Reproduce “Interpretable Models for Granger Causality Using Self-explaining Neural Networks” on OpenResearch

About this paper

Ričards Marcinkevičs, Julia E. Vogt

Generalized vector autoregression (GVAR) combines nonlinear neural forecasting with self-explaining coefficients whose magnitudes, signs, and temporal changes can be inspected directly. On synthetic Lorenz 96, fMRI, and Lotka–Volterra systems, it recovered causal graphs competitively and identified ecological interaction directions with 0.961 balanced accuracy, including 0.999 for negative effects.

Key insights

• The learned coefficients are part of the prediction equation rather than post-hoc attribution scores, allowing their signs to represent increasing or decreasing predictive effects under the model’s definition.
• In the Lotka–Volterra setting, coefficient trajectories visibly separated causal from noncausal links and recovered positive prey-to-predator and negative predator-to-prey interactions.
• Time-reversal thresholding can provide a data-driven graph-selection heuristic, but its justification for nonlinear processes is empirical rather than a formal significance test.

Summary by alphaXiv.