Fraud Detection in Non-Network Knowledge Graph

Fraud is becoming an increasingly severe problem in all industries, and one of the most widely-used techniques in fraud detection is data mining. However, most of the existing research revolve around credit card frauds, and studies on fraud detection in other fields are still extremely limited. Therefore, in this thesis, we develop an automated fraud detection system for both reporting and prediction in Contract Management, and use the contract data from Defence Construction Canada (DCC) to evaluate the approaches employed. As we lack practical training data, we use a weak-supervision approach to generate labels for our data and build two machine learning models, Logistic Regression and Random Forest. We also propose a graph-based approach that transforms our dataset into graphical representations, which results in a non-network knowledge graph, and learns both structural and statistical features of this graph to identify potential anomalies. Results from both approaches reveal relatively high recall.