Researchers propose GraphBU, a generator for mixed-integer linear programming (MILP) instances that uses local subproblems and their interfaces as its basic unit. This approach promotes coupling nodes into master constraints or boundary variables to enable compatibility-checked replacement while preserving the structure solvers rely on.
- The method ensures graph construction invariance to row-column permutations and preserves feasibility under an interface-slack condition.
- Generated instances maintain high graph-statistical similarity (approximately 0.934) and feasibility (approximately 96.7%) relative to source families.
- The generated data improves downstream Predict-and-Search training, yielding an average increase of approximately 8.0% in the main index.
GraphBU addresses the difficulty of obtaining MILP instances from private pipelines by explicitly recording how local parts of an instance couple to the rest.