Implementation of Point-Based Scoring Model for Fraud Detection

This study examined explainability in machine learning, specifically, poor parsimony in fraud detection models. Financial services organizations, especially, are subject to regulations requiring transparency in automated decision-making. The purpose of this study was to compare traditional variable selection techniques with a discriminative, parsimonious point-based scoring model for e-commerce fraud detection, a new application for this model. Data were from a secondary dataset containing real-world e-commerce transactions. Four models were compared: the point-based scoring model (AutoScore), random forest, LASSO, and logistic regression. The point-based scoring model, AutoScore, was the simplest out of the four models with thirty-seven variables. Still, AutoScore also achieved the lowest F1 score (0.2066), which was comparable to logistic regression and LASSO. The most discriminating model was the hyperparametertuned random forest, with the highest F1 score (0.30577) with 91 variables. Future practice should explore how AutoScore can be used as a supplementary or surrogate explanatory model. Besides such future practices, future research can explore if other model types can be applied to the point-based scoring framework.