Research Of Multi-view Local Subspace Learning

With the development of Internet technology, subspace learning has become a hot issue in machine learning, data mining and pattern recognition. High dimensional of data decreases the performance of model. Subspace learning aims to compute the relationship in high dimensional space and maintains the relationship in low-dimensional subspace.Feature of data has become more abundant because new collection methods and feature extraction methods were proposed. Different feature has different performance in machine learning and data mining model. It is hard to get good performance in complicated dataset with single feature. Multi-view learning is a new technology which can utilize the feature in different views simultaneously to improve the performance of model. This paper proposed a multi-view local subspace learning to improve the performance of subspace learning. The main contributions of this paper are summarized as follows:(1) This paper summarized the background, meaning and research status at home and abroad of multi-view learning. This paper introduced the core idea can research status of three kinds of multi-view learning method: co-training, multiple kernel learning and subspace learning.(2) This paper introduced the details of subspace learning and multi-view learning: locally linear embedding, Laplacian eigenmaps, co-training approach for multi-view spectral clustering and co-regularized multi-view spectral clustering.(3) Locally linear embedding is a classical subspace learning method. However the performance is easy to be affected by the noise data. In addition, the parameter is also important for algorithm. For this issue, this paper proposed a robust locally linear embedding. This paper utilized L2 norm to penalize the complexity of similarity matrix and L1 norm to guarantee the sparsity of similarity matrix. This paper took consensus principle and complementary principle into account and utilized the multi-view data to revise the influence from noise data to get a robust similarity which can improve the performance of subspace learning. This paper proposed an iteration solution for multi-view local subspace learning and utilized the auxiliary function to prove the correctness and convergence of solution. This paper compared the multi-view local subspace learning with single method, feature concatenation, multiple kernel and multi-view methods on real-word and synthetic datasets to show the efficacy of the proposed algorithm.(4) Image retrieval is a hot issue in academic area. Feature extraction method is important for image retrieval system. It is hard to get good performance with single feature because the complexity and diversity of image on Internet. This paper utilized multi-view subspace learning method to get a robust feature and applied it in image retrieval system. The results show the efficacy of our method.