The Research Of Semi-supervised Feature Selection And Stability Evaluation

Data mining is the process of acquiring knowledge of the variousmethods from the data. Increasing dimensions of datasets has brought agrowing challenge. With the increase of human knowledge, data setsbecome increasingly large, and continue to produce new types of data, forexample, the data stream, the network of genes and proteins in themicroarray, and so on. The researchers have come to realize that, featureselection is indispensable to a data mining system. Feature selection is aprocess to select a subset of features from the original features.Feature subset selection is an important approach to deal withhigh-dimensional data. But selecting the best subset of data is NP hard.So most of feature selection methods cannot handle high-dimensionaldata efficiently, or they can only obtain local optimum instead of globaloptimum. In these cases, when the data consist of both labeled andunlabeled data, semi-supervised feature selection can make full use ofdata information.This paper is to study the method of feature selection in recent yearsthe field of two hot topics, a) semi-supervised feature selection, b) stability of feature selection.In this paper, we introduce a novel semi-supervised feature selectionalgorithm, which is a filter method based on Fisher-Markov selector, thusours can achieve global optimum and computational efficiency undercertain kernelsThe stability of feature selection indicators is an often neglected, butsometimes the importance of stability as important as the classificationaccuracy.