ECCRS: An Interpretable Rule Classifier with Exceptions
Aus International Center for Computational Logic
ECCRS: An Interpretable Rule Classifier with Exceptions
Vortrag von Ruvarashe Madzime
- Veranstaltungsort: APB-2026
- Beginn: 13. November 2025 um 11:00
- Ende: 13. November 2025 um 12:00
- Forschungsgruppe: Wissensbasierte Systeme
- Event series: Research Seminar Logic and AI
- iCal
In this talk, Exception Closed Conjunctive Rule Sets (ECCRS) will be presented as a transparent rule based classifier with explicit exceptions. Exception closure induces laminarity among opposite label rules, so prediction follows a clear most specific wins principle. The learning pipeline builds minimal decisive rule bodies from training examples, enforces comparability with opposite rules, and can apply a strict global closure that yields a canonical order independent model. Two base preserving completions resolve undecided cases only when required. On standard binary classification benchmarks ECCRS is competitive and often comparable to symbolic and non symbolic baselines. Models are compact with direct reasons per decision. An ablation quantifies the trade off between closure level, model size, and time.
