Study On Unsupervised Feature Selection Based On Spectral Regression Algorithm

With the widely application of text, images, networks, gene databases, etc. high dimensional data rapidly appear in the world. And people find that some features are not so significant and useful in data mining process, and they may be redundant and even irrelevant. Most of time, data are preprocessed so as to improve data quality. Feature selection is one of the most common methods, and it can remove the redundant and the irrelevant features, and select the significant subset of original data. After dimensional selection, the speed and results of cluster are improved.According to the data being labeled or not, feature selection can be divided into two ways: supervised and unsupervised. Because the merit rating of featured subsets would be decayed by the correlation between the categories of samples and inner features, unsupervised feature selection raises up to be an interesting research focus. Traditional unsupervised featured selection methods apply manifold methods, so the original category data cannot connect the selected transformed subsets due to the absence of the original label information. The lack of correspondence between subsets and original data leads to worse results.This dissertation proposes a method of unsupervised spectral regression feature selection based on manifold learning and L1-norm. By Laplacian matrix, the original feature space data is generated. On the basis of the fitting process, the significant corresponding factors is obtained. Experimental results show that the method based on the spectral regression obtains a good effect on saving featured category data.