Extensões do algoritmo Ant-Miner para tratar o problema de bases de dados desbalanceadas

Abstract

The classification task in Data Mining uses algorithms called classifiers to find patterns on data bases. Real data bases can have an imbalance in its classes, when there are more cases of one class then the others. Classification algorithms are sensitive of this imbalance and tend to valorize the majority class and ignore de minority class, because the cases of minority class have low representation on the training set. It is a problem when the minority class is the class of interest. In this work we propose two extensions to the Ant-Miner algorithm to find better rules to the minority classes. These extensions modify, mainly, how rules are constructed and evaluated. The Ant-Miner algorithm is based on ACO (Ant Colony Optimization). The original version and others related works showed that the Ant-Miner is competitive with other standard classifiers. Moreover, we analyzed sampling techniques (undersampling and oversampling) and also a study of ROC (Receiver Operating Characteristics) analysis. The sampling techniques aim to make a new sampling of the data sets changing the classes? proportion of the training set. The ROC analysis can evaluate the results with more accurate than other metrics, mainly when the classifiers are applied in data sets with classes imbalance. Experimental results showed that the developed algorithms contribute to the rule discovery of the minority classes and also contribute with the simplicity of the rules.