RWR-RGCN : A Novel Framework for Fraud Detection via Node Context Aggregation

TMLR Paper6668 Authors

Abstract: The integrity of online reviews is crucial for businesses, yet widespread review fraud poses significant risks. This paper addresses this challenge by leveraging the power of multi-relational graph convolutional networks (RGCNs) for fraud detection. We introduce RWR-RGCN, a novel framework integrating a multi-layer RGCN architecture with Random Walks with Restart (RWR). The essential role of capturing critical connections lies in RWR generating node sequences, which can aggregate node features, enhancing the model's understanding of the local and global context within the review graph. To further refine fraud detection, we incorporate Louvain clustering for community identification, identifying high-modularity clusters indicative of coordinated fraudulent activity. Evaluated on the Yelp dataset, RWR-RGCN achieved an AUC of 82.58\% and a recall of 94.56\%, surpassing the state-of-the-art and baseline methods in AUC and recall. These results demonstrate the superior effectiveness of the proposed framework in detecting fraudulent activity within complex online review networks.

Submission Type: Long submission (more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=g9RxpOB7De

Changes Since Last Submission: Our Last Desk Reject Comments was "Incorrect format, e.g. font. Please revise, ensuring adherence to format, and resubmit" then we reviewed the format again and submit this time with latex template.

Assigned Action Editor: ~Chuxu_Zhang2

Submission Number: 6668