Anomaly detection using autoencoders with network analysis features

Fraudulent activity within a financial ecosystem often involves the coordinated efforts of several bad actors. Expressing the interactions between participants in a system as a mathematical graph allows researchers to apply social network analysis to understand the nature of these relationships better. This article proposes and extends a unified approach using an autoencoder to detect anomalies in a transactional setting. The methodology begins with a neural architecture search to determine a best autoencoder model architecture configuration. This is followed by a threshold optimisation process to find a reconstruction error that best discriminates between normal and anomalous classes. Gaussian scaling is applied to the raw anomaly scores in order to represent the output in an interpretable and universally transferable form. The unified approach is extended by selecting and including network metrics as features, for the purposes of producing a model that can detect anomalies from both standard transactional data and network data representing the relationships between users within a financial system. Applying SHAP on the model output highlighted the strongest contributing or offsetting network metric features for all anomalies detected. The PageRank and degree centrality network metrics were most significant in detecting anomalous instances within the data. Including network metrics in the feature space generated encouraging model performance results, leading to a potential low operational cost of fraud.