Causal Discovery on Sequential Data with HiBaNG
Our work addressing causal discovery on sequential data is out at TMLR 2026:
HiBaNG: Hierarchical Bayesian Nonparametric Granger Causal Discovery in Low-Data Regimes
Figuring out which variables in a time series (e.g., climate signals) causally influence each other is incredibly hard and may require models that need lots of data (among other challenges), which isn’t always available. This paper introduces HiBaNG, a statistical method built for exactly these small-data situations: it uses a Bayesian approach to both detect likely cause-and-effect links and report how confident it is about each one. Tested on synthetic, semi-synthetic, and real climate data, it beats both classical and deep-learning baselines while giving well-calibrated uncertainty estimates.