Bayesian Networks (BNs) are widely used for modelling uncertainty and causal relationships in domains such as healthcare, finance, cyber security and decision support. However, learning the optimal BN structure directly from observational data remains computationally challenging due to the super-exponential search space of possible graphs.

This project proposes a novel self-supervised deep learning framework for Bayesian Network structure learning. Instead of learning directly from real datasets, the framework generates a large synthetic training corpus from expert-designed Bayesian Networks. Each original network is transformed into multiple masked variants by introducing small structural perturbations, while corresponding datasets are also masked using modern representation learning techniques. A deep neural network is then trained to recover or predict the relationship between masked datasets and their associated masked graph structures.

The learned model aims to capture general structural patterns that traditional search algorithms cannot easily exploit, potentially providing fast and accurate BN structure estimation for unseen datasets.