Study On Algorithms For Analyzing Multi-View Data

In many applications, data can be represented in multiple ways and this kind of data is the so-called multi-view data. Using the complementary information appropriately to analyze multi-view data more effectively is becoming increasingly important in these applications. In this thesis, we study algorithms for analyzing multi-view data from the aspects of clustering and feature selection.As for clustering, we adopt the idea of co-training and require the clustering results on each view to be consensus. We generalize the single-view Spectral Embedded Clustering algorithm to multi-view scenario and propose the Co-trained Multi-View Spectral Embedded Clustering(CoSEC) algorithm, to enhance spectral clustering's ability of handling multi-view data without a clear manifold structure. Experimental results verify the effectiveness of CoSEC.To reduce the dimensionality of multi-view data by feature selection, we employ the one-by-one feature selection strategy and simultaneous feature selection respectively.For the first strategy we proposed the Discriminative Feature Selection(DFS) algorithm which combines the classical Linear Discriminate Analysis(LDA) and the 2,1-norm sparsity regularization technique to select the most discriminative features and remove the redundancy among the selected features. Under the second strategy, we aim to use the domain knowledge in the form of pairwise constraints for feature selection in cases where no label information available and propose the Multi-View Semi-Supervised Feature Selection(MVSSFS) algorithm. We solve both algorithms by alternative iteration and prove them to be convergent theoretically. The effectiveness of the above two algorithms is verified by experiments.