Answers are grounded in this paper and cite their sources.
Learning from data streams is challenging because the data distribution may change during model deployment, making previously reliable decision rules less dependable. At the same time, observations may be vague or fall within overlapping class regions, where forcing binary decisions can obscure decision reliability. Recomputing fuzzy rough models from the complete history is also computationally expensive. This paper develops IFR-3WD-CD, an incremental fuzzy rough three-way decision framework for concept-drifting data streams that addresses these challenges jointly. A vaguely quantified fuzzy rough membership is maintained over a bounded sliding window with online standardization to provide a noise-tolerant decision score. A decision-theoretic three-way rule assigns instances to accept, defer, or reject regions, while a drift adaptation layer separates detection from adaptation. A change detector triggers selective forgetting, and a bounded drift-intensity signal adaptively widens the deferral region during unstable periods. The resulting per-instance computational cost is independent of the elapsed stream length, avoiding repeated reconstruction from historical data. The framework is evaluated using a strictly chronological prequential protocol on four synthetic benchmarks representing abrupt, gradual, recurring, and incremental drift, and two real-world streams, against eight adaptive baselines. Results show competitive predictive accuracy while deferring a modest fraction of difficult instances, with committed decisions substantially more accurate than forced decisions. Ablation experiments confirm the contribution of individual components, while boundary-region analysis demonstrates expansion after drift and contraction during recovery. The study demonstrates the utility of uncertainty-aware three-way decisions for adaptive learning under evolving data distributions.
No takes yet. Share an insight, caveat, or question.
Ansari et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: