Unsupervised Feature Selection with Fuzzy Clustering and Local Structure Learning

Xiaorong Wang,Shuisheng Zhou

DOI: https://doi.org/10.1109/icnc-fskd64080.2024.10702214

2024-01-01

Abstract:Feature selection is a crucial technique for handling high-dimensional data. In clustering analysis tasks, most feature selection methods typically involve selecting features using a projection matrix first, and then performing clustering with the selected features. However, these approaches may result in the selection of features during the initial stage that are not necessarily appropriate for subsequent clustering tasks. To deal with this matter, we proposes a strategy that combines fuzzy clustering with local structure learning for selecting features. We introduce fuzzy clustering to guide the process of feature selection, rather than solely relying on structural learning for feature selection. This approach effectively integrates two-stage models into a framework, enabling the acquisition of features that can represent both local structural information of the data and are suitable for clustering tasks, thereby enhancing the model's clustering performance. Moreover, we employ an $\ell_{2,0}$ -norm constrained projection matrix for feature selection. This approach is easier to interpret than $\ell_{2,1}$ -norm selection and eliminates the need for additional sorting operations, thus streamlining the feature selection process. It is worth emphasizing that we introduce a projection relaxation term. The presence of this projection relaxation term also makes the learned graph more flexible. Experiments demonstrate that the proposed algorithm has certain advantages over some classical algorithms.