A fusion of centrality and correlation for feature selection

Abstract

The rapid development of computer and database technologies has led to the high growth of large-scale datasets. This produces an important issue for data mining applications called the curse of dimensionality, where the number of features is much higher than the number of patterns. One of the dimensionality reduction approaches is feature selection, which can increase the accuracy of these applications and reduce their computational complexity. This paper proposes a novel feature selection method to reduce the dimensionality and computational complexity in high-dimensional data processing. First, based on the centrality and Fisher score, a probabilistic strategy metric is proposed to measure the influence of features. Second, a new discriminant function is proposed to determine whether one feature should be selected. It can automatically calculate weight parameters to balance the relevance of a feature to class labels and the redundancy of the selected feature subset. Finally, a new method is proposed by combining the maximum information coefficient (MIC), total information (TI) and centrality technique, named MTC_FS. The experimental results show that the average accuracy of the MTC_FS improves by 1.8% compared with the best baseline, and the comprehensive performance of the MTC_FS is superior to all baselines on 12 public datasets. MTC_FS has a shorter runtime on all datasets than all baselines. In addition, the performance of the MTC_FS method is the most stable on the NBayes classifier.