FERL: Fast Evidential Rule Learning for Interpretable Classification with Abstention
A new method called Fast Evidential Rule Learning (FERL) has been introduced for interpretable classification with abstention. FERL learns fuzzy rule models that are both interpretable and accurate, with outputs that are evidential, meaning they provide belief, plausibility, and abstention capabilities directly from fuzzy memberships in a single deterministic pass, without auxiliary heads, held-out sets, or repeated inference. The method is theoretically shown to be Lipschitz stable, ensuring that its evidential outputs vary smoothly with input. In benchmark tests against state-of-the-art rule learners, FERL achieved statistically significant higher accuracy across a 30 tabular-dataset benchmark, with an average accuracy improvement of +2.6% over the second-best method. Its native set predictions also attained the best utility. The paper is available on arXiv under the identifier 2608.05859.
Key facts
- FERL learns interpretable, accurate fuzzy rule models with evidential outputs.
- Abstention capabilities arise directly from fuzzy memberships in a single deterministic pass.
- FERL is Lipschitz stable, ensuring smooth variation of evidential outputs with input.
- FERL outperforms state-of-the-art rule learners on a 30 tabular-dataset benchmark.
- Average accuracy improvement of +2.6% over the second-best method.
- Native set predictions attain the best utility.
- Paper available on arXiv with identifier 2608.05859.
- Method avoids post-hoc calibration and auxiliary heads.
Entities
Institutions
- arXiv