Research Assistant (Nserc)
Toronto, Ontario, Canada
Worked with Fidelity Canada on a comparative research of detecting anomalous financial journal entries using machine learning and deep learning techniques including logistic regression (LGR), support vector machine (SVM), random forest (RF), deep neural network (DNN) and deep autoencoder network (DAE). The goal was to develop an automated system to identify fraudulent/abnormal transactions in advance and therefore, reduce the risk. I conducted my research using tremendous real-world production data. Preprocessed over two million data records using techniques like clustering, one-hot encoding, normalization, etc. Employed the undersampling technique to alleviate the imbalance in the dataset, in which only about 2% of records are positive. The research showed that a compact 2-layer DNN had the best performance in our task. RF was the only state-of-the-art model that has comparable performance. However, compared to RF, DNN supported online learning better which made it more suitable to be deployed in the production environment. The final fine-tuned DNN had an 0.94 F-score and 0.99 ROC AUC on the validation set. The unsupervised DAE could achieve 75% validation accuracy. Presented a strategy to balance the precision-recall trade-off in different production scenarios. Based on the cost ratio of a false alarm versus a missing anomaly, an optimization problem to minimize the total cost could be solved to select the best decision threshold.
