Feature generation and contribution comparison for electronic fraud detection.

Area of Science:

• Computer Science
• Data Science
• Financial Technology

Background:

• Modern money transfer services face increasing fraud risks.
• Traditional rule-based fraud detection methods are insufficient.
• Machine learning is widely adopted for fraud detection, but performance varies.

Purpose of the Study:

• To evaluate the effectiveness of existing feature categories in machine learning-based fraud detection.
• To propose and validate a novel feature generation guideline based on account characteristics.
• To improve the accuracy of detecting fraudulent transactions in money transfer services.

Main Methods:

• Utilized real transaction data for analysis.
• Applied various machine learning algorithms to assess feature performance.
• Compared traditional feature categories (recency, frequency, monetary, anomaly detection) with a new guideline.
• Developed a feature generation guideline based on statistical perspectives and characteristics of non-fraudulent accounts.

Main Results:

• Existing feature categories, except monetary, showed poor performance across different machine learning models.
• Anomaly detection features performed the worst among traditional categories.
• Statistical features derived from financial knowledge had limited effectiveness.
• The proposed guideline, focusing on atypical detection characteristics of normal accounts, significantly improved fraud detection.

Conclusions:

• Traditional feature engineering methods for transaction data have limitations in fraud detection.
• A novel approach focusing on the distinct characteristics of normal accounts enhances fraud detection capabilities.
• The proposed method offers a superior alternative to existing techniques for identifying fraudulent money transfers.