Evidential Reasoning Rule Learning

In recent decades, artificial intelligence (AI), particularly machine learning (ML), has achieved remarkable advancements. However, the practical implementation of ML in critical real-world applications remains constrained due to the concerns on building reliable ML model. To address this challenge, ensuring model reliability requires a focus on three major aspects: credibility, adaptability, and interpretability. To achieve these objectives in a unified way, a new reliable learning pattern which consists of training, adaptation, testing and second decision stage was proposed to extend the traditional training-testing pattern. Based on this framework, a new evidential reasoning rule learning (ER2L) strategy was developed, where individual-reliability is defined to tune the trained model for fitting the test sample better in adaptation stage and overall-reliability is introduced to measure the credibility of the model outputs for the second decision. Furthermore, due to ER2’s nature as a probabilistic inference engine, the model is interpretable. Building on ER2L, a new ML approach termed Naïve ER2(NER2) was proposed. The experimental studies demonstrated that NER2can obtain better performance by introducing individual-reliability for model tuning, whilst the overall-reliability can evaluate the credibility of model output reasonably.