Research On Feature Selection Based On Label Structure

Feature selection is an effective way to solve "Curse of dimensionality" in data mining and machine learning.The quality of feature selection is highly correlated with classification accuracy and generalization performance.Research on efficient feature selection algorithm is very important for clustering and classification of high-dimensional data.This paper focuses on the feature selection method of embedded tags and the feature selection method under the hierarchical class structure.The specific research results are mainly embodied in the following two aspects:(1)We propose a label feature selection algorithm based on sparse clustering.Compared with the existing unsupervised feature selection algorithm,it maps the high dimensional data to the low dimensional space.Then,it obtains an embedded label and uses the low dimensional space to fit the high dimensional space.Finally,we construct the dimensionality reduction target function and propose a feature selection algorithm based on the score of each feature by the norm sparse regression.Experiments show that the proposed algorithm achieves good results under two indexes of clustering accuracy and mutual information.(2)We propose a feature selection algorithm based on label hierarchy.It deal with the problem that most existing feature selection methods ignore the hierarchical structure among categories.First,the weight of the category is calculated by orthogonal transformation.Second,the label of the sample is predicted according to the weight selection characteristics.Experiments show that the proposed algorithm achieves better experimental results in terms of classification accuracy.