Architecture Alternative Deep Multi-View Clustering

Ke Ping, Shuxiao Li, Chuhan Wu, Zhenwen Ren

Abstract: Due to the strong non-linear fitting ability of deep neural networks, deep multi-view clustering has become a popular topic in the fields of signal processing and machine learning. Multi-view clustering based on deep auto-encoder can effectively capture the nonlinear features of high-dimensional data. However, the existing methods still have the following problems: 1) most autoencoder-based deep multi-view clustering methods ignore the differences between cross-view data and lose view-diversity features by using the same encoder structure for different views, and 2) many current deep multi-view techniques rely on single-lane neural networks for extracting feature data from each view. The current approach has limitations in its capability to accurately analyze comprehensive complementary information and multilevel features. To address these issues, we introduce a new clustering method Architecture Alternative Deep Multi-view Clustering (AADMC). Specifically, AADMC proposes a dynamic encoder network to adapt the encoder structure of each view according to the diversities of different views. Subsequently, AADMC proposes utilizing the Hilbert-Schmidt Independent Criterion (HSIC) to analyze the diversity of information between each encoder output. Moreover, AADMC integrates high-order and low-order information of the data into a shared connection matrix. To be more specific, the low-rank constraint is employed in order to effectively investigate and utilize the consensus information derived from all available views. The effectiveness and superiority of AADMC are demonstrated through experimental results conducted on various public datasets.