• Título:  COST-SENSITIVE CLASSIFICATION BASED ON BREGMAN DIVERGENCES
• Autor:  Santos Rodríguez, Raúl
• Universidad:  Universidad Carlos III de Madrid
• Departamento:  TEORIA DE LA SEÑAL Y DE LAS TELECOMUNICACIONES
• Fecha de Lectura:  11/07/2011
• Programa de doctorado:  Doctorado Interuniversitario en Multimedia y Comunicaciones
• Dirección: 
  • CID SUEIRO, JESÚS (DIRECTOR)
• Tribunal: 
  • ARTES RODRIGUEZ, ANTONIO (PRESIDENTE)
  • PARRADO HERNANDEZ, EMILIO (SECRETARIO)
  • ALAIZ RODRIGUEZ, ROCÍO (VOCAL)
  • ROJO ALVAREZ, JOSÉ LUIS (VOCAL)
  • DE BIE, TIJL (VOCAL)
• Descriptores: 
  • TEORIA Y PROCESOS DE DECISION
  • ANALISIS DE DATOS
• Localización:  E-ARCHIVO, REPOSITORIO INSTITUCIONAL DE LA UNIVERSIDAD CARLOS III DE MADRID
• http://hdl.handle.net/10016/12508
• Resumen:
  The main object of this PhD. Thesis is the definition of new loss functions to address the so-called cost-sensitive classification. Many (likely, most) decision problems are cost-sensitive: if there is a (human or machine) actuator after a decision maker, the consequences of the actions taken after any decision may depend on the decision itself, on the class of the data, on the values of the observations and even on other unobserved factors that may be unpredictable before deciding. The dominating preference for cost-insensitive methods in the machine learning literature is likely a natural consequence of the fact that true costs in real applications are difficult to evaluate, and cost information is usually not available in benchmark databases. Since, in general, guessing correctly the class of the data is less costly than any decision error, designing low error decision systems is a reasonable (but suboptimal) approach. Our proposal relies on Bayes decision theory where the goal is to assign instances to the class with minimum expected cost. The decision is made involving both costs and posterior probabilities of the classes. Obtaining calibrated probability estimates at the classifier output requires a suitable learning machine, a large enough representative data set as well as an adequate loss function to be minimized during learning. The design of the loss function can be aided by the costs: classical decision theory shows that cost matrices define class boundaries determined by posterior class probability estimates. Strictly speaking, in order to make optimal decisions, accurate probability estimates are only required near the decision boundaries. This leads to probability estimates that are more robust against noisy data. It is key to point out that the election of the loss function become especially relevant when the prior knowledge about the problem is limited or the available train examples are somehow unsuitable (i.e., applications where the high-cost examples are scarce). In those cases, different loss functions lead to dramatically different posterior probabilities estimates. The losses we study belong to the family of Bregman divergences. Bregman divergences offer a rich family of proper losses that have recently become very popular. In this dissertation we consider the application of Bregman divergences as loss functions to generate finely tuned posterior probability estimates in different cost-sensitive scenarios, namely, supervised and semi-supervised learning with both deterministic and example-dependent costs.