alliGATOR : Graph Anomaly Detection through diffusion based Topology Reconstruction

Abstract

Financial crime detection concerning monetary transactions remains a critical challenge in the fight against illicit activities such as fraud and money laundering. In this thesis, we present the alliGATOR model, an unsupervised graph anomaly detection model designed to identify both node and edge anomalies. Our approach is grounded in the notion that node connections are dependent on their attributes. We hypothesised that anomalies can be detected by comparing a graph's original topology to one reconstructed from node attributes alone.

The alliGATOR model requires a novel node-guided topology reconstructor model - a discrete generative diffusion model based on a graph neural network link predictor. The model iteratively reconstructs the graph structure by adding edges conditioned on the node attributes and degrees, generating a topology that obeys the local properties of the nodes. Our node-guided topology reconstructor operates under an inductive setting and can generalise across graphs.

We evaluate the alliGATOR model on a custom synthetic dataset designed to simulate real-world behaviour, and compare its performance to the Isolation Forest baseline. Experimental results show that alliGATOR is robust and successfully identifies both node and edge anomalies.

This work establishes a foundation for applying generative diffusion models to graph anomaly detection, and it can support researchers, financial institutions, banks, and governments in developing personalised autonomous anomaly detection systems.