The Art of Saying ‘I Don’t Know’: Learned Deferment in Neuro-Fuzzy Three-Way Decision Cascade Classifiers

The three-way decision paradigm is a new and auspicious paradigm approach to classification. It introduces a non-commitment region, allowing classifiers to abstain from (defer) uncertain predictions. This is a key mechanism in cascade classification systems, where samples assigned to the non-commitment region in one classifier are passed to the next one. Fixed thresholds are often used to determine the non-commitment region, but they require fine-tuning and provide limited insight into the reasons for deferment. We propose an automatic mechanism for determining the non-commitment region using auxiliary metaclassifiers. We reframe deferment as a learnable decision problem rather than a thresholding problem. Each metaclassifier predicts whether its accompanying classifier is likely to make a correct prediction for a given sample and decides whether to return a final answer or defer the decision to the next cascade stage. In this approach, deferment is based on a broader context than a single confidence threshold, and it is tailored to the characteristics of each classifier and dataset. With neuro-fuzzy systems used as metaclassifiers, deferment decisions can be expressed as human-readable rules.