FSCLBF: Feature Selection based on Correlation Label and B-R Belief Function in Multi-label Data
Zahra Mehravaran, Javad Hamidzadeh, Reza Monsefi · Research Square · 2022
Abstract In many real-world problems, each sample has over one label, which may also have semantic correlations between labels. However, in multi-label datasets, there are many irrelevant features that impact the accuracy of multi-label learning. Using feature selection to improve accuracy is an effective method. Even though numerous papers have presented novel methods to feature selection, more effort has to be done to improve the accuracy of such methods. This paper presents a novel feature selection method based on correlation labels using a belief function called FSCLBF. Our proposed method increases the accuracy of multi-label learning by considering the correlation labels and the data uncertainty. The proposed method's first step is to select the essential features for each label using a belief function based on a rough set theory (B-R belief function). Then, we identify the equivalent labels according to the overlapping based on their essential features. In the next step, the tag of samples is determined in the set of equivalent labels. Finally, we select the essential features for each equivalent label set, using the B-R belief function by considering the sample differentiation to label. The experimental results show the superiority of FSCLBF to state-of-the-art methods in terms of Hamming loss, One-error, and Accuracy measures.