Comparison of classification algorithms using feature selection

In this paper we tackle a classification problem using 4 algorithms: C4.5, JRip, k-NN, SVM Linear. Experiments were conducted using the Divorce dataset, consisting of 170 instances, 54 features, and 2 classes. We performed feature selection using the CFS filter method, obtaining just 4 relevant features. We ran the algorithms 30 times over the train set to obtain the best model, both for the dataset using all attributes and the subset with relevant features. With the best model we ran 30 times over the test set of both datasets and we used the balanced accuracy metric to perform a Wilcoxon test. Results show that, with a significant diference, it is better to use only the relevant features rather than the whole dataset.