Avaliação de medidas de similaridade de matrizes kernel aplicadas em classicadores de larga margem para seleção de modelos

Abstract

The purpose of this work is to investigate the behavior of similarity measurements, i.e., Kernel Target Alignment (KTA) and Feature Space-based Kernel Matrix Evaluation Measure (FSM) in relation to their correlation with a large margin classifier - support vector machine, in order to propose and implement a model selection method, constructed by means of two steps: a hyper-parameter selection model and a model for feature selection. The KTA and FSM methods indicate the degree of similarity between kernel matrices determined by an alignment measure. This value of alignment is used as reference for a wrapper model selection construction using the simulated anneling as optimizer. Initial tests are depicted to verify the similarity measurements performance in relation to a large margin classi er aiming to identify the better measure to be adopted in the proposed selection model. Following, the described selection model components are tested separately and their results are exhaustively analyzed.