Feature selection based on IB theory

Machine learning and pattern recognition are confronted with the difficulty of feature selection. However, the data for clustering are unlabelled and there is no commonly accepted evaluation criterion to clustering accuracy. Therefore, feature selection has been paid little attention in unsupervised learning or clustering. This paper proposed a feature selection method based on IB theory. It selected the most effective feature subset while preserved the most information. The experimental results on selected UCI datasets showed that it not only reduced the dimension but also got better clustering accuracy. So, the method is valid.