A new feature selection algorithm based on mutual information with pairwise constraints
Jing Song, Ming Bo Yang, Genlin Ji, Wenbin Cai · 2010
Feature selection plays an important role in the area of machine learning. Class Label is often used as the supervised information for supervised feature selection algorithm while constraints are rarely used. So, an effective feature selection algorithm with pairwise constraints called Constraints Score was proposed. But its performance still is limited by neglecting the correlation between features. In this paper we improve this algorithm by considering the correlation between features and using SVM density estimation, mutual information to measure the correlation and further eliminate the feature redundancy. Experiments show the effectiveness of our improved algorithm.