prediction margins to end-user, one of its other advantages might be seen as determining and transferring low-margined classification decisions to a human expert for further investigation.

We demonstrated that GSNAc outperforms or is on par with state-of-the-art classifiers across different domains in terms of prediction performance. However, the main drawback of GSNAc is the duration of its run, since its core data structure is a ‘graph’ which has a processing complexity of O(n 2 ). Except for scalability, we are not aware of any shortcomings of this Model.

Lastly, we believe in the future this work can be effectively extended into node classification problems.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Abbreviations

Explainable artificial intelligence

Generalized social network analysis-based classifier

Graph classifier model

Social network

Social network analysis

Social network analysis-based classifier

Machine learning

Natural language processing

Chui, M. C. M. et al. Notes from the AI Frontier: Insights from Hundreds of Use Cases (McKinsey Global Institute, 2018).

Google Scholar  

Adadi, A. & Berrada, M. Peeking inside the black-box: A survey on explainable artificial intelligence (XAI). IEEE Access 6 , 52138–52160 (2018).

Article   Google Scholar  

Alhajj, R. & Rokne, J. (eds) Encyclopedia of Social Network Analysis and Mining (Springer New York, 2018).

Özyer, T., Ucer, S. & Iyidogan, T. Employing social network analysis for disease biomarker detection. Int. J. Data Min. Bioinforma. 12 (3), 343 (2015).

Üçer, S., Koçak, Y., Ozyer, T. & Alhajj, R. Social network Analysis-based classifier (SNAc): A case study on time course gene expression data. Comput. Methods Programs Biomed. 150 , 73–84 (2017).

Tabassum, S., Pereira, F. S. F., Fernandes, S. & Gama, J. Social network analysis: An overview. WIREs Data Min. Knowl. Discov. 8 (5), e1256 (2018).

Bishop, C. M. Pattern Recognition and Machine Learning (Springer, Preface p.viii, 2006).

MATH   Google Scholar  

Chen, T. & Guestrin, C. XGBoost: a scalable tree boosting system. In Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. , 785–794 (2016).

Bansal, S. Data Science Trends on Kaggle !! (Kaggle, 2022).

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521 (7553), 436–444 (2015).

Article   ADS   CAS   Google Scholar  

Kumar, R., Novak, J. & Tomkins, A. Structure and evolution of online social networks. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining—KDD ’06, Philadelphia, PA, USA, 611 (2006) (Accessed 26 Feb 2022).

Shanavas, N., Wang, H., Lin, Z. & Hawe, G. Knowledge-driven graph similarity for text classification. Int. J. Mach. Learn. Cybern. 12 (4), 1067–1081 (2021).

Zhu, X., Ghahramani, Z. & Lafferty, J. Semi-supervised learning using Gaussian fields and harmonic functions. In Proceedings of the Twentieth International Conference on International Conference on Machine Learning , 912–919, Washington, DC, USA (2003).

Belkin, M., Niyogi, P. & Sindhwani, V. Manifold regularization: A geometric framework for learning from labeled and unlabeled examples. J. Mach. Learn. Res. 7 , 2399–2434 (2006).

MathSciNet   MATH   Google Scholar  

Kipf, T. N. & Welling, M. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR) (2017).

Zhu, Y. et al. Converting tabular data into images for deep learning with convolutional neural networks. Sci. Rep. 11 (1), 11325 (2021).

Kuhn, M. & Johnson, K. Applied Predictive Modeling , 1st ed. 2013, Corr. 2nd printing 2018 edition. (Springer, 2013).

Kuhn, M. & Johnson, K. Feature Engineering and Selection: A Practical Approach for Predictive Models (CRC Press, Taylor & Francis Group, 2020).

Bhandari, A. Feature scaling|standardization vs normalization. Analytics Vidhya, (2020).

Deza, M. M. & Deza, E. Encyclopedia of Distances (Springer, 2016).

Book   Google Scholar  

Cook, S. J. et al. Whole-animal connectomes of both Caenorhabditis elegans sexes. Nature 571 (7763), 63–71 (2019).

Emmons, S. W. The beginning of connectomics: A commentary on White et al. (1986) ‘The structure of the nervous system of the nematode Caenorhabditis elegans ’. Philos. Trans. R. Soc. B Biol. Sci. 370 (1666), 20140309 (2015).

Badhwar, R. & Bagler, G. Control of neuronal network in Caenorhabditis elegans . PLoS ONE 10 (9), e0139204 (2015).

Varshney, L. R., Chen, B. L., Paniagua, E., Hall, D. H. & Chklovskii, D. B. Structural properties of the Caenorhabditis elegans neuronal network. PLoS Comput. Biol. 7 (2), e1001066 (2011).

The Insurance Company Benchmark (COIL 2000). http://kdd.ics.uci.edu/databases/tic/tic.data.html (Accessed 30 Dec 2021).

Alamsyah, A. et al. Community detection methods in social network analysis. Adv. Sci. Lett. 20 (1), 250–253 (2014).

Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. Fast unfolding of communities in large networks. J. Stat. Mech. Theory Exp. 2008 (10), P10008 (2008).

Clauset, A., Newman, M. E. J. & Moore, C. Finding community structure in very large networks. Phys. Rev. E 70 (6), 066111 (2004).

Article   ADS   Google Scholar  

Dianati, N. Unwinding the hairball graph: Pruning algorithms for weighted complex networks. Phys. Rev. E. 93 (1), 012304 (2016).

Article   ADS   MathSciNet   Google Scholar  

Edge, D., Larson, J., Mobius, M. & White, C. Trimming the hairball: Edge cutting strategies for making dense graphs usable. In 2018 IEEE International Conference on Big Data (Big Data) , (2018).

Han, J., Kamber, M. & Pei, J. Data Mining: Concepts and Techniques (2011).

“6.3. Preprocessing data,” scikit-learn. http://scikit-learn.org/stable/modules/preprocessing.html , (2021).

Pedregosa, F. et al. Scikit-learn: Machine learning in python. J. Mach. Learn. Res. 12 , 2825–2830 (2011).

Fleming, T. R. & Harrington, D. P. Counting processes and survival analysis. (Wiley-Interscience, 2005). (Accessed 13 Jan 2022).

Das, A. & Rad, P. Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey . ArXiv: 200611371 Cs (2020).

Gunning, D. et al. XAI—Explainable artificial intelligence. Sci. Robot. 4 , eaay7120 (2019).

Bi, X. et al. Explainable time–frequency convolutional neural network for microseismic waveform classification. Inf. Sci. 546 , 883–896. https://doi.org/10.1016/j.ins.2020.08.109 (2021).

Article   MathSciNet   MATH   Google Scholar  

Bi, X. et al. An uncertainty-based neural network for explainable trajectory segmentation. ACM Trans. Intell. Syst. Technol. 13 (1), 1–18. https://doi.org/10.1145/3467978 (2022).

Download references

Author information

Authors and affiliations.

The Scientific and Technological Research Council of Turkey, TUBITAK, Ankara, Turkey

Serkan Ucer

Department of Computer Engineering, Ankara Medipol University, Ankara, Turkey

Tansel Ozyer

Department of Computer Science, University of Calgary, Alberta, Canada

Reda Alhajj

Department of Computer Engineering, Istanbul Medipol University, Istanbul, Turkey

Department of Heath Informatics, University of Southern Denmark, Odense, Denmark

You can also search for this author in PubMed   Google Scholar

Contributions

All authors (S.U., T.O. and R.A.) developed the methodology, S.U. wrote the programs, run the experiments, analyzed the results and wrote the manuscript. All authors reviewed and approved the manuscript.

Corresponding author

Correspondence to Reda Alhajj .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Supplementary information., rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Ucer, S., Ozyer, T. & Alhajj, R. Explainable artificial intelligence through graph theory by generalized social network analysis-based classifier. Sci Rep 12 , 15210 (2022). https://doi.org/10.1038/s41598-022-19419-7

Download citation

Received : 13 March 2022

Accepted : 29 August 2022

Published : 08 September 2022

DOI : https://doi.org/10.1038/s41598-022-19419-7

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Recent trends in knowledge graphs: theory and practice

• Methodologies and Application
• Published: 16 April 2021
• Volume 25 , pages 8337–8355, ( 2021 )

Cite this article

• Sanju Tiwari 1 ,
• Fatima N. Al-Aswadi 2 &
• Devottam Gaurav 3  

3778 Accesses

54 Citations

6 Altmetric

Explore all metrics

With the extensive growth of data that has been joined with the thriving development of the Internet in this century, finding or getting valuable information and knowledge from these huge noisy data became harder. The Concept of Knowledge Graph (KG) is one of the concepts that has come into the public view as a result of this development. In addition, with that thriving development especially in the last two decades, the need to process and extract valuable information in a more efficient way is increased. KG presents a common framework for knowledge representation, based on the analysis and extraction of entities and relationships. Techniques for KG construction can extract information from either structured, unstructured or even semi-structured data sources, and finally organize the information into knowledge, represented in a graph. This paper presents a characterization of different types of KGs along with their construction approaches. It reviews the existing academia, industry and expert KG systems and discusses in detail about the features of it. A systematic review methodology has been followed to conduct the review. Several databases (Scopus, GS, WoS) and journals (SWJ, Applied Ontology, JWS) are analysed to collect the relevant study and filtered by using inclusion and exclusion criteria. This review includes the state-of-the-art, literature review, characterization of KGs, and the knowledge extraction techniques of KGs. In addition, this paper overviews the current KG applications, problems, and challenges as well as discuss the perspective of future research. The main aim of this paper is to analyse all existing KGs with their features, techniques, applications, problems, and challenges. To the best of our knowledge, such a characterization table among these most commonly used KGs has not been presented earlier.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Similar content being viewed by others

Knowledge Graphs: Opportunities and Challenges

Ciyuan Peng, Feng Xia, … Francesco Osborne

Artificial intelligence to automate the systematic review of scientific literature

José de la Torre-López, Aurora Ramírez & José Raúl Romero

BEAR: Revolutionizing Service Domain Knowledge Graph Construction with LLM

https://www.blog.google/products/search/introducing-knowledge-graph-things-not/ .

https://developers.google.com/freebase .

https://babelnet.org/ .

https://wiki.dbpedia.org/develop/datasets .

https://wordnet.princeton.edu/ .

http://www.globalwordnet.org/AWN/ .

http://ling.uni-konstanz.de/pages/home/mousser/files/Arabicverbnet.php .

https://queue.acm.org/detail.cfm?id=3332266 .

https://www.poolparty.biz/what-is-a-knowledge-graph .

https://www.slideshare.net/phaase/getting-started-with-knowledge-graphs .

Abouenour L, Nasri M, Bouzoubaa K, Kabbaj A, Rosso P (2014) Construction of an ontology for intelligent Arabic QA systems leveraging the conceptual graphs representation. J Intell Fuzzy Syst 27(6):2869–2881

Article   Google Scholar  

Abualigah LM, Khader AT, Hanandeh ES (2018) A novel weighting scheme applied to improve the text document clustering techniques. In: Innovative computing, optimization and its applications. Springer, Cham, pp 305–320

Abualigah LMQ (2019) Feature selection and enhanced krill herd algorithm for text document clustering. Springer, Berlin, pp 1–165

Google Scholar  

Al-Aswadi FN, Chan HY, Gan KH (2019) Automatic ontology construction from text: a review from shallow to deep learning trend. Artificial Intelligence Review 1–28

Angeli G, Manning CD (2013) Philosophers are mortal: Inferring the truth of unseen facts. In Proceedings of the seventeenth conference on computational natural language learning (pp. 133-142)

Arnold P, Rahm E (2014) Extracting semantic concept relations from wikipedia. In Proceedings of the 4th International Conference on Web Intelligence, Mining and Semantics (WIMS14) (pp. 1-11)

Baker CF, Fillmore CJ, Lowe JB (1998) The berkeley framenet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1 (pp. 86-90)

Banko M, Etzioni O (2008) The tradeoffs between open and traditional relation extraction. In Proceedings of ACL-08: HLT (pp. 28-36)

Belleau F, Nolin MA, Tourigny N, Rigault P, Morissette J (2008) Bio2RDF: towards a mashup to build bioinformatics knowledge systems. J Biomed Inf 41(5):706–716

Berners-Lee T (2006). Linked Data http://www.w3.org/DesignIssues.LinkedData.html

Berners-Lee T, Hendler J (2001) Publishing on the semantic web. Nature 410(6832):1023–1024

Bizer C, Heath T, Berners-Lee T (2011) Linked data: The story so far. In Semantic services, interoperability and web applications: emerging concepts (pp. 205-227). IGI Global

Bollacker K, Cook R, Tufts P (2007) Freebase: A shared database of structured general human knowledge. In AAAI (Vol. 7, pp. 1962-1963)

Bollacker K, Evans C, Paritosh P, Sturge T, Taylor J (2008) Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD international conference on Management of data (pp. 1247-1250)

Carlson A, Betteridge J, Kisiel B, Settles B, Hruschka E, Mitchell T (2010) Toward an architecture for never-ending language learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 24, No. 1)

Chekol MW, Pirrò G, Schoenfisch J, Stuckenschmidt H (2017) Marrying uncertainty and time in knowledge graphs. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (pp. 88-94)

Chen Y, Kuang J, Cheng D, Zheng J, Gao M, Zhou A (2019) AgriKG: an agricultural knowledge graph and its applications. In International Conference on Database Systems for Advanced Applications. Springer, Cham, pp. 533–537

Chen Y, Li W, Liu Y, Zheng D, Zhao T (2010) Exploring deep belief network for chinese relation extraction. In: CIPS-SIGHAN Joint Conference on Chinese Language Processing

Culotta A, McCallum A (2005) Joint deduplication of multiple record types in relational data. In Proceedings of the 14th ACM international conference on Information and knowledge management (pp. 257-258)

Davis R, Shrobe H, Szolovits P (1993) What is a knowledge representation? AI Mag 14(1):17

De Sa C, Ratner A, Ré C, Shin J, Wang F, Wu S, Zhang C (2016) Deepdive: declarative knowledge base construction. ACM SIGMOD Record 45(1):60–67

Dong Z, Dong Q (2003) HowNet-a hybrid language and knowledge resource. In International Conference on Natural Language Processing and Knowledge Engineering, 2003. Proceedings. 2003 (pp. 820-824). IEEE

Dong X, Gabrilovich E, Heitz G, Horn W, Lao N, Murphy K, Strohmann T, Sun S, Zhang W (2014) Knowledge vault: a web-scale approach to probabilistic knowledge fusion. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 601-610)

Etzioni O, Cafarella M, Downey D, Popescu AM, Shaked T, Soderland S, Weld DS, Yates A (2005) Unsupervised named-entity extraction from the web: an experimental study. Artif Intell 165(1):91–134

Etzioni O, Banko M, Soderland S, Weld DS (2008) Open information extraction from the web. Commun ACM 51(12):68–74

Färber M, Bartscherer F, Menne C, Rettinger A (2018) Linked data quality of dbpedia, freebase, opencyc, wikidata, and yago. Sem Web 9(1):77–129

Ferre S (2019, June). Link prediction in knowledge graphs with concepts of nearest neighbours. In European Semantic Web Conference (pp. 84-100). Springer, Cham

Fortunato S (2010) Community detection in graphs. Phys Rep 486(3–5):75–174

Article   MathSciNet   Google Scholar  

Gaurav D, Tiwari SM, Goyal A, Gandhi N, Abraham A (2020) Machine intelligence-based algorithms for spam filtering on document labeling. Soft Comput 24(13):9625–9638

Hakkani-Tür D, Heck L, Tur G (2013) Using a knowledge graph and query click logs for unsupervised learning of relation detection. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 8327-8331). IEEE

Hearst MA (1992) Automatic acquisition of hyponyms from large text corpora. In Coling 1992 volume 2: The 15th international conference on computational linguistics

Heck L, Hakkani-Tür D, Tur G (2013) Leveraging knowledge graphs for web-scale unsupervised semantic parsing

Heist N (2018) Towards knowledge graph construction from entity Co-occurrence. In EKAW (Doctoral Consortium)

Hoffart J, Suchanek FM, Berberich K, Weikum G (2013) YAGO2: a spatially and temporally enhanced knowledge base from Wikipedia. Artif Intell 194:28–61

Article   MathSciNet   MATH   Google Scholar  

Jia Y, Qi Y, Shang H, Jiang R, Li A (2018) A practical approach to constructing a knowledge graph for cybersecurity. Engineering 4(1):53–60

Ji S, Pan S, Cambria E, Marttinen P, Yu PS (2020) A survey on knowledge graphs: Representation, acquisition and applications. arXiv preprint arXiv:2002.00388

Kambhatla N (2004 ) Combining lexical, syntactic, and semantic features with maximum entropy models for extracting relations. In Proceedings of the ACL 2004 on Interactive poster and demonstration sessions (pp. 22-es)

Keele S (2007) Guidelines for performing systematic literature reviews in software engineering (Vol. 5). Technical report, Ver. 2.3 EBSE Technical Report. EBSE

Klyne G, Carroll JJ, McBride B (2004) Resource description framework (RDF): concepts and abstract syntax. W3C Recommendation, Feb. 2004

Lehmann J, Isele R, Jakob M, Jentzsch A, Kontokostas D, Mendes PN, Hellmann S, Morsey M, Van Kleef P, Auer S, Bizer C (2015) DBpedia-a large-scale, multilingual knowledge base extracted from Wikipedia. Sem Web 6(2):167–195

Li L, Wang P, Yan J, Wang Y, Li S, Jiang J, Sun Z, Tang B, Chang TH, Wang S, Liu Y (2020) Real-world data medical knowledge graph: construction and applications. Artif Intell Med 103:101817

Liben-Nowell D, Kleinberg J (2007) The link-prediction problem for social networks. J Am Soc Inf Sci Technol 58(7):1019–1031

Lin Y, Han X, Xie R, Liu Z, Sun M (2018) Knowledge representation learning: A quantitative review. arXiv preprint arXiv:1812.10901

Liu Z, Han X (2018) Deep learning in knowledge graph. Springer, Singapore

Book   Google Scholar  

Liu H, Singh P (2004) ConceptNet-a practical commonsense reasoning tool-kit. BT Technol J 22(4):211–226

Matuszek C, Witbrock M, Cabral J, DeOliveira J (2006) An introduction to the syntax and content of Cyc. UMBC Computer Science and Electrical Engineering Department Collection

Miller GA (1995) WordNet: a lexical database for English. Commun ACM 38(11):39–41

Minsky M (1974). A framework for representing knowledge

Mishra S, Jain S (2019) An intelligent knowledge treasure for military decision support. Int J Web-Based Learn Teaching Technol (IJWLTT) 14(3):55–75

Momtchev V, Peychev D, Primov T, Georgiev G (2009) Expanding the pathway and interaction knowledge in linked life data. Proc. of International Semantic Web Challenge

Nakashole N, Theobald M, Weikum G (2011) Scalable knowledge harvesting with high precision and high recall. In Proceedings of the fourth ACM international conference on Web search and data mining (pp. 227-236)

Newcombe HB, Kennedy JM, Axford SJ, James AP (1959) Automatic linkage of vital records. Science 130(3381):954–959

Newman ME (2001) The structure of scientific collaboration networks. Proc Nat Acad Sci 98(2):404–409

Nicholson DN, Greene CS (2020) Constructing knowledge graphs and their biomedical applications. Comput Struct Biotechnol J 18:1414

Nickel M, Murphy K, Tresp V, Gabrilovich E (2015) A review of relational machine learning for knowledge graphs. Proc IEEE 104(1):11–33

Niu X, Sun X, Wang H, Rong S, Qi G, Yu Y (2011) Zhishi. me-weaving chinese linking open data. In International Semantic Web Conference (pp. 205-220). Springer, Berlin, Heidelberg

Noy N, Gao Y, Jain A, Narayanan A, Patterson A, Taylor J (2019) Industry-scale knowledge graphs: lessons and challenges. Queue 17(2):48–75

Paulheim H (2017) Knowledge graph refinement: a survey of approaches and evaluation methods. Sem Web 8(3):489–508

Rahm E, Bernstein PA (2001) A survey of approaches to automatic schema matching. VLDB J 10(4):334–350

Article   MATH   Google Scholar  

Rahul M, Kohli N, Agarwal R, Mishra S (2019) Facial expression recognition using geometric features and modified hidden Markov model. Int J Grid Util Comput 10(5):488–496

Ringler D, Paulheim H (2017) One knowledge graph to rule them all? Analyzing the differences between DBpedia, YAGO, Wikidata & co. In Joint GermanAustrian Conference on Artificial Intelligence (Künstliche Intelligenz) (pp. 366-372). Springer, Cham

Ruttenberg A, Rees JA, Samwald M, Marshall MS (2009) Life sciences on the Semantic Web: the Neurocommons and beyond. Brief Bioinf 10(2):193–204

Saïs F (2019). Knowledge Graph Refinement: Link Detection, Link Invalidation, Key Discovery and Data Enrichment (Doctoral dissertation, Université Paris Sud)

Sengupta S (2013) Facebook unveils a new search tool. NY Times, New York

Singhal A (2012) Introducing the knowledge graph: things, not strings. Official google blog, 5

Sowa JF (2006) Semantic Networks [Electronic resource]. Access mode: http://www.jfsowa.com/pubs/semnet.htm

Suchanek FM, Kasneci G, Weikum G (2007) Yago: a core of semantic knowledge. In Proceedings of the 16th international conference on World Wide Web (pp. 697-706)

Suchanek FM, Sozio M, Weikum G (2009) SOFIE: a self-organizing framework for information extraction. In Proceedings of the 18th international conference on World wide web (pp. 631-640)

Suh B, Convertino G, Chi EH, Pirolli P (2009) The singularity is not near: slowing growth of Wikipedia. In Proceedings of the 5th International Symposium on Wikis and Open Collaboration (pp. 1-10)

Sun Y, Han J (2012) Mining heterogeneous information networks: principles and methodologies. Synth Lect Data Mining Knowl Discov 3(2):1–159

Tejada S, Knoblock CA, Minton S (2001) Learning object identification rules for information integration. Inf Syst 26(8):607–633

Tiwari SM, Jain S, Abraham A, Shandilya S (2018) Secure Semantic Smart HealthCare (S3HC). J Web Eng 17(8):617–646

Tiwari S, Abraham A (2020) Semantic assessment of smart healthcare ontology. International Journal of Web Information Systems

Vrandecic D (2012) Wikidata: a new platform for collaborative data collection. In Proceedings of the 21st international conference on world wide web (pp. 1063-1064)

Wang J, Liu J, Kong L (2017) Ontology construction based on deep learning. In Advances in Computer Science and Ubiquitous Computing Springer, Singapore

Wang P, Jiang H, Xu J, Zhang Q (2019) Knowledge graph construction and applications for Web search and beyond. Data Intell 1(4):333–349

Wang Z, Li J, Wang Z, Li S, Li M, Zhang D, Shi Y, Liu Y, Zhang P, Tang J (2013) XLore: A Large-scale English-Chinese Bilingual Knowledge Graph. In International semantic web conference (Posters & Demos) (Vol. 1035, pp. 121-124)

Wu T, Qi G, Li C, Wang M (2018) A survey of techniques for constructing Chinese knowledge graphs and their applications. Sustainability 10(9):3245

Wu W, Li H, Wang H, Zhu KQ (2012) Probase: A probabilistic taxonomy for text understanding. In Proceedings of the 2012 ACM SIGMOD International Conference on Management of Data (pp. 481-492)

Wu T, Wang H, Li C, Qi G, Niu X, Wang M, Li L, Shi C (2019) Knowledge graph construction from multiple online encyclopedias. World Wide Web 1–28

Xu B, Xu Y, Liang J, Xie C, Liang B, Cui W, Xiao Y (2017) CN-DBpedia: a never-ending Chinese knowledge extraction system. In International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems (pp. 428-438). Springer, Cham

Yan J, Wang C, Cheng W, Gao M, Zhou A (2018) A retrospective of knowledge graphs. Front Comput Sci 12(1):55–74

Zhang J, Liu J, Wang X (2016) Simultaneous entities and relationship extraction from unstructured text. Int J Database Theory Appl 9(6):151–160

Zhang Z, Zhuang F, Qu M, Lin F, He Q (2018) Knowledge graph embedding with hierarchical relation structure. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (pp. 3198-3207)

Zhao M, Wang H, Guo J, Liu D, Xie C, Liu Q, Cheng Z (2019) Construction of an industrial knowledge graph for unstructured Chinese text learning. Appl Sci 9(13):2720

Zhong B, Liu J, Du Y, Liaozheng Y, Pu J (2016) Extracting attributes of named entity from unstructured text with deep belief network. Int J Database Theory Appl 9(5):187–196

Zhu G, Iglesias CA (2015) Sematch: Semantic Entity Search from Knowledge Graph. In SumPre-HSWI@ ESWC

Zhu J, Nie Z, Liu X, Zhang B, Wen JR (2009) Statsnowball: a statistical approach to extracting entity relationships. In Proceedings of the 18th international conference on World wide web (pp. 101-110)

Zou X (2020) A survey on application of knowledge graph. JPhCS 1487(1):012016

Download references

Author information

Authors and affiliations.

Universidad Autonoma de Tamaulipas, Ciudad Victoria, Mexico

Sanju Tiwari

Faculty of Computer Sciences and Engineering, Hodeidah University, Hodeidah, Yemen

Fatima N. Al-Aswadi

Dept of Computer Science and Engineering, Indian Institute of Technology, Delhi, India

Devottam Gaurav

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Sanju Tiwari .

Ethics declarations

Conflict of interest.

The authors declare that they have no competing interests.

Compliance with Ethical Standards

This study was not funded by any grant. No animals were involved. This article does not contain any studies with human participants or animals performed by any of the authors. Informed consent was obtained from all individual participants included in the study.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Tiwari, S., Al-Aswadi, F.N. & Gaurav, D. Recent trends in knowledge graphs: theory and practice. Soft Comput 25 , 8337–8355 (2021). https://doi.org/10.1007/s00500-021-05756-8

Download citation

Accepted : 16 March 2021

Published : 16 April 2021

Issue Date : July 2021

DOI : https://doi.org/10.1007/s00500-021-05756-8

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Knowledge graphs
• Knowledge extraction
• Learning techniques
• Find a journal
• Publish with us
• Track your research

Help | Advanced Search

Computer Science > Discrete Mathematics

Title: graph theory.

Abstract: This is a replacement paper. There are 6 chapters. The first two chapters are introductory. The third chapter is on extremal graph theory. The fourth chapter is about algebra in graph theory. The fifth chapter is focused on algorithms. The third section of the fifth chapter deals with computable time. The sixth chapter has sections on probability and enumeration.

Submission history

Access paper:, references & citations.

• Google Scholar
• Semantic Scholar

DBLP - CS Bibliography

Bibtex formatted citation.

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

List of Publications

Pre-prints -- subject to revision.

• Random 2-SAT with prescribed literal degrees Algorithmica 48, 249-265 [Co-authors: C. Cooper and G. Sorkin]
• The cover time of sparse random graphs Random Structures and Algorithms 30, John Wiley and Sons, 1-16. [Co-author: C.Cooper]
• Random k-SAT: A tight threshold for moderately growing k Combinatorica 25, 297-305. [Co-author: N.C. Wormald]
• Fast Monte-Carlo algorithms for finding low rank approximations JACM 51, 1025-1041 [Co-authors: R.Kannan and S.Vempala]
• Avoidance of a giant component in half the edge set of a random graph Program to check some calculations Random Structures and Algorithms 25, John Wiley and Sons, 432-449. [Co-authors: T. Bohman and N. Wormald]
• The size of the largest strongly connected component of a random digraph with a given degree sequence Combinatorics, Probability and Computing 13, 319-338. [Co-author: C.Cooper]
• On the b-independence number of sparse random graphs Combinatorics, Probability and Computing 13, 295-310. [Co-author: G. Atkinson]
• The emergence of a giant component in a random subgraph of pseudo-random graphs Random Structures and Algorithms 24, John Wiley and Sons, 42-50. [Co-authors: M.Krivelevich, R. Martin]
• Randomly colouring graphs with lower bounds on girth and maximum degree Random Structures and Algorithms 23, John Wiley and Sons, 167-179. [Co-author: M.E. Dyer]
• Crawling on web graphs Internet Mathematics 1, 57-90 [Co-author: C.Cooper]
• On Randomly Generated Intersecting Hypergraphs Electronic Journal on Combinatorics, R29 [Co-authors: T. Bohman, C.Cooper, R. Martin, M. Ruszinko]
• Arc-Disjoint Paths in Expander Digraphs SIAM Journal on Computing 32, 326-344. [Co-author: T. Bohman]
• On a general model of web graphs Random Structures and Algorithms 22, John Wiley and Sons, 311-335. [Co-author: C. Cooper]
• A probabilistic analysis of randomly generated binary constraint satisfaction problems Theoretical Computer Science 290, 1815-1828 [Co-authors: M.E. Dyer and M. Molloy]
• The cover time of sparse random graphs JOURNAL VERSION Proceedings of SODA 2003, 140-147. [Co-author: C.Cooper]
• How many random edges make a dense graph Hamiltonian? Random Structures and Algorithms 22, 33-42. [Co-authors: T. Bohman, R. Martin]
• Hamilton Cycles in Random Subgraphs of Pseudo-Random Graphs Discrete Mathematics 256, 137-150 [Co-author: M. Krivelevich]
• On graph irregularity strength Journal of Graph Theory 41, 120-137. [Co-authors: R.J. Gould, M. Karonski, F. Pfender]
• Random regular graphs of non-constant degree: independence and chromatic number Combinatorics, Probability and Computing 11, 323-342. [Co-authors: C.Cooper, B.Reed, O.Riordan]
• On random symmetric travelling salesman problems Proceedings of the 43rd Annual IEEE Symposium on Foundations of Computer Science, 789-790. JOURNAL VERSION
• Random regular graphs of non-constant degree: connectivity and Hamilton cycles Combinatorics, Probability and Computing 11, 249-262. [Co-authors: C.Cooper, B.Reed]
• Probabilistic analysis of the Traveling Salesman Problem The traveling salesman problem and its variations, G. Gutin and A.P. Punnen (Eds.), Kluwer Academic Publisher, 257-308. [Co-author: J.Yukich]
• Crawling on web graphs STOC 2002, 419-427 JOURNAL VERSION [Co-author: C.Cooper]
• Random k-SAT: A tight threshold for moderately growing k Proceedings of the Fifth International Symposium on Theory and Applications of Satisfiability Testing, 1-6. JOURNAL VERSION [Co-author: N.C. Wormald]
• Optimal sequencing by hybridization in rounds Journal of Computational Biology 9, 355-369. [Co-author: B. Halldorsson]
• A New Rounding Procedure for the Assignment Problem with Applications to Dense Graph Arrangement Problems Mathematical Programming A 92, 1-36. [Co-authors: S.Arora and H.Kaplan]
• Multi-coloured Hamilton cycles in randomly coloured random graphs Combinatorics, Probability and Computing 11, 129--134. [Co-author: C. Cooper]
• A note on random 2-SAT with prescribed literal degrees Proceedings of SODA 2002, 316-320. JOURNAL VERSION [Co-authors: C. Cooper and G. Sorkin]
• Balls and Bins Models with Feedback Proceedings of SODA 2002, 308-315. [Co-authors: E. Drinea and M. Mitzenmacher]
• Vertex covers by edge disjoint cliques Combinatorica 21 171-197. [Co-authors: T.Bohman, M.Ruszinko, L.Thoma]
• G-intersecting families Combinatorics, Probability and Computing 10, 367-384. [Co-authors: T.Bohman, M.Ruszinko, L.Thoma]
• Arc-Disjoint Paths in Expander Digraphs Proceedings of FOCS 2001, 558-567. [Co-author: T. Bohman] JOURNAL VERSION
• On a general model of web graphs Proceedings of ESA 2001, 500-511. [Co-author: C. Cooper] JOURNAL VERSION
• Randomly colouring graphs with lower bounds on girth and maximum degree Proceedings of FOCS 2001, 579-587. JOURNAL VERSION [Co-author: M.E. Dyer]
• Avoiding a giant component Random Structures and Algorithms 19, John Wiley and Sons, 75-85 [Co-author: T. Bohman]
• Optimal sequencing by hybridization in rounds Proceedings of RECOMB2001, 141-148. [Co-author: B. Halldorsson] JOURNAL VERSION
• Edge disjoint paths in expander graphs SIAM Journal on Computing 30, 1790-1801.
• On Markov chains for randomly H-colouring a graph Journal of Algorithms 39, 117-134 [Co-authors: C.Cooper and M.E.Dyer]
• Hamilton cycles in the union of random permutations Random Structures and Algorithms 18, John Wiley and Sons, 83-94.
• The probabilistic relationship between the assignment and asymmetric traveling salesman problems Proceedings of SODA 2001, 652-660. [Co-author: G. Sorkin] JOURNAL VERSION
• On the number of perfect matchings and Hamilton cycles in epsilon-regular non-bipartite graphs Electronic Journal of Combinatorics 7, R57.
• Mixing properties of the Swendsen-Wang process on classes of graphs II Journal of Mathematical Physics 4, 1499-1527. [Co-authors: C.Cooper, M.E.Dyer and R.Rue]
• A note on random minimum length spanning trees Electronic Journal of Combinatorics 7, R41. [Co-authors: M.Ruszinko, L.Thoma]
• Optimal Construction of Edge-Disjoint Paths in Random Regular Graphs Combinatorics, Probability and Computing 9, 241-264. [Co-author: L.Zhao]
• Hamilton cycles in random graphs and directed graphs Random Structures and Algorithms 16, John Wiley and Sons, 369-401. [Co-author: C.Cooper]
• Edge disjoint paths in expander graphs Proceedings of SODA 2000, 717-725. JOURNAL VERSION
• Average-case analysis of shortest-paths algorithms in the vertex potential model Random Structures and Algorithms 16, John Wiley and Sons, 33-46. [Co-authors: C.Cooper, V.Priebe and K.Mehlhorn]
• A Note on Sparse Random Graphs and Cover Graphs Electronic Journal of Combinatorics 7, R19. [Co-authors: T.Bohman, M.Ruszinko, L.Thoma]
• On Hamilton cycles in sparse random graphs with minimum degree at least k Journal of Graph Theory 34, John Wiley and Sons, 42-59. [Co-authors: B.Bollobas, C.Cooper and T.I.Fenner]
• Min-Wise independent linear permutations Electronic Journal of Combinatorics 7, R26. [Co-authors: T.Bohman and C.Cooper]
• On counting independent sets in sparse graphs Proceedings of FOCS '99, 210-217. [Co-authors: M.E.Dyer and M.R.Jerrum] JOURNAL VERSION
• On the Power of Universal Bases in Sequencing by Hybridization Proceedings of RECOMB '99 Journal version appeared as Optimal Reconstruction of a Sequence from its Probes in Journal of Computational Biology 6, 361-368. [Co-authors: F.Preparata and E.Upfal]
• Mixing Properties of the Swendsen-Wang Process on Classes of Graphs Random Structures and Algorithms 15, John Wiley and Sons, 242-261. [Co-author: C.Cooper]
• Splitting an expander graph Journal of Algorithms 33, 166-172. [Co-author: M.Molloy]
• Existence and construction of edge low congestion paths on expander graphs Random Structures and Algorithms 14, John Wiley and Sons, 87-109. [Co-authors: A.Broder and E.Upfal]
• Optimal Construction of Edge-Disjoint Paths in Random Regular Graphs Proceedings of SODA '99, 346-355. JOURNAL VERSION [Co-author: L.Zhao]
• Torpid mixing of some MCMC algorithms in Statistical Physics Proceedings of FOCS '99, 218-229. [Co-authors: C.Borgs, J.Chayes, J.H.Kim, P.Tetali, E.Vigoda and V.Vu]
• Clustering in large graphs and matrices Proceedings of SODA '99, 291-299 [Co-authors: P.Drineas, R.Kannan, S.Vempala and V.Vinay] JOURNAL VERSION
• A simple algorithm for constructing Szemeredi's Regularity Partition Electronic Journal of Combinatorics, 6(1) R17. [Co-author: R.Kannan]
• Optimal construction of edge-disjoint paths in random graphs SIAM Journal on Computing 28, 541-574. [Co-authors: A.Z.Broder, S.Suen and E.Upfal]
• On Perfect Matchings and Hamiltonian Cycles in Sums of Random Trees SIAM Journal on Discrete Mathematics 12, 208-216. [Co-authors: M.Karonski and L.Thoma]
• Log-Sobolev inequalities and sampling from log-concave distributions Annals of Applied Probability 9, 14-26. [Co-author: R. Kannan]
• Quick approximation to matrices and applications Combinatorica 19, 175-220.  [Co-author: R. Kannan]
• Maximum matchings in sparse random graphs: Karp-Sipser revisited Random Structures and Algorithms 12, John Wiley and Sons ,111-178. [Co-authors: J.Aronson and B.Pittel]
• Greedy algorithms for the shortest common superstring that are asymptotically optimal Algorithmica 21, 21-36. [Co-author: W.Szpankowski]
• Min-Wise Independent permutations Proceedings of the 30th Annual ACM Symposium on Theory of Computing, 327-336. [Co-authors: A.Broder, M.Charikar, M.Mitzenmacher]
• Approximately counting Hamilton cycles in dense graphs SIAM Journal on Computing 27, 1262-1272. [Co-authors: M.E.Dyer and M.R.Jerrum]
• On balls and bins with deletions Proceedings of Random '98, Lecture Notes in Computer Science 1518, Springer, 145-158. [Co-authors: M.Cole, B.Maggs, M.Mitzenmacher, A.Richa, R.Sitaraman]
• Disjoint Paths in Expander Graphs via Random Walks: a Short Survey Proceedings of Random '98, Lecture Notes in Computer Science 1518, Springer, 1-14.
• A polynomial-time algorithm for learning noisy linear threshold functions Algorithmica 22, 35-52. [Co-authors: A.Blum, R.Kannan and S.Vempala]
• Average case analysis of the merging algorithm of Hwang and Lin Algorithmica 22, 483-489. [Co-authors: W.Fernandez de la Vega and M.Santha]
• Fast Monte-Carlo algorithms for finding low rank approximations JOURNAL VERSION Proceedings of 1998 IEEE Symposium on Foundations of Computer Science, 370-378. [Co-authors: R.Kannan and S.Vempala]
• Random minimum length spanning trees in regular graphs Combinatorica 18, 311-333. [Co-author: A.Beveridge and C.McDiarmid]
• Existence and construction of edge low congestion paths on expander graphs Proceedings of the 29th Annual ACM Symposium on Theory of Computing, 531-539. [Co-authors: A.Broder and E.Upfal] JOURNAL VERSION
• Average-case analysis of shortest-paths algorithms in the vertex potential model. Randomization and approximation techniques in Computer Science (Proceedings of RANDOM '97) Lecture Notes in Computer Science 1269, 15-26. [Co-authors: C.Cooper, V.Priebe and K.Mehlhorn] JOURNAL VERSION
• Improved approximation algorithms for MAX k-CUT and MAX BISECTION Algorithmica 18, 61-77. [Co-author: M.R.Jerrum]
• On the connectivity of random k-th nearest neighbour graphs Combinatorics, Probability and Computing 4, 343-362. [Co-author: C.Cooper]
• Analysis of Two Simple Heuristics on a Random Instance of k-SAT Journal of Algorithms 20, 312-355. [Co-author: S.Suen]
• Perfect matchings in random r-regular, s-uniform hypergraphs Combinatorics, Probability and Computing 5, 1-15. [Co-authors: C.Cooper, M.Molloy and B.Reed]
• Greedy algorithms for the shortest common superstring that are asymptotically optimal Proceedings of ESA96, Springer Lecture Notes in Computer Science 1136, 194-207. [Co-author: W.Szpankowski] JOURNAL VERSION
• Analysis of Parallel Algorithms for Finding A Maximal Independent Set in A Random Hypergraph Random Structures and Algorithms 9, John Wiley and Sons , 359-378. [Co-author: H.Chen]
• Generating and counting Hamilton cycles in random regular graphs Journal of Algorithms 21, 176-198. [Co-authors: M.R.Jerrum, M.Molloy, R.Robinson and N.C.Wormald]
• A New Rounding Procedure for the Assignment Problem with Applications to Dense Graph Arrangement Problems Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, 21-30 [Co-authors: S.Arora and H.Kaplan] JOURNAL VERSION The journal version of this paper is in preparation.
• A polynomial-time algorithm for learning noisy linear threshold functions Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, 330-338. [Co-authors: A.Blum, R.Kannan and S.Vempala] JOURNAL VERSION
• Learning linear transformations Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, 359-368. [Co-authors: R.Kannan and M.R.Jerrum]
• The regularity lemma and approximation schemes for dense problems Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, 12-20. [Co-author: R.Kannan] JOURNAL VERSION
• A general approach to dynamic packet routing with bounded buffers Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, 390-399. [Co-authors: A.Z.Broder and E.Upfal] JOURNAL VERSION
• Coloring Bipartite Hypergraphs IPCO '96, 345-358. [Co-author: H.Chen]
• An efficient algorithm for the vertex-disjoint paths problem in random graphs. Proceedings of SODA '96, 261-268. [Co-authors: A.Z.Broder, S.Suen and E.Upfal]
• On the best case of heapsort Journal of Algorithms 20, 205-217. [Co-authors: B.Bollobas and T.I.Fenner]
• An Analysis of a Monte Carlo Algorithm for Estimating the Permanent Combinatorica 15, 67-83. [Co-author: M.R.Jerrum]
• Multicoloured Hamilton Cycles Electronic Journal of Combinatorics 2, R10 [Co-authors: M.Albert and B.Reed]
• Ordering Clone Libraries in Computational Biology Journal of Computational Biology 2, 207--218 [Co-authors: M.E.Dyer and S.Suen]
• When is the Assignment Bound Tight for the Asymmetric Traveling-Salesman Problem ? SIAM Journal on Computing 24, 484-493. [Co-authors: R.M.Karp and B.Reed]
• Polynomial Time Randomised Approximation Schemes for Tutte-Grothendieck Invariants: The Dense Case Random Structures and Algorithms 6, John Wiley and Sons , 459-478. [Co-authors: N.Alon and D.Welsh]
• Analysis of a simple greedy matching algorithm on random cubic graphs Combinatorics, Probability and Computing 4, 47-66. [Co-authors: A.J.Radcliffe and S.Suen]
• Perfect Matchings in Random s-Uniform Hypergraphs Random Structures and Algorithms 7, John Wiley and Sons , 41-57. [Co-author: S.Janson]
• Covering the edges of a random graph by cliques Combinatorica 15, 1-9. [Co-author: B.Reed]
• Multicoloured Hamilton cycles in random graphs: an anti-Ramsey threshold. Electronic Journal of Combinatorics 2, R19. [Co-author: C.Cooper]
• On key storage in secure networks Journal of Cryptology 8, 189-200. [Co-authors: M.E.Dyer, T.I.Fenner and A.Thomason]
• The worst-case running time of the random simplex algorithm is exponential in the height Information Processing Letters 56, 79-81. [Co-authors: A.Z.Broder, M.E.Dyer, P.Raghavan and E.Upfal]
• Improved approximation algorithms for MAX k-CUT and MAX BISECTION Proceedings of the fourth Integer Programming and Combinatorial Optimization Conference (IPCO4)}, Springer-Verlag Lecture Notes in Computer Science 920, 1-13. [Co-author: M.R.Jerrum] JOURNAL VERSION
• Probabilistic analysis of an algorithm in the theory of markets in indivisible goods Annals of Applied Probability 5, 768-808. [Co-author: B.Pittel]
• Random walks, totally unimodular matrices and a randomised dual simplex algorithm. Mathematical Programming 64, 1-16. [Co-author: M.E.Dyer]
• Multicoloured trees in random graphs Random Structures and Algorithms 5, John Wiley and Sons , 45-56. [Co-author: B.D.Mckay]
• Randomised greedy matching II Proceedings of 4th Annual Symposium on Discrete Algorithms (1994) 141-149. [Co-authors: J.Aronson, M.E.Dyer and S.Suen] JOURNAL VERSION
• Optimal construction of edge disjoint paths in random graphs Proceedings of 4th Annual Symposium on Discrete Algorithms (1994) 603-612 [Co-authors: A.Z.Broder, S.Suen and E.Upfal] JOURNAL VERSION
• Approximately counting Hamilton cycles in dense graphs Proceedings of 4th Annual ACM/SIAM Symposium on Discrete Algorithms 336-343 [Co-authors: M.E.Dyer and M.R.Jerrum] JOURNAL VERSION
• On the Problem of Approximating the Number of Bases of a Matroid Information Processing letters 50, 9-11. [Co-authors: Y.Azar and A.Broder]
• The probability of unique solutions of sequencing by hybridization Journal of Computational Biology 1, 105-110. [Co-authors: M.E.Dyer and S.Suen]
• Hamilton cycles in random regular digraphs Combinatorics, Probability and Computing 3, 39-50. [Co-authors: C.Cooper and M.J.Molloy]
• Finding hidden Hamilton cycles Random Structures and Algorithms 5, John Wiley and Sons , 395-410. [Co-authors: A.Broder and E.Shamir]
• Near-perfect token distribution Random Structures and Algorithms 5, John Wiley and Sons , 559-572 [Co-authors: A.Broder, E.Shamir and E.Upfal]
• Broadcasting in random graphs, Discrete Applied Mathematics 54, 77-79. [Co-author: M.Molloy]
• Hamilton cycles in a class of random directed graphs Journal of Combinatorial Theory B 62, 151-163 [Co-author: C.Cooper]
• Polynomial time randomised approximation schemes for the Tutte polynomial of dense graphs. Proceedings of the 35th Annual IEEE Symposium on Foundations of Computer Science 24-35. [Co-authors: N.Alon and D.Welsh] JOURNAL VERSION
• On the complexity of computing the diameter of a polyhedron Journal of Computational Complexity 4, 207-219. [Co-author: S.Teng]
• Existence and construction of edge disjoint paths on expander graphs SIAM Journal of Computing 23, 976-989. [Co-authors: A.Broder and E.Upfal]
• On the satisfiability and maximum satisfiability of random 3-CNF formulas Proceedings of the Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, 322-330. [Co-authors: A.Broder and E.Upfal]
• Analysis of a simple greedy matching algorithm on random cubic graphs Proceedings of the Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, 341-351. [Co-authors: A.J.Radcliffe and S.Suen] JOURNAL VERSION
• The average performance of the greedy matching algorithm Annals of Applied Probability 3, 526-552. [Co-authors: M.E.Dyer and B.Pittel]
• Polychromatic Hamilton cycles Discrete Mathematics118, 69-74. [Co-author: B.Reed]
• A mildly exponential algorithm for estimating the number of knapsack solutions. Combinatorics, Probability and Computing 2, 271-284. [Co-authors: M.E.Dyer,R.Kannan, A. Kapoor, L.Perkovic and U.Vazirani]
• On the independence and chromatic numbers of random regular graphs Journal of Combinatorial Theory 54 , 123-132. [Co-author: T.Luczak]
• On the expected performance of a parallel algorithm for finding maximal independent sets Random Structures and Algorithms 3, John Wiley and Sons , 215 - 222. [Co-authors: N.Calkin and L.Kucera]
• On small subgraphs of random graphs Proceedings of Random Graphs 1989, Poznan, John Wiley and Sons 67 - 90.
• Counting Hamilton cycles in random directed graphs Random Structures and Algorithms 3, John Wiley and Sons , 235-242. [Co-author: S.Suen]
• Probabilistic analysis of the generalised assignment problem Mathematical Programming 55, 169-181. [Co-author: M.E.Dyer]
• Random walks, totally unimodular matrices and a randomised dual simplex algorithm. Proceedings of IPCO '92, 72-84. [Co-author: M.E.Dyer] JOURNAL VERSION
• When is the assignment bound asymptotically tight for the asymmetric traveling-salesman problem? Proceedings of IPCO '92, 453-461. [Co-author: R.M.Karp and B.Reed] JOURNAL VERSION
• Existence and construction of edge disjoint paths on expander graphs Proceedings of the 24th Annual ACM Symposium on Theory of Computing, 140-149. [Co-authors: A.Broder and E.Upfal] JOURNAL VERSION
• On subgraph sizes in random graphs Combinatorics, Probability and Computing 1 , 123-134. [Co-authors: N.J.Calkin and B.D.Mckay]
• Separator based parallel divide and conquer in computational geometry Proceedings of the 4th Annual ACM Symposium on Parallel Algorithms and Architecture [Co-authors: G.L.Miller and S.Teng]
• Randomised greedy matching Random Structures and Algorithms 2, John Wiley and Sons , 29-46. [Co-author: M.E.Dyer]
• Occupancy problems and random algebras Discrete Mathematics 87 , 1-8. [Co-author: M.Albert]
• A random polynomial time algorithm for approximating the volume of convex bodies Journal of the Association for Computing Machinery 38 , 1-17. [Co-authors: M.E.Dyer and R.Kannan]
• Spanning maximal planar subgraphs of random graphs Random Structures and Algorithms 2, John Wiley and Sons , 225-232. [Co-author: B.Bollobas]
• A parallel algorithm for finding the lexicographically first depth first search tree in a random graph Random Structures and Algorithms 2, John Wiley and Sons , 233-240. [Co-author: M.E.Dyer]
• On the length of the longest monotone subsequence of a random permutation The Annals of Applied Probability 1 , 301-305.
• On hidden Hamilton cycles Proceedings of the 23rd Annual ACM Symposium on Theory of Computing,182-189. [Co-authors: A.Broder and E.Shamir] JOURNAL VERSION
• Computing the volume of a convex body: a case where randomness provably helps Proceedings of AMS Symposium on Probabilistic Combinatorics and Its Applications , 123-170. [Co-author: M.E.Dyer]
• On a conjecture of Bondy and Fan Ars Combinatoria 11, 191-205. [Co-authors: C.J.H.McDiarmid and B.Reed]
• Edge disjoint trees in random graphs Periodica Mathematica Hungarica 21 , 28-30 [Co-author: T.Luczak]
• Probabilistic analysis of a parallel algorithm for finding maximal independent sets Random Structures and Algorithms 1, John Wiley and Sons , 39-50. [Co-author: N.Calkin]
• The limiting probability that a-in, b-out is strongly connected Journal of Combinatorial Theory 48 , 117-134. [Co-author: C. Cooper]
• On an optimisation problem with nested constraints Discrete Applied Mathematics 26 , 159-173. [Co-author: M.E. Dyer]
• Probabilistic analysis of graph algorithms Computing Supplement 7, Computational Graph Theory, Edited by Tinhofer, Mayr, Noltemeier and Syslo, Springer-Verlag , 209-234.
• Probabilistic analysis of the generalised assignment problem Proceedings of Integer Programming and Combinatorial Optimization 1, 189 - 200. [Co-author: M.E.Dyer] JOURNAL VERSION
• On the independence number of random graphs Discrete Mathematics 81 , 171-176.
• On patching algorithms for random asymmetric travelling salesman problems Mathematical Programming 46 , 361-378. [Co-author: M.E. Dyer]
• Pancyclic random graphs in Proceedings of Random Graphs '87, Edited by M.Karonski, J.Jaworski and A.Rucinski, John Wiley and Sons , 29-39. [Co-author: C. Cooper]
• Parallel colouring of random graphs in Proceedings of Random Graphs '87, Edited by M.Karonski, J.Jaworski and A.Rucinski, John Wiley and Sons , 41-52. [Co-author: L. Kucera]
• Hamiltonian cycles in a class of random graphs: one step further in Proceedings of Random Graphs '87, Edited by M.Karonski, J.Jaworski and A.Rucinski, John Wiley and Sons , 53-59. [Co-author: T. Luczak]
• Greedy matching on the line SIAM Journal on Computing 19 , 666-672. [Co-authors: C.J.H.McDiarmid and B.Reed]
• Hamilton cycles in random graphs with minimal degree at least k in A tribute to Paul Erdos, edited by A.Baker, B.Bollobas and A.Hajnal, 59 - 96. [Co-authors: B.Bollobas and T.I.Fenner]
• A new integer programming formulation of the permutation flowshop problem European Journal of Operations Research 40 , 90-98 [Co-author: J. Yadegar]
• Probabilistic analysis of random m-dimensional knapsack problems Mathematics of Operations Research 14 , 162-176. [Co-author: M.E. Dyer]
• A randomized algorithm for fixed dimensional linear programming Mathematical Programming 44 , 203-212. [Co-author: M.E. Dyer]
• Algorithms for assignment problems on an array processor Parallel Computing 11 , 151-162 [Co-authors: J.Yadegar, D.Parkinson and S.El-Horbaty]
• Algorithms for shortest paths problems on an array processor Proceedings of the Fourth International Conference on Supercomputing, Santa Clara CA. [Co-authors: J.Yadegar, D.Parkinson and S.El-Horbaty]
• Random graph orders Order 6 , 19-30. [Co-author: M.Albert]
• The solution of some random NP-hard problems in polynomial expected time Journal of Algorithms 10 , 451-489. [Co-author: M.E. Dyer]
• Survival time of a random graph Combinatorica 9 , 133-143.
• On the number of hamilton cycles in a random graph Journal of Graph Theory 13 , 719-735. [Co-author: C.Cooper]
• On matchings and Hamilton cycles in random graphs Surveys in Combinatorics, 1989, Proceedings of British Combinatorial Conference, 84-114.
• On random minimum length spanning trees Combinatorica 9 , 363-374. [Co-author: C.McDiarmid]
• A random polynomial time algorithm for approximating the volume of convex bodies Proceedings of 21st ACM Symposium on Theory of Computing, 375-381. [Co-authors: M.E.Dyer and R.Kannan] JOURNAL VERSION
• Probabilistic analysis of a relaxation for the k-median problem Mathematics of Operations Research 13, 1-31. [Co-authors: S. Ahn, C. Cooper and G. Cornuejols]
• On the random construction of heaps Information Processing Letters 27, 103-109.
• Reconstructing truncated integer variables satsifying linear congruences SIAM Journal on Computing 17, 262-280. [Co-authors: J. Hastad, R. Kannan, J.C. Lagarias, A. Shamir]
• Finding hamilton cycles in sparse random graphs Journal of Combinatorial Theory B 44, 230-250.
• An algorithm for finding hamilton cycles in random digraphs Journal of Algorithms 9 181-204.
• Partitioning random graphs into large cycles Discrete Mathematics 70 149-158.
• On the complexity of computing the volume of a polyhedron SIAM Journal on Computing 17 , 967-974. [Co-author: M.E. Dyer]
• Edge-colouring random graphs Journal of Combinatorial Theory B 45 , 135-149. [Co-authors: B. Jackson, C. Mc.Diarmid and B. Reed]
• On large induced trees in sparse random graphs Journal of Cominatorial Theory B42, 181-195. [Co-author: B. Jackson]
• Parallel algorithms for finding hamilton cycles in random graphs Information Processing Letters 25, 111-117.
• On the strength of connectivity of random subgraphs of the n-cube Annals of Discrete Mathematics, 33, 17-40. [Co-authors: M.E. Dyer and L.R. Foulds]
• A probabilistic analysis of the next fit decreasing bin packing heuristic Operations Research Letters 5, 233-236. [Co-authors: J. Csirik, J.B.G. Frenk, G. Galambos and A.H.G. Rinnooy Kan]
• On the exact solution of random travelling salesman problems with medium-sized integer costs SIAM Journal on Computing, 16, 1052-1072.
• Large holes in sparse random graphs Combinatorica 7, 265-274. [Co-author: B. Jackson]
• An algorithm for finding hamilton paths and cycles in random graphs Combinatorica 7, 327-341. [Co-authors: B.Bollobas and T.I.Fenner]
• On large matchings and cycles in sparse random graphs Discrete Mathematics 59, 243-256.
• On the Lagarias-Odlyzko algorithm for the subset-sum problem SIAM Journal on Computing 15, 536-539.
• Maximum matchings in a class of random graphs Journal of Combinatorial Theory Series B 40, 196-212.
• Planar 3DM is NP-Complete Journal of Algorithms 7, 174-184. [Co-author: M.E. Dyer]
• Linear programs with random costs Mathematical Programming 35, 3-16. [Co-authors: M.E. Dyer and C.J.H. Mc.Diarmid]
• Fast algorithms for some random NP-hard problems Extended abstract in the proceedings of the 27th Annual IEEE Symposium on the Foundations of Computer Science. [Co-author: M.E. Dyer]
• Expected behaviour of line balancing heuristics I.M.A. Journal on Mathematics in Management 1, 1-11. [Co-author: I. Baybars]
• On the value of a random minimum spanning tree problem Discrete Applied Mathematics 10, 47-56.
• An algorithm for finding hamiltonian cycles in random graphs Proceedings of the 17th Annual ACM Symposium on Theory of Computing, 430-439. [Co-authors: B. Bollobas and T.I. Fenner] JOURNAL VERSION
• The shortest path problem for graphs with random arc-lengths Discrete Applied Mathematics 10, 57-77. [Co-author: G.R. Grimmett]
• On the complexity of partitioning graphs into connected subgraphs Discrete Applied Mathematics 10, 139-153. [Co-author: M.E. Dyer]
• An analysis of some graph theoretic heuristics for the facilities layout problem European Journal of Operational Research 20, 102-114. [Co-authors: M.E. Dyer and L.R. Foulds]
• A parallel algorithm for all-pairs shortest paths in a random graph Proceedings of the 22nd Allerton Conference on Communication, Control and Computing, 663-670. [Co-author: L. Rudolph]
• A simple heuristic for the p-centre problem Operations Research Letters 3, 285-288. [Co-author: M.E. Dyer]
• An algorithm for finding a matroid basis that maximises the product of the weights of the elements BIT 25, 434-438. [Co-author: T.I. Fenner]
• On matchings and hamiltonian cycles in random graphs Annals of Discrete Mathematics 28, 23-46. [Co-author: B. Bollobas]
• Limit distribution for the existence of hamiltonian cycles in random bipartite graphs European Journal of Combinatorics 6, 327-334.
• Approximation algorithms for the m-dimensional 0-1 knapsack problem: worst-case and probabilistic analysis European Journal of Operations Research 15, 100-109. [Co-author: M.R.B. Clarke]
• A partitioning algorithm for minimum weighted euclidean matching Information Processing Letters 18, 59-62. [Co-author: M.E.Dyer]
• Linear congruential generators do not produce random sequences Proceedings of 25th IEEE Conference on the Foundations of Computer Science, 480-484. [Co-authors: R.Kannan and J.C.Lagarias] JOURNAL VERSION
• Long cycles in sparse random graphs Graph theory and combinatorics, 59-64. Proceedings of Cambridge Combinatorial Conference in honour of Paul Erdos. [Co-authors: B.Bollobas and T.I.Fenner]
• Algebraic flows Annals of Discrete Mathematics 19, 135-146.
• Hamiltonian cycles in random regular graphs Journal of Combinatorial Theory, Series B, 103-112. [Co-author: T.I. Fenner]
• On the existence of polychromatic sets of edges in graphs and digraphs Progress in Graph Theory, Edited by J.A. Bondy and U.S.R. Murty, Academic Press, 219-232. [Co-author: T.I. Fenner]
• Partitioning heuristics for two geometric maximisation problems Operations Research Letters 3, 267-270. [Co-authors: M.E. Dyer and C.J.H. Mc.Diarmid]
• On the quadratic assignment problem Discrete Applied Mathematics 5, 89-98. [Co-author: J. Yadegar]
• An extension of Christofides Heuristic to the k-person travelling salesman problem Discrete Applied Mathematics 6, 79-83.
• On the existence of hamiltonian cycles in a class of random graphs Discrete Mathematics 45, 301-305. [Co-author: T.I. Fenner]
• Complexity of a 3-dimensional assignment problem European Journal of Operational Research 13, 161-164.
• On the worst-case performance of some algorithms for the asymmetric travelling salesman problem Networks 12, 23-39. [Co-authors: G. Galbiati, F. Maffioli]
• Algebraic linear programming Mathematics of Operations Research 7, 172-182.
• K-Greedy Algorithms Proceedings of CO81 Conference on Combinatorial Optimisation, Stirling 66-83.
• On the connectivity of random m-orientable graphs and digraphs Combinatorica 2, 347-359. [Co-author: T.I. Fenner]
• An algorithm for solving 3-dimensional assignment problems with application to scheduling a teaching practice Journal of the Operational Research Society 32, 989-995. [Co-author: J. Yadegar]
• Probabilistic analysis of some Euclidean clustering problems Discrete Applied Mathematics 2, 295-309.
• Worst-case analysis of algorithms for travelling salesman problems Methods of Operations Research 32, 93-112.
• An algorithm for algebraic assignment problems Discrete Applied Mathematics 1, 253-259.
• A partitioned inverse in linear programming Journal of the Operational Research Society 29, 383-388.
• Minimum paths in directed graphs Operational Research Quarterly 28, 339-346.
• A generalisation of the greedy algorithm for matroids Proceedings of CP77, Conference on Combinatorial Programming at Liverpool University.
• Shortest path algorithms for knapsack type problems Mathematical Programming 11, 150-157.
• Bottleneck linear programming Operational Research Quarterly 26, 871-874.
• A cost function property for plant location problems Mathematical Programming 7, 245-248.
• A bilinear programming formulation of the 3-dimensional assignment problem Mathematical Programming 7, 376-379.

graph labelling Recently Published Documents

Total documents.

• Latest Documents
• Most Cited Documents
• Contributed Authors
• Related Sources
• Related Keywords

Job Assignment Problem Using Union Group Graph Labelling

On a graph labelling conjecture involving coloured labels, modular sumset labelling of graphs.

Graph labelling is an assignment of labels or weights to the vertices and/or edges of a graph. For a ground set X of integers, a sumset labelling of a graph is an injective map f:VG→PX such that the induced function f⊕:EG→PX is defined by f+uv=fu+fv, for all uv∈EG, where fu+fv is the sumset of the set-label, the vertices u and v. In this chapter, we discuss a special type of sumset labelling of a graph, called modular sumset labelling and its variations. We also discuss some interesting characteristics and structural properties of the graphs which admit these new types of graph labellings.

Embedding rainbow trees with applications to graph labelling and decomposition

Sumset valuations of graphs and their applications.

Graph labelling is an assignment of labels to the vertices and/or edges of a graph with respect to certain restrictions and in accordance with certain predefined rules. The sumset of two non-empty sets A and B, denoted by A+B, is defined by A+B=\{a=b: a\inA, b\inB\}. Let X be a non-empty subset of the set \Z and \sP(X) be its power set. An \textit{sumset labelling} of a given graph G is an injective set-valued function f: V(G)\to\sP_0(X), which induces a function f+: E(G)\to\sP_0(X) defined by f+(uv)=f(u)+f(v), where f(u)+f(v) is the sumset of the set-labels of the vertices u and v. This chapter discusses different types of sumset labeling of graphs and their structural characterizations. The properties and characterizations of certain hypergraphs and signed graphs, which are induced by the sumset-labeling of given graphs, are also done in this chapter.

Fuzzy Magic and Bi-magic Labelling of Intuitionistic Path Graph

A graph labelling is an assignment of integers to the vertices or edges, or both, subject to certain conditions. Fuzzy labelling models precision, flexibility and compatibility to the classical models. The objective is to discuss about the magic and bi-magic labelling of fuzzy graphs. In the beginning the magic labelling of fuzzy path graph is discussed followed by the bi-magic labelling of fuzzy path graph. The next main purpose is to introduce the Intuitionistic path graph and examine the existence of fuzzy magic and bi-magic labelling.

Semi-supervised graph labelling reveals increasing partisanship in the United States Congress

Link prediction evaluation using palette weisfeiler-lehman graph labelling algorithm.

Link prediction is gaining interest in the community of machine learning due to its popularity in the applications such as in social networking and e-commerce. This paper aims to present the performance of link prediction using a set of predictive models. In link prediction modelling, feature extraction is a challenging issue and some simple heuristics such as common-neighbors and Katz index were commonly used. Here, palette weisfeiler-lehman graph labelling algorithms have been used, which has a few advantages such as it has order-preserving properties and provides better computational efficiency. Whereas, other feature extraction algorithms cannot preserve the order of the vertices in the subgraph, and also take more computational time. The features were extracted in two ways with the number of vertices in each subgraph, say K = 10 and K = 15. The extracted features were fitted to a range of classifiers. Further, the performance has been obtained on the basis of the area under the curve (AUC) measure. Comparative analysis of all the classifiers based on the AUC results has been presented to determine which predictive model provides better performance across all the networks. This leads to the conclusion that ADABoost, Bagging and Adaptive Logistic Regression performed well almost on all the network. Lastly, comparative analysis of 12 existing methods with three best predictive models has been done to show that link prediction with predictive models performs well across different kinds of networks.

GLaSS: Semi-supervised Graph Labelling with Markov Random Walks to Absorption

Graph labelling techniques, export citation format, share document.

• News & Media
• Chemical Biology
• Computational Biology
• Ecosystem Science
• Cancer Biology
• Exposure Science & Pathogen Biology
• Metabolic Inflammatory Diseases
• Advanced Metabolomics
• Mass Spectrometry-Based Measurement Technologies
• Spatial and Single-Cell Proteomics
• Structural Biology
• Biofuels & Bioproducts
• Human Microbiome
• Soil Microbiome
• Synthetic Biology
• Computational Chemistry
• Chemical Separations
• Chemical Physics
• Atmospheric Aerosols
• Human-Earth System Interactions
• Modeling Earth Systems
• Coastal Science
• Plant Science
• Subsurface Science
• Terrestrial Aquatics
• Materials in Extreme Environments
• Precision Materials by Design
• Science of Interfaces
• Friction Stir Welding & Processing
• Dark Matter
• Flavor Physics
• Fusion Energy Science
• Neutrino Physics
• Quantum Information Sciences
• Emergency Response
• AGM Program
• Tools and Capabilities
• Grid Architecture
• Grid Cybersecurity
• Grid Energy Storage
• Earth System Modeling
• Energy System Modeling
• Transmission
• Distribution
• Appliance and Equipment Standards
• Building Energy Codes
• Advanced Building Controls
• Advanced Lighting
• Building-Grid Integration
• Building and Grid Modeling
• Commercial Buildings
• Federal Performance Optimization
• Resilience and Security
• Grid Resilience and Decarbonization
• Building America Solution Center
• Energy Efficient Technology Integration
• Home Energy Score
• Electrochemical Energy Storage
• Flexible Loads and Generation
• Grid Integration, Controls, and Architecture
• Regulation, Policy, and Valuation
• Science Supporting Energy Storage
• Chemical Energy Storage
• Waste Processing
• Radiation Measurement
• Environmental Remediation
• Subsurface Energy Systems
• Carbon Capture
• Carbon Storage
• Carbon Utilization
• Advanced Hydrocarbon Conversion
• Fuel Cycle Research
• Advanced Reactors
• Reactor Operations
• Reactor Licensing
• Solar Energy
• Wind Resource Characterization
• Wildlife and Wind
• Community Values and Ocean Co-Use
• Wind Systems Integration
• Wind Data Management
• Distributed Wind
• Energy Equity & Health
• Environmental Monitoring for Marine Energy
• Marine Biofouling and Corrosion
• Marine Energy Resource Characterization
• Testing for Marine Energy
• The Blue Economy
• Environmental Performance of Hydropower
• Hydropower Cybersecurity and Digitalization
• Hydropower and the Electric Grid
• Materials Science for Hydropower
• Pumped Storage Hydropower
• Water + Hydropower Planning
• Grid Integration of Renewable Energy
• Geothermal Energy
• Algal Biofuels
• Aviation Biofuels
• Waste-to-Energy and Products
• Hydrogen & Fuel Cells
• Emission Control
• Energy-Efficient Mobility Systems
• Lightweight Materials
• Vehicle Electrification
• Vehicle Grid Integration
• Contraband Detection
• Pathogen Science & Detection
• Explosives Detection
• Threat-Agnostic Biodefense
• Discovery and Insight
• Proactive Defense
• Trusted Systems
• Nuclear Material Science
• Radiological & Nuclear Detection
• Nuclear Forensics
• Ultra-Sensitive Nuclear Measurements
• Nuclear Explosion Monitoring
• Global Nuclear & Radiological Security
• Disaster Recovery
• Global Collaborations
• Legislative and Regulatory Analysis
• Technical Training
• Additive Manufacturing
• Deployed Technologies
• Rapid Prototyping
• Systems Engineering
• 5G Security
• RF Signal Detection & Exploitation
• Climate Security
• Internet of Things
• Maritime Security
• Artificial Intelligence
• Graph and Data Analytics
• Software Engineering
• Computational Mathematics & Statistics
• High-Performance Computing
• Visual Analytics
• Lab Objectives
• Publications & Reports
• Featured Research
• Diversity, Equity, Inclusion & Accessibility
• Lab Leadership
• Lab Fellows
• Staff Accomplishments
• Undergraduate Students
• Graduate Students
• Post-graduate Students
• University Faculty
• University Partnerships
• K-12 Educators and Students
• STEM Workforce Development
• STEM Outreach
• Meet the Team
• Internships
• Regional Impact
• Philanthropy
• Volunteering
• Available Technologies
• Industry Partnerships
• Licensing & Technology Transfer
• Entrepreneurial Leave
• Atmospheric Radiation Measurement User Facility
• Electricity Infrastructure Operations Center
• Energy Sciences Center
• Environmental Molecular Sciences Laboratory
• Grid Storage Launchpad
• Institute for Integrated Catalysis
• Interdiction Technology and Integration Laboratory
• PNNL Portland Research Center
• PNNL Seattle Research Center
• PNNL-Sequim (Marine and Coastal Research)
• Radiochemical Processing Laboratory
• Shallow Underground Laboratory

Hybrid Attack Graph Generation with Graph Convolutional Deep-Q Learning

Published: April 13, 2024

Research topics

Parkinson's Disease: New theory on the disease's origins and spread

The nose or the gut? For the past two decades, the scientific community has debated the wellspring of the toxic proteins at the source of Parkinson's disease. In 2003, a German pathologist, Heiko Braak, MD, first proposed that the disease begins outside the brain. More recently, Per Borghammer, MD, with Aarhus University Hospital in Denmark, and his colleagues argue that the disease is the result of processes that start in either the brain's smell center (brain-first) or the body's intestinal tract (body-first).

A new hypothesis paper appearing in the Journal of Parkinson's Disease on World Parkinson's Day unites the brain- and body-first models with some of the likely causes of the disease-environmental toxicants that are either inhaled or ingested. The authors of the new study, who include Borghammer, argue that inhalation of certain pesticides, common dry cleaning chemicals, and air pollution predispose to a brain-first model of the disease. Other ingested toxicants, such as tainted food and contaminated drinking water, lead to body-first model of the disease.

"In both the brain-first and body-first scenarios the pathology arises in structures in the body closely connected to the outside world," said Ray Dorsey, MD, a professor of Neurology at the University of Rochester Medical Center and co-author of the piece. "Here we propose that Parkinson's is a systemic disease and that its initial roots likely begin in the nose and in the gut and are tied to environmental factors increasingly recognized as major contributors, if not causes, of the disease. This further reinforces the idea that Parkinson's, the world's fastest growing brain disease, may be fueled by toxicants and is therefore largely preventable."

Different pathways to the brain, different forms of disease

A misfolded protein called alpha-synuclein has been in scientists' sights for the last 25 years as one of the driving forces behind Parkinson's. Over time, the protein accumulates in the brain in clumps, called Lewy bodies, and causes progressive dysfunction and death of many types of nerve cells, including those in the dopamine-producing regions of the brain that control motor function. When first proposed, Braak thought that an unidentified pathogen, such as a virus, may be responsible for the disease.

The new piece argues that toxins encountered in the environment, specifically the dry cleaning and degreasing chemicals trichloroethylene (TCE) and perchloroethylene (PCE), the weed killer paraquat, and air pollution, could be common causes for the formation of toxic alpha-synuclein. TCE and PCE contaminates thousands of former industrial, commercial, and military sites, most notably the Marine Corps base Camp Lejeune, and paraquat is one of the most widely used herbicides in the US, despite being banned for safety concerns in more than 30 countries, including the European Union and China. Air pollution was at toxic levels in nineteenth century London when James Parkinson, whose 269th birthday is celebrated today, first described the condition.

The nose and the gut are lined with a soft permeable tissue, and both have well established connections to the brain. In the brain-first model, the chemicals are inhaled and may enter the brain via the nerve responsible for smell. From the brain's smell center, alpha-synuclein spreads to other parts of the brain principally on one side, including regions with concentrations of dopamine-producing neurons. The death of these cells is a hallmark of Parkinson's disease. The disease may cause asymmetric tremor and slowness in movement and, a slower rate of progression after diagnosis, and only much later, significant cognitive impairment or dementia.

When ingested, the chemicals pass through the lining of the gastrointestinal tract. Initial alpha-synuclein pathology may begin in the gut's own nervous system from where it can spread to both sides of the brain and spinal cord. This body-first pathway is often associated with Lewy body dementia, a disease in the same family as Parkinson's, which is characterized by early constipation and sleep disturbance, followed by more symmetric slowing in movements and earlier dementia, as the disease spreads through both brain hemispheres.

New models to understand and study brain diseases

"These environmental toxicants are widespread and not everyone has Parkinson's disease," said Dorsey. "The timing, dose, and duration of exposure and interactions with genetic and other environmental factors are probably key to determining who ultimately develops Parkinson's. In most instances, these exposures likely occurred years or decades before symptoms develop."

Pointing to a growing body of research linking environmental exposure to Parkinson's disease, the authors believe the new models may enable the scientific community to connect specific exposures to specific forms of the disease. This effort will be aided by increasing public awareness of the adverse health effects of many chemicals in our environment. The authors conclude that their hypothesis "may explain many of the mysteries of Parkinson's disease and open the door toward the ultimate goal-prevention."

In addition to Parkinson's, these models of environmental exposure may advance understanding of how toxicants contribute to other brain disorders, including autism in children, ALS in adults, and Alzheimer's in seniors. Dorsey and his colleagues at the University of Rochester have organized a symposium on the Brain and the Environment in Washington, DC, on May 20 that will examine the role toxicants in our food, water, and air are playing in all these brain diseases.

Additional authors of the hypothesis paper include Briana De Miranda, PhD, with the University of Alabama at Birmingham, and Jacob Horsager, MD, PhD, with Aarhus University Hospital in Denmark.

• Parkinson's Research
• Chronic Illness
• Brain Tumor
• Diseases and Conditions
• Parkinson's
• Disorders and Syndromes
• Brain-Computer Interfaces
• Parkinson's disease
• Deep brain stimulation
• Homosexuality
• Dopamine hypothesis of schizophrenia
• Excitotoxicity and cell damage

Story Source:

Materials provided by University of Rochester Medical Center . Original written by Mark Michaud. Note: Content may be edited for style and length.

Journal Reference :

• E. Ray Dorsey, Briana R. De Miranda, Jacob Horsager, Per Borghammer. The Body, the Brain, the Environment, and Parkinson’s Disease . Journal of Parkinson's Disease , 2024; 1 DOI: 10.3233/JPD-240019

Cite This Page :

Explore More

• Genes for Strong Muscles: Healthy Long Life
• Brightest Gamma-Ray Burst
• Stellar Winds of Three Sun-Like Stars Detected
• Fences Causing Genetic Problems for Mammals
• Ozone Removes Mating Barriers Between Fly ...
• Parkinson's: New Theory On Origins and Spread
• Clash of Stars Solves Stellar Mystery
• Secure Quantum Computing at Home
• Ocean Currents: Collapse of Antarctic Ice ...
• Pacific Cities Much Older Than Previously ...

Trending Topics

Strange & offbeat.