RL-Net: Interpretable Rule Learning with Neural Networks

As there is a need for interpretable classification models in many application domains, symbolic, interpretable classification models have been studied for many years in the literature. Rule-based models are an important class of such models. However, most of the common algorithms for learning rule-based models rely on heuristic search strategies developed for specific rule-learning settings. These search strategies are very different from those used in neural forms of machine learning, where gradient-based approaches are used. Attempting to combine neural and symbolic machine learning, recent studies have therefore explored gradient-based rule learning using neural network architectures. These new proposals make it possible to apply approaches for learning neural networks to rule learning. However, these past studies focus on unordered rule sets for classification tasks, while many common rule-learning algorithms learn rule sets with an order. In this work, we propose RL-Net, an approach for learning ordered rule lists based on neural networks. We demonstrate that the performance we obtain on classification tasks is similar to the state-of-the-art algorithms for rule learning in binary and multi-class classification settings. Moreover, we show that our model can easily be adapted to multi-label learning tasks