Feature selection for label distribution learning based on the statistical distribution of data and fuzzy mutual information

Abstract

Label distribution learning (LDL) is an emerging framework in machine learning. Fuzzy mutual information is mutual information under a fuzzy environment and plays an important role in handling uncertainty. This paper explores feature selection for LDL data based on the statistical distribution of data and fuzzy mutual information. The similarity between the feature values in the feature space is first defined by means of the statistical distribution of the data, and a threshold is introduced to control the similarity. Then, the fuzzy similarity relation for each feature subset is established via the similarity. This method utilizes adjustable fuzzy similarity radii to establish a fuzzy similarity relation and improve the classification ability of the data. The decision relation in the label space is then presented, and the decision class of each sample is constructed. Subsequently, two feature selection algorithms based on fuzzy mutual information are designed to remove the irrelevant features by employing a strategy that considers the correlation between the features and labels as well as the redundancy between the features in the LDL data. Finally, the experimental results show that the designed algorithms can effectively measure the uncertainty of LDL data and outperform four state-of-the-art feature selection algorithms. Specifically, our algorithms, LDFM and LDFMR, demonstrate their superiority by achieving overall average ranking improvements of 63.64% and 58.52%, respectively, across six evaluation metrics compared to the other four algorithms.