Internet fraud transaction detection based on temporal-aware heterogeneous graph oversampling and attention fusion...

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• E-commerce (EC) faces escalating Internet fraud challenges.
• Existing fraud detection methods struggle with complex transaction data and imbalance.

Purpose of the Study:

• To develop an advanced method for detecting Internet fraud in e-commerce transactions.
• To enhance fraud detection accuracy and enable active fraud prevention.

Main Methods:

• Abstracting transaction data into a heterogeneous graph.
• Utilizing Gated Recurrent Unit (GRU) for temporal dynamics and Graph Neural Network (GNN) for topology.
• Implementing an improved Graph-based Synthetic Minority Oversampling Technique (GraphSMOTE) for data imbalance.
• Employing a multi-layer attention mechanism for active fraud prevention.

Main Results:

• THG-OAFN achieved an Area Under the Curve (AUC) of 96.56% on the Amazon dataset, a 7.78% improvement.
• Achieved recall of 95.21% and F1-score of 94.72% on the Amazon dataset.
• On the YelpChi dataset, THG-OAFN outperformed existing GNN models with AUC of 90.43%, recall of 89.51%, and F1-score of 90.31%.

Conclusions:

• THG-OAFN provides a deployable solution for dynamic fraud detection and active defense in e-commerce.
• The proposed method effectively handles data imbalance and captures complex transaction patterns.
• This approach significantly advances the state-of-the-art in graph-based fraud detection.