Orthogonal Semi-Nonnegative Tensor Factorization Based Multi-View Clustering

Abstract

Non-negative matrix factorization(NMF) is a representative technique for multi-view clustering, achieving fast and impressive experimental results. However, existing NMF-based multi-view methods process each view sequentially and then fuse data from multiple views at the decision level, lacking sufficient exploration of the hidden information and leading to poor clustering performance. Besides, many graph clustering methods rely heavily on the input anchor graph. To address these issues, we propose Orthogonal Semi-Nonnegative Tensor Factorization based Multi-view Clustering (OSNTF). First, we extend the orthogonal semi-non-negative matrix factorization to the orthogonal semi-non-negative tensor factorization, and give a unified framework for multi-view scenes, which has the better discrimination ability. Second, for the multi-view clustering scene, combing the unified framework with the tensor Schatten p-norm regularization, we integrate the multi-view information both at the data level and decision level to fully exploit the complementary information between views. Finally, adaptive learning of the anchor graph can obtain the more accurate relationship between data points and anchor points, avoiding the impact of the predetermined graph. Comprehensive experiments across multiple benchmark datasets show that our proposed approach delivers effective clustering results.

Keywords: Multi-View Clustering, Graph fusion, Sparse representation