RB-Net: Rule-Based Neural Networks with Interpretable Models
Zhehang Xue, Yangfei Zheng · 2024
In the current landscape of machine learning, black-box models face certain limitations in applications such as healthcare and law due to their opacity. Consequently, there’s a growing preference for interpretable models to address these issues, with rule-based models being a significant category among them. Traditional rule-based models often rely on heuristic search techniques customized for particular rule-learning scenarios. These methods contrast sharply with the gradient-based approaches typical in neural machine learning. Recent efforts to bridge the gap between neural and symbolic machine learning have led to the exploration of gradient-based rule learning within neural network frameworks. These innovative approaches enable the application of neural network learning techniques to the realm of rule learning. In this work, we propose RB-Net, a neural network-based rule-generating interpretable model. We demonstrate its competitive performance in classification tasks compared to similar algorithms, particularly excelling when the number of rules is limited.