Beyond Incremental Processing: Tracking Concept Drift

Abstract

Learning in complex, changing environments requires methods that are able to tolerate noise (less than perfect feedback) and drift (concepts that change over time). These two aspects of complex environments interact with each other: when some particular learned predictor fails to correctly predict the expected outcome (or when the outcome occurs without having been preceded by the learned predictor), a learner must be able to determine whether the situation is an instance of noise or an indication that the concept is beginning to drift. We present a learning method that is able to learn complex Boolean characterizations while tolerating noise and drift. An analysis of the algorithm illustrates why it has these desirable behaviors, and empirical results from an implementation (called STAGGER) are presented to show its ability to track changing concepts over time.