[5M3-GS-2h] Machine learning

2:15 PM - 2:30 PM JST(5:15 AM - 5:30 AM UTC)

[5M3-GS-2h-02] Online Incremental Drift Adaptation for Mitigating Accuracy Degradation

〇Ryuta Matsuno¹ (1. NEC Corporation)

Keywords:

Data Drift,Incremental Drift,Distribution Shift,Machine Learning

This paper addresses the problem of incremental drift adaptation, which aims to achieve accurate online prediction from past data streams under incremental drift. Specifically, incremental drift is modeled as temporal changes in the mixing weights over fixed base parameters within the parameter space. In this paper, we propose a novel incremental drift adaptation method. We theoretically analyze the problem and introduce fundamental assumptions for optimal adaptation. Based on this analysis, our method learns base parameters and time-dependent mixing weights from offline training data, and then adapts to the online data stream by updating only the mixing weights. By learning drift within the parameter space of the joint distribution, our method is more general and flexible than existing methods. Experiments on six drifting real-world datasets show that our method reduces the cumulative squared loss by up to 51.2%, demonstrating its superiority over baseline methods.