Attribute reduction for multi-label learning with fuzzy rough set

摘要

In multi-label learning, each sample is related to multiple labels simultaneously, and attribute space of samples is with high-dimensionality. Therefore, the key issue for attribute reduction in multi-label data is to measure the quality of each attribute with respect to a set of labels. Stimulated by fuzzy rough set theory, which allows different fuzzy relations to measure the similarity between samples under different labels. In this paper, we propose a novel fuzzy rough set model for attribute reduction in multi-label learning. Different from single-label attribute reduction, a bottleneck of fuzzy rough set for multi-label attribute reduction is to find the true different classes’ samples for the target sample, which deeply affects the robustness of fuzzy upper and lower approximations. We first define the score vector of each sample to evaluate the probability of being different class’s sample with respect to the target sample. Then, local sampling is leveraged to construct a robust distance between samples. It can implement the robustness against noisy information when calculating the fuzzy lower and upper approximations under the whole label space. Moreover, multi-label fuzzy rough set model is proposed, and some related properties are discussed. Finally, the significance measure of a candidate attribute is defined, and a greedy forward attribute selection algorithm is designed. Extensive experiments are carried out to verify the effectiveness of the proposed algorithm by comparing it with some state-of-the-art approaches on eight publicly available data sets.