Reducing Dimensionality of Variables for a Classification Problem: Fraud Detection

Abstract

Fraud detection can be considered as a classification task since we can use datasets with labelled instances as fraud cases and legal cases. Although, many classifiers were applied to this problem, the data pre-processing related to the reduction of values of each variable is an uncommon approach. We explore a method to reduce the cardinality of the variables in a dataset of fraud transaction to identify improvement in this classification problem. Our best result indicated an improvement of $+$ 31.8% in terms of F1-measure when we reduce the cardinality to detect fraud cases.