The AIM-3D team, Valentina Kuskova, Dmitry Zaytsev, and Michael Coppedge, published their paper, “Function-Valued Causal Influence in Nonlinear Time Series,” in the Proceedings of Machine Learning Research (PMLR) as part of the 43rd International Conference on Machine Learning (ICML 2026). ICML is one of the world’s leading and most selective conferences in machine learning, widely recognized alongside NeurIPS as a premier venue for research in artificial intelligence and machine learning.
The paper develops a new approach to understanding causal relationships learned by nonlinear machine-learning models. Rather than representing causal influence with a single numerical score, the authors model it as a function that can vary across different states and contexts. Using Neural Additive Vector Autoregression (NAVAR) and Individual Conditional Expectation, the approach makes it possible to identify nonlinear patterns such as threshold effects, saturation, and changes in the direction of an effect.
The study demonstrates the approach through synthetic experiments and an application to democratic development, where function-valued analysis reveals regime-specific and asymmetric causal relationships that conventional scalar measures can miss. The research contributes directly to AIM-3D’s broader goal of developing interpretable AI-based methods for modeling, explaining, and forecasting complex processes of democratic development and decline.
