Low-Rank Transfer Learning for Multi-stream Data Classification

Leveraging the labeled data from source stream to handle unlabeled data from target stream is a promising method for multi-stream classification. However, there are still two main issues, concept drifts and concept revolution, needed to be addressed. A couple of methods is proposed to solve them, such as adding a change detection module, modifying sliding window size and setting concept confidence. However, the majority of these approaches did not consider shared data structures among multiple data streams, which can be employed to handle these two issues. At the same time, a lot of algorithms assume that the source data stream and the target data stream are all come from the same domain. In reality, however, most data streams are from different domains with different data distributions. In this paper, therefore, we propose a novel framework Low-Rank Transfer Learning (LRTL) to address cross-domain multi-stream classification. LRTL is mainly composed of three modules, i.e., domain adaptation module, change detection module and classification module. Different from these existing multi-stream classification frameworks, we design a novel domain adaptation module, which utilizes low-rank representation and graph embedding to preserve data structures that have many benefits in dealing with concept drifts and concept revolution. Extensive experiments are carried out on the domain adaptation module and the proposed framework LRTL on multiple benchmarks, which verifies that our domain adaptation methods and frameworks can achieve the state-of-the-art results with significant improvements.