Learn-to-adapt: Concept drift adaptation for hybrid multiple streams

Abstract

Existing concept drift adaptation (CDA) methods aim to continually update outdated classifiers in a single-labeled stream scenario. However, real-world data streams are massive, with hybrids of labeled and unlabeled streams. In this paper, we discuss CDA in multiple data streams that may contain unlabeled drifting streams. To address this realistic and complex problem, we rethink the concept drift problem by adopting a meta-learning approach and introduce a Learn-to-Adapt framework (L2A). The L2A framework simultaneously 1) makes adaptations for drifting labeled streams, and 2) leverages knowledge from labeled drifting streams to make adaptations for unlabeled stream prediction. In L2A, a meta-representor with an adapter in the meta-training stage is designed to learn the invariant representations for drifting streams, enabling the model to quickly produce a good generalization of new concepts with limited training samples. In the online stage, the meta-representor will be adapted continually under the control of the adapter and will contribute to adapting the classifiers for unlabeled drifting stream prediction. Compared to existing CDA methods which mostly only adapt the classifiers, L2A adapts the feature extractor and classifier in a feedback process, which is advanced in dealing with more complex and high-dimensional data streams.