Reliable Semi-Supervised Learning on Imbalanced Evolving Data Stream

Abstract

Existing semi-supervised learning (SSL) algorithms often heavily depend on some assumptions (e.g., cluster assumption) and usually work on class-balanced static datasets. If the assumption(s) does (do) not hold, the biased prediction of unlabeled data may even hurt accuracy. This issue becomes more problematic in the context of streaming data due to the existence of concept drift. Therefore, it’s of great importance to enhance the reliability of the SSL algorithms on imbalanced concept-drifting data streams. In this paper, we propose a reliable and scalable SSL framework on imbalanced evolving data stream. Instead of relaxing different assumptions, we apply a novel sampling strategy and an additional balanced classifier to reduce the impact of imbalance and introduce the deep metric learning loss to enlarge the class margin to increase the degree of discrimination. We further maintain a small set of reliable micro-clusters dynamically in that embedding space and employ different strategies to update their reliabilities to maintain the most recent concept and cope with concept drifts. We conducted some experiments on real and synthetic stream datasets to evaluate the effectiveness of our proposed model.