Reproduce “NIMO: a Nonlinear Interpretable MOdel” on OpenResearch

About this paper

Shijian Xu, Marcello Massimo Negri, Volker Roth

NIMO introduces a nonlinear interpretable model that combines the expressive power of neural networks with the inherent interpretability of linear models from the University of Basel. It consistently outperforms various baselines on synthetic datasets and maintains competitive predictive accuracy on real-world data while providing clear, intrinsic global and local explanations.

Key insights

• Structuring a nonlinear model to ensure its global linear coefficients exactly represent Marginal Effects at the Mean enables inherent interpretability without relying on post-hoc approximations.
• Masking the primary feature from its own nonlinear correction term (`x_j` from `g_u_j(x_minus_j)`) and using positional encoding are crucial architectural choices for maintaining interpretable marginal effects and avoiding confounding.
• Parameter elimination combined with adaptive ridge regression offers a robust and efficient optimization method for models with entangled linear and nonlinear components, facilitating sparsity and interpretability.

Summary by alphaXiv.