Community Detection Method Based on GNN and Improved K-means

Abstract

As the scale of networks continues to expand, their structures become increasingly complex, and the diversity of node information within these networks intensifies. Traditional community detection methods are no longer capable of meeting the growing demands. Thus, the design of a powerful community detection method with low complexity and cost is an urgent issue to be addressed. This paper proposes a community detection method based on Graph Neural Networks (GNN) and an improved K-means algorithm. Initially, the original adjacency matrix is reconstructed based on the importance of nodes, resulting in a new node similarity matrix. Subsequently, a graph autoencoder is employed to extract structural information from the network graph. The decoder part utilizes the form of vector inner product to restore the similarity matrix using the structural information from the hidden layer. The features extracted by the graph autoencoder are then utilized by the improved K-means algorithm to ultimately achieve community division. Finally, this paper summarizes the shortcomings of community detection and provides a prospect for future research directions.