Steering Neural Network Training through Interpretable Constraints Based on Partial Dependence
2026-07-09 • Machine Learning
Machine Learning
AI summaryⓘ
The authors explore how to teach neural networks to make predictions while following rules or knowledge that people already understand about a problem, especially for predicting numbers (regression). They introduce a way to guide the model so its average behavior with certain input features matches known real-world relationships. Their experiments show that models trained this way not only make better predictions but also need less data and give explanations that fit the prior knowledge better than normal models. This approach is different from most prior work that mostly focuses on classification and assumes exact important features beforehand.
Neural networksInterpretabilityPartial dependenceRegressionExplanation-guided learningDynamical systems forecastingModel alignmentData efficiency
Authors
Yann Claes, Pierre Geurts, Vân Anh Huynh-Thu
Abstract
Over the last few years, there has been an increased interest in making machine learning models more interpretable. Although a great deal of effort goes into developing techniques for interpreting the interactions learned by a given model, fewer studies focus on assessing the quality of such explanations. Even fewer focus on how to adjust the model to produce explanations faithful to prior knowledge, a process known as explanation-guided learning. Furthermore, most approaches in this area focus on classification problems and usually assume prior knowledge about which input features or regions are most important. In this work, we introduce a new approach to steering neural networks based on partial dependence, such that their average response to certain features aligns with specific functional domain knowledge about the problem. We empirically demonstrate on a range of regression problems, including dynamical systems forecasting, that models whose training has been controlled using our method perform better than unconstrained models and are more data-efficient. Moreover, we highlight that interpretations obtained from the former actually align with the user-provided knowledge, whereas those obtained from the latter do not.