PRISM-DM: Drift-Aware, Imbalance-Resilient Ensemble Learning for Adaptive Data Stream Mining
Shodhshauryam International Scientific Refereed Research · 2022
DOI: 10.32628/shisrrj221240 ↗Kurz gefasst
Fachartikel in der Zeitschrift Shodhshauryam International Scientific Refereed Research, 2022. Verfasst von Prudvi Saisaran Ponduru Prudvi Saisaran Ponduru. Die Arbeit nennt 33 Literaturangaben.
Abstract (Originalsprache)
Artificial-intelligence systems that learn from continuously arriving data face two coupled difficulties: the statistical concept can change while the class distribution can become severely imbalanced. A model that reacts aggressively to drift may forget useful structure, whereas a model optimized for aggregate accuracy can ignore rare but operationally important classes. This paper introduces PRISM-DM, a prequential risk- and imbalance-sensitive ensemble for adaptive data-stream mining. PRISM-DM combines a two-horizon balanced-error monitor, dynamic inverse-frequency and difficulty weighting, a bounded class-balanced selective memory, drift-triggered pruning and weakest-member reset, and diversity-aware ensemble weighting. The method is evaluated under a strict test-then-train protocol on six evolving streams: abrupt, gradual, and recurring synthetic hyperplane shifts; a changing-prior multiclass stream; and controlled replay streams derived from the Digits and Breast Cancer datasets. Five locked seeds produce 30 matched stream-seed evaluations per method. Across those evaluations, PRISM-DM obtains 0.779 mean balanced accuracy, 0.745 macro-F1, 0.774 G-mean, and 0.694 mean worst-class recall. Its balanced-accuracy advantage is +15.44 percentage points over a static ensemble, +9.30 over online SGD, +9.19 over cost-sensitive SGD, and +10.28 over an Oza-style online bagging baseline; exact paired Wilcoxon tests remain significant after Holm correction. Ablation results identify drift response as the dominant component, while balanced memory and diversity weighting provide smaller, stream-dependent effects. PRISM-DM is not uniformly superior: on the Digits replay stream, cost-sensitive SGD attains higher balanced accuracy, exposing a boundary where conservative ensemble adaptation can lag a simpler learner. The results support a reproducible design for AI systems that must mine nonstationary, imbalanced streams without treating majority-class accuracy as the only objective.
Zitieren
APA
Ponduru, P. S. P. P. S. (2022). PRISM-DM: Drift-Aware, Imbalance-Resilient Ensemble Learning for Adaptive Data Stream Mining. Shodhshauryam International Scientific Refereed Research. https://doi.org/10.32628/shisrrj221240
Vancouver
Ponduru PSPPS. PRISM-DM: Drift-Aware, Imbalance-Resilient Ensemble Learning for Adaptive Data Stream Mining. Shodhshauryam International Scientific Refereed Research. 2022. doi:10.32628/shisrrj221240
BibTeX
@article{Ponduru2022,
author = {Ponduru, Prudvi Saisaran Ponduru Prudvi Saisaran},
title = {PRISM-DM: Drift-Aware, Imbalance-Resilient Ensemble Learning for Adaptive Data Stream Mining},
journal = {Shodhshauryam International Scientific Refereed Research},
year = {2022},
doi = {10.32628/shisrrj221240}
}
EUVIMED