Community detection in complex networks via semi-supervised learning: application of link prediction

Community detection is a fundamental problem in network learning. Existing semi-supervised approaches often underutilize labeled data and semantic structures, and their performance drops significantly when labeled samples are scarce. Many methods rely on known samples to learn community structures, but structural differences can cause mismatches with target communities, reducing detection quality. Ignoring community structure while neglecting user feature information remains a major challenge for improving community detection accuracy. Motivated by the need to effectively integrate network structure and node features, this study proposes a semi-supervised and semi-local community detection model based on a graph convolutional neural network (SLC-GCN). Our approach extracts a semi-local subgraph around the target node and leverages this subgraph to identify communities with similar structural patterns. SLC-GCN simultaneously utilizes both topological structure and user feature information and employs an ensemble clustering approach to integrate the results into a coherent community structure. Furthermore, we extend the identified communities for link prediction in multilayer networks. For this purpose, we introduce a novel similarity metric based on local random walks and reliable weighted paths, which effectively addresses the link prediction problem in multilayer networks while incorporating community information. The effectiveness of SLC-GCN is evaluated on both real and synthetic multilayer networks for the joint community detection and link prediction task. Comparative evaluations reveal that SLC-GCN achieves superior results relative to current leading methods on a wide range of performance metrics.