Partial Multi-View Clustering Based on Sparse Embedding Framework

Abstract

In this paper, we propose a novel partial multi-view clustering method based on the sparse embedding framework, which can handle incomplete view data well and obtain good clustering performance. Most real-world datasets are often comprised of different views, which provide complementary information for each other for clustering or classification tasks. Existing multi-view clustering methods assume that each example appears in all views or all examples contained at least one view. But each view of real data is often subject to various degrees of damage resulting in partial examples. There are some works on partial view clustering, in which almost all are based on standard non-negative matrix factorization (NMF) to solve the problem that did not consider the sparseness degree of coefficient and unsatisfactory basic matrices learned by NMF. Besides, most existing clustering methods only carry out dimensionality reduction before learning the models, which is unable to make full use of the discriminative information in raw data. Thus, partial multi-view clustering based on sparse embedding framework is proposed, in which dimensionality reduction and dictionary learning are taken into consideration simultaneously to learn better sparse coefficients and dictionaries. The proposed method preserves as much useful information as possible in original space by constraining projection matrix to be orthogonal and imposes Fisher discrimination analysis on dictionaries rather than sparse coefficients, which makes the learned dictionaries more discriminating and promotes sparsity of coefficients. Finally, the synthetic dataset, the extended Yale B dataset, the MNISIT Handwritten Digit dataset and the large dataset-Caltech101 are employed to perform clustering tasks, and the experimental results show that the proposed method achieved a better clustering performance than the other state-of-the-art clustering methods.