Feature selection algorithm for dynamically evolving feature spaces

Feng Wang

doi:10.1016/j.neucom.2025.131747

Feature selection algorithm for dynamically evolving feature spaces

2025-10-11

0

PRE

AI

DOI：10.1016/j.neucom.2025.131747

Original

Original request for help

Save

Abstract

En 中文

In many applications, indications abound that most of the real-world data vary with time, such as medical research and weather data. Especially with the coming of big data, data vary dynamically not only in size but also in dimension at an unprecedented speed. In response to the issue of dimensionality change, it is not merely the increase in dimensionality but also potential alterations in the feature space that need to be considered. Thus, it becomes very difficult or even hardly feasible to use existing incremental techniques to handle such dynamic data. To address this issue, this paper proposes an efficient dynamic feature selection algorithm tailored for data exhibiting data distribution variation due to evolving feature spaces. In this study, a comprehensive set of novel strategies for measuring, detecting, and tracking is introduced to capture variations in data dimension distribution firstly. On this basis, a dynamic algorithm for selecting informative features from multiple sub-tables is proposed. Finally, a multi-layered fusion mechanism is proposed to integrate all the results derived from the granules and determine the final outcome. To comprehensively demonstrate the efficacy of the proposed algorithm, several state-of-the-art incremental and classical feature selection algorithms are selected as benchmark methods for comparison. Experiments are conducted on multiple high-dimensional data sets. The experimental results further validate the feasibility and efficiency of the proposed algorithm.

Feature selection algorithm for dynamically evolving feature spaces

Abstract

Journal

Organization