Joint Use of Time Series and Graph Data for Fake Comment Detection in CafeBazaar Dataset

Fake comments pose a significant challenge for e-commerce, as they manipulate customer perception and harm legitimate businesses. We propose a novel approach using recurrent neural networks to detect anomalies in user behavior patterns. We model user activity on Cafebazaar e-commerce as a multivariate time series and employ a long short-term memory autoencoder to determine periods of abnormal activity. By incorporating our model with existing methods, we observe significant improvement in the detection accuracy of fraudulent users, reaching 99 percent precision with the dataset. Our proposed model can enhance other fake comment detection methods, particularly those that struggle with the cold start problem, which often lacks information about the majority of users. The proposed system has been successfully implemented in a real-life application, removing 1.9 million fake comments from the Cafebazaar platform. We also provide a new dataset for fake comment detection that researchers can use to evaluate existing methods or develop new approaches.