Self-Paced Encoding with Adaptive Graph Regularization for Multi-view Clustering

ICLR 2026 Conference Submission17343 Authors

Keywords: multi-view clustering

Abstract: Multi-view graph clustering is a powerful technique for learning discriminative node representations by integrating complementary information from diverse views. However, existing methods often suffer from rigid fusion schemes, ignore sample difficulty during training, and struggle to capture both global semantics and local structures through graph-based regularization. To address these issues, we propose SPEAG, a novel framework for Self-Paced Encoding with Adaptive Graph Regularization. SPEAG combines view-specific graph autoencoders with a unified learning objective that incorporates self-paced training, adaptive view fusion, and structure-aware regularization. Specifically, a self-paced neighborhood expansion strategy is introduced, where the $k$-nearest neighbor graph is gradually densified to learn from easy instances first and hard ones later. Meanwhile, each view’s embedding is adaptively weighted based on its importance, and a fusion representation is formed for global consistency. To encourage distributional alignment and enhance cluster compactness, SPEAG integrates a Maximum Mean Discrepancy (MMD) loss across views and a self-supervised clustering objective based on soft assignment refinement. Extensive experiments on real-world datasets demonstrate that SPEAG achieves superior clustering accuracy and robustness compared to existing multi-view graph clustering methods.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17343