Embedding-Assisted Auto-Tuned Community Detection

State-of-the-art techniques for community detection in graphs either use feature-based clustering algorithms or structure-driven spectral clustering approaches. The former takes into consideration only the semantic relations based on node features, and the latter uses structural information for an analysis in eigenspace. Hence, neither of these two approaches can provide a very accurate detection of communities. The existing node2vec algorithm translates structural information into an embedding space. We propose a novel approach that combines all of these and obtains encouraging results in large graphs. The proposed approach in this work brings in few innovations in the form of an auto-tuning framework to address the existing shortcomings of these methods. It stabilizes an important hyper-parameter of spectral clustering and brings in a step-heuristic along with a proposed BinKMeans (binary search-driven k-means) algorithm to better detect a distinct drop in eigenvalue while using spectral clustering. Additionally, the pipeline implements NodeFeat2Vec to bring in feature information in node2Vec. An in-depth evaluation with real-world graph data showed that this proposed approach obtained superior results when evaluated using appropriate metrics on communities in labelled real-world graphs.