What is structure learning?
Structure learning is the task of learning the shape of a model from data rather than only its parameters. In a Bayesian network it means discovering which variables depend on which, that is, where the arrows go. In graph learning more broadly it means inferring which entities are connected and how strongly, when the edges are not given. The output is a graph that people can inspect, which separates it from most machine learning, where the structure is fixed in advance.
Two families of algorithms exist. Constraint-based methods run statistical independence tests between pairs of variables and remove edges that the tests rule out. Score-based methods propose candidate graphs, score each by how well it explains the data while penalizing complexity, and search for the best. Both benefit from expert constraints, such as forbidding an arrow from an effect back to its cause.
Structure learning in practice
A specialty chemicals plant records several hundred process variables every minute, but nobody knows the full web of dependencies between them. Structure learning over a year of data proposes a dependency graph, which process engineers prune and correct. The resulting Bayesian network supports root cause analysis when a batch goes out of specification. Treat learned edges as correlation until confirmed, because several different graphs can explain the same data equally well.