Rademacher analysis of infimum classifiers
Amiran Ambroladze, Emilio Parrado-Hernandez and John Shawe-Taylor
Unpublished, Southampton, UK.
This paper addresses the problem of analysing the performance of classifiers obtained as the infimum of a set of k weak learners. The main result consists in a bound on the error of these classifiers based on the Rademacher complexity of the weak learners.
This theory motivates the development of a classification algorithm called Soft-output Set Cover Machine as an extension of the Set Cover Machine that forms the classifier as a combination of real valued features. The goodness of the Soft-output Set Cover Machine is benchmarked in several real data problems, in comparison with the original Set Cover Machine.