RLENN: A Rule Loss Embedded Neural Network Framework for Knowledge Discovery and Integration

In the era of booming artificial intelligence, data is crucial in driving research innovation with its core value in underlying scientific knowledge. Machine learning (ML)-based knowledge discovery and integration have seen complementary development. However, there are still challenges in mining and integrating knowledge without prior domain knowledge guidance, as well as the problem of lack of physical consistency in neural network models. To address these, we propose a Rule Loss Embedded Neural Network (RLENN) framework in this paper. By introducing rule information into the loss function during the model training stage, the model's adaptability is enhanced. Specifically, explicit rules are extracted from the data by subgroup discovery, providing a source of knowledge for the subsequent construction of rule loss. Then, these rules are converted into rule loss and incorporated into the loss function of the neural network to guide model learning. The experiments on two datasets demonstrate the effectiveness of our proposed framework.