Algorithm Of Constrained Subspace Clustering

With the rapid development of science and technology, we can get more and more data and the dimension of data is more and more higher. How to mine useful information from complex and diverse data is becoming more important. Cluster analysis as an effective tool for data analysis has become a hot topic. On one hand, it is often difficult to obtain good clustering results in high-dimensional space. On the other hand, a prior knowledge is beneficial information to improve the clustering effect. Thus looking for high dimensional data clustering algorithms and fully utilizing the existing background knowledge has become an important topic clustering field. In this thesis, from two perspectives the high-dimensional of data clustering and prior knowledge proposed two improved algorithms. The main content and innovation are as follows:1. For cop-kmeans clustering ignores the local structure information of samples, resulting in clustering results is not very good, so introducing local structural information on constrained sample spread labels. Constrained k-means clustering (Neighborhood Label Propagation Constrained K-means Clustering, NLPCC) based label propagation neighborhood is proposed. To some extent, this algorithm improves the accuracy of the clustering. And through experiments demonstrate the effectiveness of the proposed algorithm.2. For dimensionality reduction and clustering of mutual independence, leading to the clustering effect is not very good, so proposed an adaptive neighborhood embedding dimension reduction clustering algorithm LDA-CNPkm. The algorithm first initializes projection direction of the sample projection by neighborhood-based label propagating constrained clustering algorithm to cluster, get a sample label, and then use the tag information of samples to solve linear discriminant direction, so as to update the projection direction, the two alternately, until clustering results change very little. To some extent, this algorithm improves the accuracy of the clustering. And through experiments demonstrate the effectiveness of the proposed algorithm.