Sometimes you don’t know which risks drive which. Structure learning tries to read that shape from the data — while being honest about what data alone can never tell you.
So far the diagram — which factor points to which — was drawn by hand from domain knowledge. Often that is right and proper. But sometimes the relationships are exactly what you’re unsure of: does the design maturity drive the rework, or do both simply track an aggressive schedule? Structure learning asks whether the data itself can suggest the graph, rather than assuming it.
1Scoring a candidate structure
The core move is to define a score for any proposed structure and then prefer high-scoring ones. A good score balances two forces in tension. It rewards fit — how well the structure explains the observed data — and it penalises complexity — how many links and parameters the structure demands. The penalty is essential: a fully connected graph can always fit the data best, yet it memorises noise and predicts the future worst. Standard scores (the Bayesian score, or the closely related information-criterion penalties) formalise this trade-off.
2Searching an astronomical space
You cannot score every possible structure — the number of valid graphs grows faster than exponentially in the number of variables. Instead you search: start somewhere, then repeatedly try small local edits — add an edge, remove an edge, reverse an edge — and keep any change that improves the score. It’s a greedy climb toward a good-enough structure, not a guaranteed-best one, and that pragmatic trade is exactly why it’s usable on real datasets.
3The limit you must respect: Markov equivalence
Here is the caution that matters most for risk. Several different diagrams can encode identical statistical relationships. From observational data alone, “A causes B” and “B causes A” can be completely indistinguishable — both predict the same correlations. Such indistinguishable structures form a Markov equivalence class. Learning can recover the skeleton (which factors are linked) and orient some arrows, but it cannot, in general, tell you the direction of every link without either a controlled intervention or genuine domain knowledge.
Data can propose the wiring of a risk model and prune the implausible, but it cannot always tell you which way the influence runs. The honest output of structure learning is often a shape with some arrows left deliberately undirected.
Use data to challenge your assumed risk drivers — you may find links you never registered, or discover that two “separate” risks are really one. But treat any causal claim the data hands you with discipline: if the direction of influence would change your decision, you probably need a deliberate test or hard domain reasoning to establish it, not another correlation. Structure learning is a superb way to generate hypotheses about what drives your risk, and a poor way to prove them.