Self-weighted Contrastive Fusion for Deep Multi-view scRNA-Seq Clustering

Single-cell RNA sequencing data clustering is inherently challenged by ultra-high dimensionality, extreme sparsity, and technical noise, collectively presenting formidable computational burdens. Compounding this difficulty, the popular GNN-based clustering methods designed to address these issues are fundamentally limited by single-scale topology and the over-smoothing effect, which significantly hinders robust cell subtype identification in large datasets. To overcome these issues, we propose Self-Weighted Contrastive Learning (SCSW), a novel framework utilizing decoupled multi-view contrastive learning. SCSW constructs complementary dual-graph representations: the KNN graph captures local neighborhood structure, while the Diffusion Map graph models global manifold topology and long-range relationships. Our innovative decoupled encoder architecture separates representation learning from contrastive optimization, which bypasses GNN over-smoothing and integrates a cross-view contrastive mechanism to ensure feature discriminative consistency. The final robust embeddings are generated by a self-adaptive weighted fusion strategy that dynamically down-weights noisy or unreliable views. Extensive experiments on nine diverse scRNA-seq datasets confirm that SCSW consistently and significantly outperforms all tested on critical clustering metrics.