Deep k-Means Clustering Based on Graph Neural Networks: Leveraging Cohesion and Separation in Graph Nodes

In graph node clustering utilizing Graph Neural Networks (GNNs), nodes are typically embedded into low-dimensional vectors by GNN s, followed by the application of clustering algorithms such as k-means. However, a limitation of these two steps operating independently is that GNN embeddings may not consider the ultimate clustering objective. To address this issue, “Deep Node Clustering” has been introduced. This aims to jointly optimize node embedding and clustering to enhance overall clustering quality. In this paper, we introduce a novel algorithm, “Deep k-means Node Clustering,” as part of the Deep Clustering approach. This method emphasizes within-cluster cohesion and between-cluster separation, proposing an extended loss function that combines GNN loss and k-means clustering loss. The extended loss function includes the sum of squared errors (SSE) for cohesion and a modified silhouette score for separation. The proposed approach iteratively retrains the GNN using the extended loss function to enhance clustering results. Additionally, performance evaluation on real graph data demonstrates the effectiveness of this method in improving node clustering.