Research Of Hybrid Manifold Clustering Algorithms Based On Spectral Clustering

With the development of information and technology, we have witnessed an explosion in the availability of data from multiple modalities. Facing a lot of high-dimension data, we can effectively deal with and find the useful information we really need, which is a mutual problem in some fields, such as the pattern recognition, the computer vision. Manifold learning is a kind of effective tool to cope with data, which dig some useful information out original information.For the multi manifold data, due to the complexity of the distribution of data, especially different manifolds of data cross stack, we traditionally considers the data link between points, does not take into account the overlapping data points because the Euclidean distance near, which make originally the two clusters of data points is partitioned into a cluster. Therefore, basing on this, we consider the other characteristics of data points. For example:the data points of the local tangent space.To the traditional K-means clustering, data sets must be n dimensional vector. Therefore, the calculation speed is slow and consume a lot of memory. Spectral clustering is based on solving this problem, which is used to reduce the dimension of high dimensional data sets, and to approximate the original data with high dimension by using some feature vectors of it. Therefore, spectral clustering has a fast calculation speed, and it is not easy to be affected by some noise and boundary, and it has strong robustness. In the article, from similarity matrix and choose of nearest neighbor, we put forward some different manifold clustering algorithm to effectively solve the hybrid manifold clustering problem. In a nutshell, in the article, we not only study manifold learning methods, but put these methods into some practical problems of clustering analysis. In addition, we not only have deep research on these algorithm theories, but also apply these algorithms to real applications, such as face recognition, image segmentation, text clustering and so on.Another aspects of this article, for some of the current algorithm based on spectral clustering were carefully studied, but there are still some problems not solved very well. For example:the choice of Laplacian matrix form, the number of manifold clustering of automatic selection, the selection of feature vectors, and so on. To solve these problems still need our further research and exploration.