Explanation-Guided Training Steers Models Toward Faithful Partial-Dependence Explanations

arXiv·low signal

Most interpretability work develops techniques to explain a trained model's learned interactions, but far less focuses on adjusting the model so its explanations stay faithful to prior knowledge — explanation-guided learning. This paper steers neural-network training using interpretable constraints based on partial-dependence functions, encoding prior knowledge directly into the training objective. Aimed at settings where an explanation being trustworthy matters as much as accuracy.