Deep manifold regularized semi-nonnegative matrix factorization for Multi-view Clustering

With the development and application of modern technology, multi-view data exists widely. The deep matrix factorization model is widely used because it mines the hierarchical information of the samples, but it lacks a geometric structure protection mechanism. Moreover, in multi-view representation learning, strong representations have consensus from multi-view data and are more likely to exist in the structural relationships of data objects. To address these issues, we propose Deep Manifold Regularized Semi-Nonnegative Matrix Factorization for Multi-view Clustering (DMRMF_MVC), which preserves the geometric structure of data objects through multi-layer embedding graph regularization. The graph regularizer combines the nearest neighbor relationship with the exclusive relationship of data objects to express the internal manifold structure of the data fully. Furthermore, to distinguish the value of each view data, the adaptive weight is associated with representation learning to maximize the complementarity between the views. Extensive experimental results show that DMRMF_MVC outperforms other state-of-the-art algorithms based on multiple indicators on various datasets.(c) 2022 Elsevier B.V. All rights reserved.