Unsupervised anomaly detection in financial transactions

Abstract

The aim of the project is to devise a process which can detect anomalies in vendor ledger entries for Össur hf. Due to the data being unlabeled two unsupervised machine learning model architectures, Isolation forest & Autoencoder, are utilized. Two datasets are extracted from the databases of Össur hf. on which pre-processing and feature engineering is performed to generate useful features. These features are then tested on the ledger entry data which is fed into each model. The models are evaluated through manual review of the anomalies returned by each model, resulting in an iterative model building process with a human-in-the-loop. Additionally, the models are explained using Shapley values in order to gather insights into which features contribute the most to classifying ledger entries as anomalous or regular. Lastly, the models are set up to run on Össur's Azure Machine Learning environment in order to facilitate future model development. The final result being an end-to-end machine learning solution to detect anomalies which can be refined and iterated upon based on manual reviews conducted by auditors within Össur hf.