New Analysis and Algorithm for Learning with Drifting Distributions
This work addresses the problem of adapting machine learning models to changing data distributions over time, which is incremental as it builds on existing discrepancy-based methods.
The paper tackles learning with drifting distributions by introducing a new analysis using discrepancy, proving tighter learning bounds than previous methods based on L1 distance, and presenting a new algorithm formulated as a quadratic program that shows benefits in preliminary experiments.
We present a new analysis of the problem of learning with drifting distributions in the batch setting using the notion of discrepancy. We prove learning bounds based on the Rademacher complexity of the hypothesis set and the discrepancy of distributions both for a drifting PAC scenario and a tracking scenario. Our bounds are always tighter and in some cases substantially improve upon previous ones based on the $L_1$ distance. We also present a generalization of the standard on-line to batch conversion to the drifting scenario in terms of the discrepancy and arbitrary convex combinations of hypotheses. We introduce a new algorithm exploiting these learning guarantees, which we show can be formulated as a simple QP. Finally, we report the results of preliminary experiments demonstrating the benefits of this algorithm.