bn-fit-modify

Structure learning recovers the DAG (directed acyclic graph) of a Bayesian Network from observational data. pgmpy and similar libraries use constraint-based or score-based algorithms to identify conditional independence relationships and infer edge direction, producing a network topology you can then fit with parameters.

Yes. After structure learning, parameter estimation fits CPDs (conditional probability distributions) to your data using model.fit(data). This estimates the probability tables for each node conditioned on its parents, completing a fully parameterized Bayesian Network ready for inference.

Do-calculus operations implement interventions by removing incoming edges to target nodes, simulating forced variable assignments. This lets you answer causal questions and sample from intervened networks, distinguishing observational inference from counterfactual scenarios.

Observational sampling follows the learned network structure as-is. Interventional sampling applies do-calculus modifications—removing edges and fixing node values—to generate samples that reflect causal interventions rather than passively observing data.

Yes. Structure learning supports constraints that restrict which edges can appear, enforce domain knowledge about forbidden or required relationships, and guide the algorithm toward plausible DAGs that align with your prior beliefs about the system.

Structure learning can run unsupervised on raw data, but incorporating domain constraints—forbidden edges, required relationships, or prior beliefs—improves accuracy and interpretability. pgmpy supports both approaches depending on your confidence in the data.