Bayesian optimization driven strategy for detecting credit card fraud with Extremely Randomized Trees.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Rising credit card usage increases fraud risks.
• Conventional fraud detection models struggle with complex, imbalanced data, leading to overfitting.
• Advanced machine learning is crucial for effective credit card fraud detection (CCFD).

Purpose of the Study:

• To propose a novel Bayesian-optimized Extremely Randomized Trees (TP-ERT) model for enhanced credit card fraud detection.
• To improve model generalization and capture diverse transaction patterns.
• To evaluate TP-ERT's performance against existing CCFD systems.

Main Methods:

• Utilized Extremely Randomized Trees with enhanced randomness in split points and feature selection.
• Employed Tree-structured Parzen Estimator (TPE), a Bayesian optimization strategy, for hyperparameter tuning.
• Assessed model performance on a real-world credit card transaction dataset.

Main Results:

• The proposed TP-ERT model demonstrated superior performance over other CCFD systems.
• TP-ERT achieved a higher F1 score, validating its effectiveness.
• Bayesian optimization using TPE proved more efficient than other techniques for hyperparameter tuning.

Conclusions:

• The optimized Extremely Randomized Trees model is a viable artificial intelligence tool for detecting credit card fraud.
• Hyperparameter tuning via Tree-structured Parzen Estimator effectively enhances model performance by capturing intricate transaction patterns.
• Empirical findings confirm the proposed approach's superiority on real-world data.