A general framework for cost-sensitive boosting

Boosting algorithms have been widely used to tackle a plethora of problems. Among them, cost-sensitive classification stands out as one of the scenarios in which Boosting is most frequently applied in practice. In the last few years, a lot of approaches have been proposed in the literature to provide standard AdaBoost with asymmetric capabilities, each with a different focus. However, for the researcher, these algorithms shape a confusing heap with diffuse differences and properties, lacking a unified framework to jointly compare, classify, analyze and discuss the approaches on a common basis. Motivated by the preeminent role of AdaBoost in the Viola-Jones framework for object detection in images, a markedly asymmetric learning problem, in this thesis we try to untangle the different Cost-Sensitive AdaBoost alternatives presented in the literature, demystifying some preconceptions and making novel proposals (Cost- Generalized AdaBoost and AdaBoostDB) with a full theoretical derivation. We try to classify, analyze, compare and discuss this family of algorithms in order to build a general framework unifying them. Our final goal is, thus, being able to find a definitive scheme to translate any cost-sensitive learning problem to the AdaBoost framework while shedding light on which algorithm ensures the best performance and formal guarantees.