The AIM-3D team continues to advance the frontier of AI-driven causal discovery and dynamic inference. Valentina Kuskova presented new research at the 39th International Florida Artificial Intelligence Research Society Conference (FLAIRS-39), introducing a novel framework for distinguishing genuinely important causal relationships from those that merely appear influential in complex nonlinear AI models.
The presentation, titled “Beyond Coefficients: Forecast-Necessity Testing for Interpretable Causal Discovery in Nonlinear Time-Series Models,” addresses a growing challenge in artificial intelligence: determining whether a relationship identified by a neural network is truly necessary for prediction or simply receives a high importance score due to statistical artifacts such as persistence or redundancy.
Developed by Valentina Kuskova, Dmitry Zaytsev, and Michael Coppedge, the research introduces a new formulation of the concept called forecast necessity, which evaluates causal relevance by measuring how much predictive performance deteriorates when a specific relationship is removed from a trained model. Rather than relying solely on model coefficients or importance scores, the framework uses edge ablation combined with Diebold-Mariano statistical testing to determine whether a relationship is indispensable for forecasting.
The study applies the methodology to a panel of 139 countries using democracy indicators from the V-Dem dataset spanning 1990–2024. Results demonstrate that relationships with larger neural causal scores are not necessarily the most important for forecasting. In one example, a predictor with a smaller causal score proved essential for prediction, while a higher-scoring predictor was found to be largely redundant.
According to the researchers, the findings highlight the importance of shifting from parameter-based interpretations toward behavior-based evaluations in nonlinear machine-learning systems. The work has implications for AI interpretability, causal discovery, social science forecasting, and other domains where decisions may depend on understanding the drivers of complex dynamic systems.
The presentation is part of the broader AIM-3D research program at the University of Notre Dame, which applies advanced artificial intelligence methods to study democratic development, political stability, and institutional change. The project’s recent publication pipeline includes forthcoming papers at top computer science conferences, KDD 2026 and ICML 2026, reflecting growing interest in combining causal inference and machine learning for the analysis of complex social systems.
