On Adaboost and optimal betting strategies

Malacaria, F. Smeraldi, in: Proc. of the 5th International Conference on Data Mining (DMIN/WORLDCOMP), CSREA Press, pp 326-332, July 2009

Abstract

We explore the relation between the Adaboost weight update procedure and Kelly's theory of betting. Specifically, we show that Adaboost can be seen as an intuitive optimal betting strategy in the sense of Kelly. Technically this is achieved as the solution of the dual of the classical formulation of the Adaboost minimisation problem. Using this equivalence with betting strategies we derive a substantial simplification of Adaboost and of a related family of boosting algorithms. Besides allowing a natural derivation of the simplification, our approach establishes a connection between Adaboost and Information Theory that we believe may cast a new light on some boosting concepts.

Full paper (PDF)

Backlinks: Publications , Home