Impact of Feature Selection Methods on the Perfromance of Credit Risk Classification Algorithms

This paper presents ensembling of filter features selection algorithms for classification problem in the context of assessment of risk of credit for a financial institution. Feature selection is one of the most important aspect of data mining, and machine learning algorithm. The main objective of feature selection is reduction in computing resources, reduction in future data collection cost, reducing complexities of the model, avoiding overfitting, and increasing the performance of machine learning algorithms. In this paper the set of available variables are firstly reduced using filter feature selection methods such as chi- square, gain ratio, information gain, relief F, and symmetric uncertainly. In addition, ensemble feature selection of the input variables based on these individual methods is also used. The impact of feature selection is measured by fitting seven classification algorithms, i.e., Random Forest, C4.5, PART, C5.0, Bagging, Boosting, and SVM Linear. The performance of the models is compared by calculating parameters such as accuracy, sensitivity, specificity, positive predictive values, negatively predictive values, and AUC. The data used is German bank data of 1000 records and 20 features and one target variable.