Joint Adaptive Graph and Structured Sparsity Regularization for Unsupervised Feature Selection

TL;DR

This paper introduces JASFS, an unsupervised feature selection method that adaptively learns data structure and selects feature groups automatically, improving over existing independent selection methods.

Contribution

The paper proposes a novel joint adaptive graph and structured sparsity regularization framework for unsupervised feature selection, with automatic feature subset size determination.

Findings

Outperforms several state-of-the-art methods on benchmark datasets.

Selects feature groups rather than individual features.

Automatically determines the optimal number of features.

Abstract

Feature selection is an important data preprocessing in data mining and machine learning which can be used to reduce the feature dimension without deteriorating model's performance. Since obtaining annotated data is laborious or even infeasible in many cases, unsupervised feature selection is more practical in reality. Though lots of methods for unsupervised feature selection have been proposed, these methods select features independently, thus it is no guarantee that the group of selected features is optimal. What's more, the number of selected features must be tuned carefully to obtain a satisfactory result. To tackle these problems, we propose a joint adaptive graph and structured sparsity regularization unsupervised feature selection (JASFS) method in this paper, in which a $l_{2,0}$-norm regularization term with respect to transformation matrix is imposed in the manifold learning for feature selection, and a graph regularization term is incorporated into the learning model to learn the local geometric structure of data adaptively. An efficient and simple iterative algorithm is designed to solve the proposed optimization problem with the analysis of computational complexity. After optimized, a subset of optimal features will be selected in group, and the number of selected features will be determined automatically. Experimental results on eight benchmarks demonstrate the effectiveness and efficiency of the proposed method compared with several state-of-the-art approaches.