Bayesian Tabular Few-shot Learning with Causal Information

Recently, the task of tabular few-shot prediction has received considerable attention, with most existing methods incorporating contextual information using language models. In contrast, we present a purely Bayesian method for tabular few-shot prediction based on providing a probability distribution over possible causal structures as an additional model input. The model then performs Bayesian inference, weighing different data generation hypotheses by considering both their likelihood of generating the provided data and their compatibility with the provided distribution over causal structures. In our evaluations, we find that our method significantly increases model performance even on real world data when suitable causal information is provided.