Detecting communities in social networks with graph convolutional neural networks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The latent community structures in the social networks have now become a cornerstone problem in the scientific study of networks, and has extensive implications in recommendation systems, epidemiology, fraud detection, and social behavior studies. The conventional community detection algorithms, most of which are based on spectral clustering, modularity maximization, and random walk heuristic, do not scale and generalize when there are heterogeneous, multi-relational graph structures that define contemporary online social networks. Over the past few years, graph representation learning and deep neural architecture have converged, providing intriguing new directions of community discovery. The article introduces a new, end-to-end framework that combines heterogeneous node embeddings as Metapath2Vec with a multi-layer Graph Convolutional Network (GCN) with multi-head attention and adaptive neighbourhood aggregation. The embedding module by embedding semantically meaningful metapaths, Author–Paper–Author (A–P–A), Author–Paper–Conference–Paper–Author (A–P–C–P–A), and Author–Paper–Keyword–Paper–Author (A–P–K–P–A), defines the heterogeneous relational structure of academic networks. These enhanced embedding are further optimized by three stacked GCN layers with LeakyReLU activation, batch normalization and dropout regularization to avert overfitting. The output node representations are categorized through a softmax layer which gives the probability of membership to a community. Extensive experiments on the two heterogeneous network benchmarks, DBLP and ACM, prove that the proposed method is accurate (87.1) and has an F1-score of 86.0 on DBLP, and 89.3 accuracy with F1-score 88.7 on ACM, outperforming a dozen baselines, including recent 20,232,024 state-of-the-art The reasons as to why the three inductive biases provided by Metapath2Vec, GCN aggregation, and multi-head attention are complementary to each other are supported by a theoretical analysis. Ablation experiments verify the value of each of the components, a case study demonstrates how the model solves the ambiguous boundaries of communities and sensitivity analysis confirms the strength of the hyperparameter selections.