SCGC-JWAF: Shallow contrastive graph clustering with Jaccard-based weighted adaptive filtering

Node clustering has gained significant attention due to its effectiveness in uncovering hidden structures in graph-structured data. However, real-world graphs often contain noise and are sparsely connected, which negatively impacts clustering accuracy. Moreover, unweighted graphs cannot reflect the varying strengths of connections between nodes, resulting in less discriminative representations and poor clustering outcomes. To address this, we propose SCGC-JWAF, a shallow contrastive node clustering model that introduces a weighted adaptive filter based on node intimacy-driven graph. First, we construct the weighted graph by combining graph’s structural data with jaccard similarity information. Then, we design a weighted adaptive filter with a new weight factor and the modified graph that incorporates high-pass and low-pass filters. By using a weighted Jaccard-based graph in the filtering process, nodes with higher Jaccard similarity contribute more to smoothing, while those with low overlap contribute less, reducing noisy influences. This filter refines nodes’ features and enhances the discriminative power of the filtered features by encoding fine-grained topological similarity. Finally, a proximity-level contrastive mechanism improves the topological structure of the graph and strengthens the discriminability of the learned graph by drawing similar nodes closer and pushing dissimilar ones apart. Extensive experiments demonstrate that our proposed shallow model achieves strong accuracy across multiple datasets–75.96% (Cora), 73.34% (Citeseer), 62.84% (Texas), 53% (Cornell), 68.69% (Pubmed), 62.57% (Wiki), and 64.14% (Wisconsin)–while outperforming 26 state-of-the-art graph clustering methods.