A novel relation aware wrapper method for feature selection

Feature selection, aiming at eliminating irrelevant and redundant features, is an important data preprocessing technology for downstream tasks, e.g., classification. With the explosive growth of data in various fields, some data are high-dimensional and contain critical and complex hidden relationships, which brings new challenges to feature selection: i) How to find out the underlying available relationships from the data, and ii) how to use the learned relations to better select features? To deal with these challenges, we propose a novel wrapper feature selection method named R e lation Awa r e Fe a ture S election M e thod (ERASE), which can learn and use the underlying sample relations and feature relations for feature selection. Different from existing methods, our method jointly learns sample relationships and feature relationships through a graph of samples and trees of features. Furthermore, it uses the relations to select the optimal feature subset according to the new proposed Relation-based Sequence Floating Selection Strategy. Extensive experimental results on nine datasets from different domains demonstrate that our method achieves the best performance in most cases compared with other feature selection methods, including state-of-the-art wrapper methods.

Automated summary

In this paper, a novel wrapper feature selection method named ERASE is proposed, which learns and utilizes sample relations and feature relations for feature selection. Experimental results demonstrate that our method outperforms other feature selection methods in most cases.