Contextual Multi-View Graph Community Detection Using Graph Neural Networks

In the last years, in order to better accommodate the increase of data volume and complexity, different data types have been presented. One of the most popular abstract data types (or known as data structure) is Graph. It is a complex data type that can capture and detail complex system components and relations. Graph data structure is used in numerous industrial scenarios and domains such as social networks (Wu et al., 2020, IEEE Access, 8, 96016–96026), e-commerce (Kim et al., 2006, Expert Systems with Applications, 31, 101–107), marketing, and chemistry. Graph clustering is an analysis technique that aims to regroup a set of related vertices (nodes) in a graph. This technique plays an important role in various applications, for instance, community detection. Community detection is a graph clustering technique that enables the recognition of densely connected communities (clusters) within a graph. However, the classical methods, that consider only single-view features and neglect the context, are insufficient. Then in industrial cases, graphs often contain multiple views and the context surrounding the graph may influence the community preference for vertices. To address this challenging problem, we introduce Contextual Graph Clustering (CGC), a novel community detection approach for discovering communities within multi-view graphs with context consideration. This approach uses a graph auto-encoder to transform every view feature into a lower-space encoding. It combines the embeddings and the corresponding context using a fusion and context block. This block will apply the context directly on the embedding, then aggregate the embeddings using an aggregation function. As a last step, the aggregated embeddings will be used to detect the communities. CGC enabled the detection of various communities from a multi-view graph under a dynamic context.