Label enhancement-based feature selection via fuzzy neighborhood discrimination index

摘要

As an essential preprocessing step, the feature selection for multi-label classification is an efficient tool to solve the high-dimensionality in training data under multiple semantics. However, in multi-label learning, the assumption of a uniform distribution is not practicable for some real applications where labels are related to the instance with different relative importance degrees. Motivated by this, the label enhancement algorithm is utilized to transform the logical values in multiple labels into label distributions with real values. To maintain the supervised information of label distributions, the discrimination index is combined with granular theory to describe the vagueness in feature and label spaces, and the fuzzy neighborhood discrimination index is presented to assess the discrimination information in features with continuous values. Furthermore, an innovative label enhancement-based feature selection approach is designed to handle the challenge caused by high dimensionality with multiple semantics. In this algorithm, the fuzzy conditional discrimination index is proposed by combining the fuzzy neighborhood discrimination index under the feature subset with the label similarity matrix of label distributions, and the significance of the feature is measured by the fuzzy conditional discrimination index, which is utilized to assess the discernibility information of labels under the feature. Furthermore, to validate the efficacy of our algorithm, a collection of experiments are conducted with five representative multi-label feature selection methods under twelve benchmark datasets. With six widespread evaluation indicators of multi-label classification, the experimental results indicate that our algorithm obtains superior performance against other algorithms.