Playing the Large Margin Preference Game

Mirko Polato, Guglielmo Faggioli, Ivano Lauriola, Fabio Aiolli

Abstract: We propose a large margin preference learning model based on game theory to solve the label ranking problem. Specifically, we show the proposed formulation is able to perform multiclass classification by solving a single convex optimization problem. Generally, such formulation, although theoretically well-founded, requires to learn a large number of parameters. To reduce the computational complexity, we propose a strategy based on the solution of smaller subproblems, that can be further optimized by exploiting techniques borrowed from multi-armed bandits literature. Finally, we show how the proposed framework exhibits state-of-the-art results on many benchmark datasets.