A0419
Title: A dimensionality expansion methodology for loss optimization in cost sensitive problems
Authors: Jorge C-Rella - University of A Coruna (Spain) [presenting]
Ricardo Cao - Universidade da Coruna (Spain)
Juan Vilar Fernandez - Universidade da Coruna (Spain)
Abstract: In all the current cost-sensitive classification models there is an intrinsic loss of information due to a blur in the consideration of the exogenous variable that defines gains/losses. The most extended approaches consider decision rules based on an estimated fraud probability, considering a cost matrix. To tackle this blur, two refinements are introduced. First, a loss function is considered a performance metric, which provides a more realistic measure for the expected results in practice. Second, a new decision space is constructed considering the estimated probability and the exogenous variable influencing the loss. This space permits a more flexible search for the decision region. An algorithm (2-DDR) is proposed to optimize the loss function over the decision space with plenty of freedom. The estimated decision region includes, as particular cases, classical approaches. As a consequence, an improvement is always achieved. This is tested and assessed with a wide range of simulations with varying difficulty and structure, showing the algorithm robustness and the systematic improvement with respect to previous approaches.
