Research On Missing View Completion For Multi-view Data

ABSTRACT:With the rapid development of information technology, massive amounts of multi-view data are constantly emerging in our daily life. To cope with such situation, multi-view learning has received much attention in the field of machine learning to promote the ability of data understanding. However, due to the difficulty, high cost, and equipment failure in multi-view data collection, we often confront with an obstacle that part or all of observed values from one view can’t be available, which makes some traditional multi-view learning algorithms won’t work well as expected. Meanwhile, the multi-view missing data will not only increase the difficulty of data mining, but also affects the result of data analysis. Thus, it becomes especially important that how to effectively complete the multi-view missing data, and further enhance the depth analysis and understanding of object that is depicted based on multiple views. It is also valuable research topic in the current field of multi-view data analysis.In order to eliminate the effects of multi-view missing data, this thesis focuses on performing missing view completion by excavating the view compatibility and view semantics consistency of multi-view data. The main contributions of this dissertation are as follows:(1) Confronting with the problem that all attributes of the view data are missing, the traditional single-view missing completion method does not considerate the complementarity of multi-view data. Thus, we propose a method of view missing completion based on kernel ridge regression for multi-view data. The method establishes a nonlinear relationship between the views, and completes missing view;(2) We focused on the missing view completion for multi-view data and propose a view compatibility based completion method. With the multiple shared subspaces by supervised learning, the view compatibility discrimination model is established. Meanwhile, assuming that the reconstruction error of each of view of multi-view data in the shared subspace takes the independent identical distribution, the preliminary completion of missing view can be performed. Furthermore, the multiple linear regression technique is implemented to obtain a more accurate completion. In addition, we also extend the proposed missing view completion method to deal with the denoising of noise-polluted multi-view data(3) For the problem of the data missing completion in the case of multi-view data disappear in pair, we propose a multi-view factor analysis method. The association of the semantics consistency between views is established by excavating shared latent factor. With the multi-view factor analysis model by supervised learning, the view semantics consistency discrimination criterion is established. Thus, the view missing completion can be effectively performed.