Predicting Concept Drift in Data Streams Using Metadata Clustering
Robert Anderson, Yun Sing Koh, Gillian Dobbie · 2018
Concept drifts cause data streams to change over time, making classification harder. Drift detection signals when change occurs so remedial action can be taken. However, existing approaches do not allow drift detectors to improve drift detection by using patterns associated with drift across multiple stream metadata. Metadata Drift Predictor (MDP) is a novel framework which allows this by adapting drift sensitivity mid-stream to reduce errors in drift detection. MDP clusters metadata patterns at drift points and uses recurring patterns to encourage drift detection where metadata looks similar to past drift points and/or discourage drift detection where metadata looks different from past drift points. It avoids relying on statistical assumptions of underlying drift patterns. Our experiments demonstrate that MDP works consistently across traditional stream classifiers and drift detectors to reduce delay and false positive drift detection and/or reduce false negative drift detection. We show that MDP can be used to reduce drifts detected in real-world datasets while improving the accuracy that each detected drift provides when compared to other popular meta-learning frameworks.