Advancing Community Detection with Graph Convolutional Neural Networks: Bridging Topological and Attributive Cohesion

TL;DR

This work tackles community detection in attributed networks by addressing two key drawbacks of existing GCN-based approaches: suboptimal modularity when training to maximize $Q$, and poor topological coherence when relying on attribute-driven human labels. It introduces TAS-Com, a GCN framework with a novel loss $L = L_M + \mu L_R$ that combines a Leiden-based modularity objective with a refinement of human-labeled communities to ensure connectivity. The method learns node embeddings $X^{(e)}$ and applies BIRCH clustering to obtain final communities, guided by two complementary losses: $L_M$ aligns embeddings with Leiden’s high-modularity partitions, and $L_R$ aligns embeddings with refined, connected human-labeled structures. Empirical results on six benchmark networks show TAS-Com achieves superior trade-offs between modularity and $NMI$, along with favorable conductance and F1 metrics, outperforming a broad set of baselines and demonstrating robust connectivity within communities. The approach offers a scalable, principled way to fuse topological and attributive signals for high-quality community detection with practical implications for social networks and related domains.

Abstract

Community detection, a vital technology for real-world applications, uncovers cohesive node groups (communities) by leveraging both topological and attribute similarities in social networks. However, existing Graph Convolutional Networks (GCNs) trained to maximize modularity often converge to suboptimal solutions. Additionally, directly using human-labeled communities for training can undermine topological cohesiveness by grouping disconnected nodes based solely on node attributes. We address these issues by proposing a novel Topological and Attributive Similarity-based Community detection (TAS-Com) method. TAS-Com introduces a novel loss function that exploits the highly effective and scalable Leiden algorithm to detect community structures with global optimal modularity. Leiden is further utilized to refine human-labeled communities to ensure connectivity within each community, enabling TAS-Com to detect community structures with desirable trade-offs between modularity and compliance with human labels. Experimental results on multiple benchmark networks confirm that TAS-Com can significantly outperform several state-of-the-art algorithms.

Figures (5)

• Figure 1: The overall design of the proposed TAS-Com approach.
• Figure 2: Comparison of $O_c(CS)$ achieved by TAS-Com, DGCluster, and DyFSS across all benchmark networks.
• Figure 3: Ablation study on the Cora network evaluating the impact of $L_M$ and $L_R$ in the loss function $L$. Scores ($Q$ and $N\!M\!I$) are scaled by 100. $L_M+L_R$ achieves the best total score.
• Figure 4: Sensitivity analysis of parameter $\mu$. Modularity $Q$ and $N\!M\!I$ scores are scaled by a factor of 100 for ease of reading.
• Figure 5: Graph Convolutional Neural Network architecture of TAS-Com.