Rule Inference Network for Classification.

To alleviate the low interpretability of algorithms in neural networks, we propose a rule inference network based on rule-based system using the evidential reasoning approach (RIMER), which is interpretable by the rules in belief rule base (BRB). Considering the influence of data distribution on attribute weights, we optimize attribute weights based on the Sigmoid activation function to ensure that the normalization process adapts the overall data distribution. A rule inference network for classification is constructed including the framework and the learning algorithm. A comparison with the other algorithms demonstrates that the proposed rule inference network for classification has advantages in interpretability and learning capability.