ADVERSARIAL ATTENTION-BASED VARIATIONAL GRAPH AUTOENCODER FOR FRAUD DETECTION IN ONLINE FINANCIAL TRANSACTION

• Main Author: Alibasyah Wiriaatmadja, Nur
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73208
• Institution: Institut Teknologi Bandung

The significant growth of online financial transactions also raises threats of fraud in transactions, which are done by identity thief to pass authentication as consumer. Fraud transaction can harm both online transaction provider companies and consumers. Several machine learning models for classification have been used for fraud detection and have shown outstanding results. However, as the volume and variety of transaction data expand, it becomes difficult for traditional machine learning algorithms to detect fraud. This research aims to apply the Adversarial Attention-Based Variational Graph Autoencoder (AAVGA) for fraud detection by modeling online transaction data as a graph and use the graph embedding for classification task. AAVGA is one of the most recent autoencoder-based graphs embedding methods, which use Graph Attention Network (GAN) in the encoder to generate latent variable and employs an adversarial strategy to enhance the generalization performance of a graph embedding model. The generated latent variables are then used in various classification and clustering algorithm to detect fraudulent transaction. As the result, our approach obtains perfect accuracy when combined with XGBoost algorithm on dataset with 8000 nodes. Similar result are obtained by direct application of Decision Tree algorithm. Our approach outperforms direct clustering algorithm with predicting all fraudulent transaction correctly. Furthermore, the graph modeling approach shows promising results to be enhanced with broad capabilities of graph deep learning in future works.