Structural SVMs for Defining Payment Rules in Combinatorial Auctions

A major challenge in combinatorial auctions is to construct payment rules to make a winner-determination algorithm maximally incentive compatible. A general purpose solution is available only when winner case for reasons of computational complexity or other economic considerations. We propose to use structural SVMs for this pricing problem. The decision boundaries for each target class (with a class corresponding to the allocation of a bundle to an agent) combine to define a payment rule. An exact classifier produces a payment rule that makes the auction incentive compatible. Minimizing regularized empirical risk in solving the structural SVM training problem minimizes a regularized upper bound on empirical regret.