Abstract

Machine learning systems deployed in dynamic environments frequently operate under non-stationary data distributions, where controlled distribution shift can progressively degrade predictive performance. However, many widely used tabular benchmark datasets lack explicit temporal structure, limiting reproducible evaluation of drift adaptation methods. This work proposes a cluster-induced distribution shift simulation framework that transforms static tabular datasets into controlled evolving data streams through structured perturbations across feature-space partitions.

Using this framework, six adaptation strategies are systematically evaluated: static learning, sliding-window retraining, global ADWIN retraining, cluster-local ADWIN retraining, random subspace drift detection, and feature-partitioned drift detection. Experiments are conducted on five benchmark datasets covering both classification and regression tasks using diverse predictive model families, including linear models, k-Nearest Neighbours, tree ensembles, boosting methods, and adaptive online learners.

The results show that sliding-window retraining often achieves the strongest predictive performance but incurs substantial computational and labelling cost due to frequent global updates. Global ADWIN retraining reduces update frequency while preserving competitive accuracy. In contrast, the proposed cluster-local adaptation strategy consistently achieves a stronger balance between predictive performance and computational efficiency across both classification and regression settings. Under structured distribution-shift scenarios, the method reduces retraining effort by up to 75% while maintaining competitive predictive performance relative to continuously adaptive baselines. For nonlinear regression models, cluster-local adaptation preserves competitive R2 performance while substantially lowering update frequency and training effort.

The proposed framework provides a reproducible benchmark for evaluating adaptation under controlled distribution shift on static tabular datasets and demonstrates that cluster-aware drift/shift detection constitutes an effective and computationally efficient alternative to uniform retraining strategies under heterogeneous distributional shift conditions.