Dynamic selection of machine learning models for time-series data

Abstract

Adapting to changes in time-series data is a major challenge in machine learning. This problem is particularly acute in the case of limited computing power (e.g., edge devices) that does not enable the independent training of new models. While multiple studies attempt to adequately train a machine learning model on rapidly shifting data (i.e., concept drift), the challenge of dynamically selecting the most effective machine learning model for future time steps has been largely overlooked. In this study, we propose Adaptive machine learning for Dynamic ENvironments (ADEN ), a method for analyzing future trends in time-series data and selecting the most suitable ML model. Our approach models multiple aspects of the analyzed data and analyzes the behavior of multiple machine learning models on earlier time steps. By training ADEN on multiple datasets, we can deploy a zero-shot model that does not require additional training when applied to new datasets. Our evaluation, conducted on 46 time-series classification datasets, shows that ADEN not only outperforms all baselines in terms of average performance but is also capable of avoiding sudden drops in performance that characterize all other evaluated algorithms.