Multi-label Online Streaming Feature Selection Algorithms via Extending Alpha-Investing Strategy

Multi-label learning is a special supervised pattern classification issue, in which an instance is possibly associated with multiple class labels simultaneously. As various real world applications emerge continuously in the big data field, more attention has been paid to streaming data forms recently, i.e., instance, feature and label streams. In this paper, we focus on multi-label online streaming feature selection (OSFS) problem, whose features arrive sequentially over time, and instances and labels are given, to choose an optimal subset of features dynamically. Alpha-investing method is one of the most cited embedded-type single-label OSFS techniques, which mainly involves the linear regression model as its classifier. In this paper, we generalize such a technique to build two new multi-label OSFS algorithms (simply ML-AIBR and ML-AIMOR), which are based on binary relevance (BR) decomposition way and multioutput regression (MOR), respectively. To the best of our knowledge, such two algorithms are the first proposed embedded-type OSFS technique for multi-label streaming features so far. Our extensive experiments conducted on six benchmark data sets demonstrate that our two proposed methods performs better than three existing algorithms. Specially, our ML-AIMOR could filter out more irrelevant and redundant features effectively.