Visual Techniques for Clustering Tendency Assessment of Networks

Graph datasets are gaining popularity in various applications. Analyzing graph data using data clustering could help reveal underlying patterns and understand them better. However, most popular graph clustering algorithms require an estimate of the number of clusters or some other metadata to obtain information about the community structure, a difficult ask for novel datasets. This paper introduces two novel approaches for community clustering of generic graph networks using the VAT family of algorithms, even without prior knowledge about the community structure. Two proposed approaches: Spec-GVAT and Jacc-GVAT, reconstructs the community structure in graphs effectively with a satisfactory run time and high level of accuracy. Spec-GVAT is spectral graph theory based, while Jacc-GVAT incorporates a pairwise dissimilarity score based on Jaccard similarity between two sets. Numerical experiments performed on five benchmark datasets: Zachary karate club network, Girvan-Newman graph, planted l-partition graph, American college football network, and Bottlenose dolphins network, validate the supremacy of proposed approaches compared to conventional methods.