Small dataset augmentation with radial basis function approximation for causal discovery using constraint-based method

Abstract

Causal analysis involves analysis and discovery. We consider causal discovery, which implies learning and discovering causal structures from available data, owing to the significance of interpreting causal relationships in various fields. Research on causal discovery has been primarily focused on constraint- and score-based interpretable methods rather than on methods based on complex deep learning models. However, identifying causal relationships in real-world datasets remains challenging. Numerous studies have been conducted using small datasets with established ground truths. Moreover, constraint-based methods are based on conditional independence tests. However, such tests have a lower statistical power when applied to small datasets. To solve the small sample size problem, we propose a model that generates a continuous function from available samples using radial basis function approximation. We address the problem by extracting data from the generated continuous function and evaluate the proposed method on both real and synthetic datasets generated by structural equation modeling. The proposed method outperforms constraint-based methods using only small datasets.