Feature selection for label distribution learning using Dempster-Shafer evidence theory

期刊论文
发表日期：

2025年1月3日

查看全文

Zhengwei Zhao,
Rongrong Wang,
Wei Pang,
Zhaowen Li

摘要

In the contemporary epoch of massive data, the fuzziness of labels and the high dimensionality of feature space are prevalent characteristics of data. As a mathematical methodology for managing uncertainty, Dempster-Shafer evidence theory has found widespread applications in artificial intelligence, pattern recognition, and decision analysis. However, it has not garnered adequate attention in label distribution learning (LDL). This paper studies feature selection for LDL using Dempster-Shafer evidence theory. First, for a LDL data, distance maps in the feature space and in the label space are given, respectively. Furthermore, a tunable parameter to regulate the proximity level of features or labels is implemented. Then, the $\alpha $ $α$ -upper and $\alpha $ $α$ -lower approximations in the LDL data are put forward. Subsequently, to alleviate the influence of uncertainty on classification performance, robust feature evaluation measures for a LDL data, namely, “belief map" and “plausibility map" are defined, and they are based on the approximations. Next, feature selection algorithms utilizing belief and plausibility maps are specially designed. Finally, experimental results and statistical analyses demonstrate that the defined belief and plausibility maps can effectively measure the indeterminacy of LDL data, and the designed feature selection algorithms outperform five existing algorithms regarding classification performance.