A Study Of Multi-view Learning And Improved Algorithms

With the development of data acquisition techniques, the acquired data often has multiple perspectives, thus forming multi-view data, and it is a challenging task to effectively utilize multi-view data for learning. In this paper, transfer learning, data dimensionality reduction and data clustering problem on multi-view data are studied, and several improved methods are proposed.Firstly, transfer learning algorithm with multiple views and multiple sources are studied, and a multi-view transfer learning algorithm based on view consistency is proposed. The algorithm uses gaussian mixture model to estimate density ratio between source and target domains instead of gaussian process, and it describes the distribution difference between source and target domains more accurately. Simultaneously, the algorithm selects some source domain samples more consistent in labels by adding the difference item of prediction in the sample selection formula, and introduces the label consistency information of multi-view samples. Therefore, the improved algorithm improves classification performance of multi-view data.Secondly, the multi-view semi-supervised dimensionality reduction method(MVSSDR) is studied in this paper. The method is suitable for multi-view data dimensionality reduction which obtains consensus low-dimensional representation by minimizing the difference between different views. This method requires that data of different views of a sample should be fully matched, so it cannot deal with semi-paired data dimensionality reduction problem. In this paper, some improvements based on MVSSDR are proposed, which use only a small amount of paired data to calculate the difference between different views to avoid the negative effects brought about by forced pairing. Meanwhile, in order to obtain a more separable consensus low-dimensional representation, we make transfer matris of each view sparse, and propose the sparse multi-view dimensionality reduction method based on dictionary learning. This method improves classification performance of the low-dimensional representation obtained by the multi-view dimensionality reduction method.Finally, the spectral clustering algorithm based on Markov chain is studied and its existing problems are analyzed. According to the problem that this algorithm cannot handle multi-view clustering, shared latent similarity matrix of multi-view data are obtained by maximizing the similarity between different views. At the same time, the traditional Euclidean distance measure cannot properly reflect the similarity between different views, and we use the angular based similarity measure. In addition, multi-view data may be corrupted by noise, which will significantly degrade multi-view clustering results if not addressed. To this end, the similarity matrix of each view is decomposed into a shared latent similarity matrix and a deviation error matrix, and sparse constraints are imposed on the offset error matrix. This method takes advantage of the complementary information of multi-view data more effectively, and improves clustering performance.For the improved algorithms proposed in this paper, comparative experiments have been done and verfy the validity of the improved algorithms.