New Framework for Incremental Learning with Evolving Concepts

Researchers have developed a novel framework for incremental learning that specifically addresses scenarios where the definitions of target concepts change over time. Unlike traditional concept drift methods that infer changes from data observations, this approach focuses on "rule-induced concept shift," where the underlying policies or rules defining a concept are explicitly revised. This revision means that previously labeled data instances might now have different semantic labels, even if their raw data hasn't changed. The "provenance-guided incremental learning" framework works by compiling consecutive concept definitions into a structured "rule delta." It then traces these changes through historical data provenance, identifying and certifying records whose previous labels remain valid. This allows the system to restrict reevaluation and relabeling to only a localized candidate region of affected data, rather than reprocessing the entire historical collection. The framework automates relabeling for executable revisions, handles ambiguous cases through selective supervision, and uses these changes for incremental predictor repair. A new benchmark, RuleShift-Bench, demonstrates that this provenance-guided repair achieves high accuracy and F1 scores while significantly reducing the amount of historical data reprocessed and drastically cutting update latency compared to full relabeling and retraining.