A Novel Online Multi-label Feature Selection Approach for Multi-dimensional Streaming Data

Online feature selection for streaming data has attracted much attention in the field of multi-label learning. Most of the existing online approaches can efficiently deal with the single-dimensional variation of a multi-label information system. However, multi-dimensional variations often occur in real-time streaming applications. Based on the improved Fisher score model for multi-label learning and feature redundancy analysis using symmetric uncertainty, we propose a novel online multi-label feature selection framework for both streaming feature and label spaces. For the situation of streaming features, the proposed framework calculates the Fisher score to obtain the importance of the feature, determines the redundancy of the newly arrived feature based on the symmetry uncertainty, and then obtains the position of the feature in the final feature rank list. For the newly arrived labels, we recalculates the weights of all current labels and updates the total Fisher score to update the current feature rank list. In the experiments, we compare the performance of our approach with four representative online feature selection algorithms for streaming features and labels, respectively. The extensive experimental results on nine multi-label benchmark datasets by using two evaluation metrics commonly used in multi-label classification demonstrate the effectiveness of the proposed framework.