Automações não-supervisionadas na abordagem de seleção dinâmica de atributos baseada na fronteira de pareto

Abstract

Many feature selection strategies have been developed in the past decades, using different criteria to select the most relevant features. The use of dynamic feature selection, however, has shown that using multiple criteria simultaneously to determine the best subset of features for similar instances can provide encouraging results. Although the use of dynamic selection has alleviated some of the limitations found in traditional selection methods, the exclusive use of supervised evaluation criteria and the manual definition of the amount of groups to be used, lead to limitations of complex problem analysis in unsupervised settings. In this context, this thesis proposes three strands of the dynamic feature selection approach based on the pareto frontier, in the preprocessing context and one strand in the classification context. The first is related to the inclusion of unsupervised criteria in the base version of PF-DFS/M. The second (PF-DFS/P) and third (PF-DFS/A) strands are variations of the base version, where they include, respectively, partial and full automation of the definition of the number of groups to be used in the preprocessing process through the use of an internal validation index committee. The automation of the hyperparameter concerning the number of groups allows, instead of arbitrary choice, mechanisms to be used that can help researchers deal with unlabeled databases, or even constitute an analysis under labeled databases. The last strand proposes the use of a dynamic clustering weighting mechanism to allow that instead of considering only one group of features to train classifiers and test instances, each instance can select a portion of features based on the proportion of similarity to all feature groups. Both real and artificial datasets were used in the investigative analyses. The results found in the empirical analyses employed in this thesis are promising, demonstrating that PF-DFS, with partial and complete automation of the definition of the number of groups to be used and the use of dynamic clustering weighting strategies, can obtain superior results to the feature selection methods used as a comparative basis, as well as when compared to the original dataset.