Multi-granularity unsupervised feature selection based on the entropy ball model

Xia Ji

doi:10.1016/j.ins.2025.122781

2025-11-01

PRE

DOI：10.1016/j.ins.2025.122781

原文链接

摘要

En 中文

Feature selection is a core research topic in machine learning and data mining. While existing feature selection algorithms have produced useful research results, most of them focus on labeled data with in supervised scenarios. However, in real-world applications, most data is unlabeled.Therefore, the research results of unsupervised feature selection algorithms have important application value, and further enrichment and development are urgently needed. Existing unsupervised feature selection algorithms mostly perform feature selection from a single granularity level, that is, they evaluate feature importance and select features in the units of the original sample layer of the dataset. However, analyzing data at only a single granularity level ignores information at the other granularity levels, thus limiting the ability of algorithms to understand complex data, affecting their assessment of feature importance, and impairing their performance. To address this problem, we propose an entropy ball model for multi-granularity data representation to capture the underlying multi-level structural information of the data. Subsequently, we construct an extended fuzzy rough set model based on the entropy ball model to address mixed types of data. We further design a multi-granularity unsupervised feature selection algorithm (MGUFS-EB) based on the entropy ball model and define a feature contribution evaluation metric to measure the importance of features. Then, we conduct comprehensive comparative experiments on the performance of nine advanced unsupervised feature selection algorithms in classification, clustering, and outlier detection. Experimental results demonstrate that MGUFS-EB achieves superior performance and high robustness.

Keyword:

Entropy ball model

Fuzzy rough set

Multi-granularity

Unsupervised feature selection