Shahin Jabbari

When several similarly performing models disagree about the same person, recourse computed for one model can fail on another. ORPM formulates recourse as budget-constrained minimization of the worst-case validity loss over a competing set of models, and solves it by alternating an approximate oracle with multiplicative-weights updates, with proven convergence to an approximate equilibrium when losses are bounded and the oracle's global error is controlled. Across one synthetic and five real datasets, linear, neural, mixed and decision-tree competing sets reach high joint validity in many settings; on COMPAS with five decision trees at budget 5, ORPM reaches 62.5% joint validity versus 46.0% for the RobX-constrained adaptation.

When several similarly performing models disagree about the same person, recourse computed for one model can fail on another. ORPM formulates recourse as budget-constrained minimization of the worst-case validity loss over a competing set of models, and solves it by alternating an approximate oracle with multiplicative-weights updates, with proven convergence to an approximate equilibrium when losses are bounded and the oracle's global error is controlled. Across one synthetic and five real datasets, linear, neural, mixed and decision-tree competing sets reach high joint validity in many settings; on COMPAS with five decision trees at budget 5, ORPM reaches 62.5% joint validity versus 46.0% for the RobX-constrained adaptation.