Abstract
Community detection in graph networks is a fundamental yet challenging task due to limitations in existing methods. Traditional Graph Convolutional Network (GCN)-based models often suffer from over-smoothing, which results in indistinguishable node representations after excessive aggregation. Additionally, many models face computational inefficiency, restricting their scalability on large-scale networks. To address these challenges, we propose GATFELPA, a hybrid community detection model that combines a Graph Attention Network (GAT) with an enhanced label propagation algorithm (DPCELPA). GATFELPA employs an adaptive strategy to dynamically determine the optimal number of aggregation layers, effectively mitigating over-smoothing by balancing intra-community compactness and inter-community separability. A novel similarity preservation module further enhances the model's ability to differentiate communities by retaining local and global dissimilarities in heterogeneous networks. Comparative experiments on four real-world datasets, including large-scale networks like ogbn-arxiv, demonstrate that GATFELPA achieves superior performance across most metrics, particularly excelling in accuracy, robustness, and scalability. These results highlight GATFELPA as a promising approach for addressing complex community detection tasks.