Guided Sparse Mixture-of-Experts for Multi-View Clustering in Hyperspherical Prototype Space
Abstract
Multi-view clustering (MVC) aims to discover cluster structure from multiple views by exploiting shared semantics while preserving complementary information. Recent mixture-of-experts (MoE) based MVC methods have demonstrated the value of shared sparse experts for collaborative representation learning, yet they still leave sparse expert learning weakly coupled to clustering. In particular, when clustering is performed in a post-hoc manner on learned embeddings, the clustering module acts only as a readout and therefore cannot provide in-training guidance to routed features. To address this limitation, we propose GSMVC, a framework that integrates guided sparse expert learning, shared hyperspherical prototype clustering, and LLM-based regime calibration. Specifically, we first employ a shared sparse MoE backbone to learn collaborative multi-view representations. We then introduce a shared hyperspherical prototype bank that produces per-view prototype logits and unified cluster assignments, enabling unified prototype supervision to participate in routed representation learning through prototype-space contrastive learning. Building on this backbone, we further use a large language model as a high-level controller that maps structured summaries of routing and hyperspherical prototype states to expert-prior calibration signals, instead of directly making per-sample clustering decisions. Experiments on MVC benchmarks show that GSMVC achieves better clustering performance than existing deep MVC methods.