Real-time anomaly detection using deep learning in e-commerce platform

As e-commerce systems are inundated with fast increasing fraudulent and anomalous activity, real-time anomaly detection provides security and operational efficiency for e-commerce systems. This study investigates real-time anomaly detection approaches that use deep learning (DL) to enhance the security and operational efficiency of e-commerce systems. The proposed framework consists of Autoencoders(AEs), Long Short-Term Memory(LSTM) models and Graph Neural Networks(GNN) taking advantages of their functionalities: AEs learn to compressed representation of what counts as ‘normal’ behaviour (in addition to determining the outliers), LSTM learn the causal/temporal dependencies among sequences, and GNN can model the complicated relational structure of behaviors on the users, transactions, and items. The ensemble uses the outcasts of all three types of models and uses an extended gradient boosting classifier to produce aggregated anomaly scores. The ensemble was evaluated on the Kaggle Online Retail Dataset and produced the highest Service Quality evaluation metrics of accuracy - 96%, precision - 92%, recall - 89%, and F1-Score - 90%, for the ensemble model against the other individual models evaluated on all measures. This ultimately provides evidence for the complementary and robust solutions an integrated framework represents when potentially combining models designed for representation learning, temporal learning, and relational learning. Furthermore, the framework can be implemented with low latency, scalability and be deployed for real-time detection in the cloud using streaming delivery services. Overall, the implemented framework represents a compelling method for detecting fraudulent and anomalous activity in e-commerce systems that can be adapted as e-commerce systems evolve in variability and malicious behavior patterns.