Feature selection based on distance correlation: a filter algorithm

Abstract

Feature selection (FS) is one of the most powerful techniques to cope with the curse of dimensionality. In the study, a new filter approach to feature selection based on distance correlation is presented (DCFS, for short), which keeps the model-free advantage without any pre-specified parameters. Our method consists of two steps: hard step (forward selection) and soft step (backward selection). In the hard step, two types of associations, between univariate feature and the classes and between group feature and the classes, are involved to pick out the most relevant features with respect to the target classes. Due to the strict screening condition in the first step, some of the useful features are likely removed. Therefore, in the soft step, a feature-relationship gain (like feature score) based on the distance correlation is introduced, which is concerned with five kinds of associations. We sort the feature gain values and implement the backward selection procedure until the errors stop declining. The simulation results show that our method becomes more competitive on several datasets compared with some of the representative feature selection methods based on several classification models.

Keywords:

• Feature selection
• distance correlation
• S-correlation
• feature relevance
• feature redundancy

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 http://mlr.cs.umass.edu/ml/datasets.html

Additional information

Funding

This work is supported by the National Natural Science Foundation of China: Managing Uncertainty in Service Based Software with Intelligent Adaptive Architecture (No. 61732019).