Feature Importance in Credit Card Fraud Detection: A Comparative Analysis of Location and Temporal Features

Abstract

Fraud detection in financial transactions is a critical challenge requiring robust machine learning techniques. In order to identify fraudulent credit card activity, this study assesses models such as Long Short-Term Memory (LSTM) networks, Random Forest, Decision Trees, and Logistic Regression. Key features such as transaction location (lat, long, merch_lat, merch_long), merchant details (zip, distance), and temporal data (unix_time) were crucial in identifying fraud patterns. Experimental results showed that tree-based models, particularly Random Forest, achieved superior performance with 99.94% accuracy, while LSTM effectively captured sequential data patterns. Random Forest’s ability to handle feature interactions and imbalances made it the most reliable. Analysis of ROC curves highlighted models’ learning behavior and generalization. This research emphasizes integrating spatial and temporal features to advance adaptive, real-time fraud prevention systems.