Multi-View Clustering via Block-Diagonal Low-Rank Graph Filtering

Authors:

Abstract
In recent years, graph-filter-based multi-view clustering methods have attracted increasing attention due to their robust feature learning capability. However, most existing approaches directly impose manifold smoothness on noisy initial graphs, which leads to insufficient feature discriminability. Moreover, graph filtering and clustering are typically performed independently, making it difficult for predefined filters to adapt to specific clustering objectives. To address these issues, we propose a Multi-View Clustering via Block-Diagonal Low-Rank Graph Filtering framework (BDLR‐GF). The core idea is to construct an ideal affinity matrix with a block-diagonal low-rank structure via the outer product of spectral embeddings, which replaces the original adjacency matrix directly used in conventional manifold regularization, thereby enabling the graph filter to learn from more reliable geometric constraints. BDLR‐GF jointly optimizes consensus graph construction and adaptive graph filtering by exploiting this purified affinity matrix, forming a closed-loop framework that dynamically adapts the filter to clustering needs and ultimately produces a consensus graph highly aligned with the clustering objective. Extensive experiments on multiple benchmark datasets demonstrate that BDLR‐GF outperforms state-of-the-art methods in terms of clustering accuracy and other metrics, verifying its superiority.

This publication has 29 references indexed in Scilit: