An Unsupervised Feature Selection Method Based On The Degree Of Feature Cooperation

As a branch of dimension reduction, feature selection plays an important role in improving the accuracy of machine learning and reducing the time complexity. Although it has been widely studied in supervised configuration, in unsupervised learning the task seems rather difficult due to the lack of real class labels. Most of the current feature selection methods tend to achieve the goal by means of eliminating noisy and redundant features, but the two kinds of features cannot be always removed simultaneously. The paper reinterprets the basic idea of feature selection and considers that directly selecting the useful features instead can not only remove the above two kinds of features but also model the interactions between features which induces a more explicit concept of feature selection. On the basis of this idea, the paper assumes the whole feature space could be approximately represented by any set of complementary features and then proposes an unsupervised feature selection method based on the degree of feature cooperation to select one of them in the direct manner. By defining the concept of the degree of feature cooperation, the paper describes the interaction between features and discriminates the complimentary features from the rest at the first step. Next, a framework based on this concept and the essential idea of hierarchical clustering is proposed, aiming at selecting a feature subset with maximal degree of feature cooperation and considering it as a complementary feature subset. In the following section of this paper, the analysis about the suppression of noisy and redundant features is given to demonstrate the idea in this paper is essentially consistent with the previous work. In the comparative experiments, the proposed method and some other state-of-art unsupervised feature selection methods are evaluated on nine different datasets and the result show the effectiveness of our method. Finally, the paper summarizes the full text and gives some possible directions for further improvement and some potential research topics related to this field.