MDIE-AutoML: Multi-Detector Drift Detection with Dynamic Weighted Ensemble for Streaming AutoML

Abstract

Non-stationary data streams suffer from simultaneous data and concept drifts that degrade model generalization. Conventional Automated Machine Learning for data streams, AML4S, relies on single-pipeline architecture with univariate ADWIN detection and exhaustive post-drift reconstruction, causing knowledge waste and insufficient ensemble robustness under high-dimensional composite drift. This work proposes MDIE-AutoML, integrating PCA-aided multi-detector joint drift identification, warm-start incremental retraining, DWE-based Top-K weighted ensemble, and a drift-category-aware memory bank. Evaluation on LoanDataset and four benchmarks spanning natural, cyclic, gradual, and high-dimensional multi-class drift demonstrates that the MDIE-Top5 variant achieves 0.9311 ± 0.0097 overall accuracy on the 20,000-sample main experiment, with roughly half the standard deviation of the AML4S baseline, and ranks first in seven of nine cross-scenario tests. These results confirm that MDIE achieves superior prediction stability with manageable computational overhead across diverse non-stationary scenarios.