Researchers introduce a method to steer neural network training by enforcing constraints derived from partial dependence, ensuring that a model's average response to specific features aligns with known functional domain knowledge.

  • The approach addresses the gap in explanation-guided learning by adjusting models to produce explanations faithful to prior knowledge rather than just interpreting existing interactions.
  • It is demonstrated on regression problems, including dynamical systems forecasting, where constrained models outperform unconstrained ones and exhibit greater data efficiency.
  • Interpretations from the constrained models align with user-provided knowledge, whereas those from unconstrained models do not.

This method allows for the creation of more interpretable and data-efficient models by directly incorporating domain expertise into the training process.