Research On Feature Learning In Clustering

We have to discern different kinds of objects and analyze similarity, if we want to know our real-world well. Clustering is a form of unsupervised classification according to similarity of objects, and it is also a analysis tool under without any hypothesis. Clustering has been used widely in different subjects. Considering the importance of feature selection and feature weighting, feature selection and feature learning algorithms available working in supervise classification mainly, the feature learning clustering has been studied and accomplished this contributions as follows.(1) Research and analyze the problems and state of clustering algorithms. Illuminate clustering algorithms based on division and the methods of feature learning.(2) Because clustering algorithm are sensitive to the initial centers of clusters, which affects feature learning and clustering quality, a initial clustering center method (the longest distance summation) is produced. This method can separate initial clustering centers in different clusters, combine to many division clustering algorithms better.(3) In order to embody different contribution of each feature, we research and analyze feature criterion and its weakness based on Relief algorithms. This efforts builds the foundation of the feature criterion function and point out the research direction. With lower complexity and easier understanding, the feature criterion function can combine many clustering algorithms in practical application.(4) In the dissertation, based on the novel feature criterion function, we apply the function into feature learning clustering to counteract the negative affects by given feature weights wrongly. And we extend the feature learning clustering algorithm into database with categorical attributions. Some experiments are conduced and compare the results to that of tradition clustering algorithms. These test experiments demonstrate the novel feature learning clustering algorithm is effective and feasibility in promoting clustering precision and feature learning.