Multiscale Graph-Guided Hybrid Contrastive Learning for Multiview Clustering

Abstract

Multiview clustering (MVC) is crucial for integrating complementary information, making it particularly well-suited for Earth observation due to the multiscale and multispectral characteristics of remote sensing imagery. However, most existing graph-based MVC approaches rely on static graphs constructed solely from final-layer embeddings, overlooking the rich hierarchical semantics encoded across intermediate network layers. Consequently, this single-scale paradigm isolates global semantics from fine-grained topologies, yielding biased graph representations that blur cluster boundaries and degrade consensus clustering. To address these issues, we propose a multiscale graph-guided hybrid contrastive (MGHC) learning framework that integrates hierarchical topological guidance with a multilevel contrastive paradigm to learn discriminative representations. In particular, MGHC extracts hierarchical representations from view-specific autoencoders to construct layer-wise consensus graphs, capturing structural dependencies at multiple granularities. Guided by these graphs, our hybrid contrastive paradigm performs alignment across three distinct levels. Graph-level alignment enforces structural consistency across different semantic layers. Instance-level alignment leverages the learned topologies and clustering predictions to mitigate false negatives and enhance cluster discrimination. Label-level alignment synchronizes soft clustering assignments through layer-specific predictors to promote hierarchical semantic consistency. Evaluations across eight general benchmarks demonstrate that MGHC predominantly outperforms state-of-the-art MVC methods, validating the strong generalization of the proposed framework. More importantly, its superior performance on the EuroSAT dataset confirms its efficacy in complex remote sensing applications. The code is released at https://github.com/JiayueLuo/MGHC