Search

In real-world databases sometimes one of the classes is more difficult to classify then the others. This can happen, for instance, when one of the class is much bigger than the other. To cope with this problem, this paper proposes to assign specific weights to each class in the model evaluation criterion of the feature selection algorithm. The proposed technique is applied to a real world classification problem: the prediction of bankruptcy. The data set used in this study has missing values and extreme values. The data set also presents a much smaller bankruptcy class than the not bankruptcy class. Feature selection is used to choose the number of relevant features, using the so called correctly classified criterion. In the case study, less features are selected using this criterion, consequently less computational time is taken. Moreover, this paper compares five different fuzzy clustering algorithms in terms of model accuracy and computational burden. These clustering algorithms are also used and compared during the feature selection procedure. For this bankruptcy data set, the classification rate with only 16 features is 80% for the companies that bankrupt, whereas the percentage of companies that do not bankrupt is 95.4%. If we take in consideration that the number of obtained rules is 6, only 16 features are used, as opposed to 70 rules and 35 features as in, where the accuracy is 81%, with all of the interpretability issues that this carries, then it can be considered that these are promising results. more...