Local interpretable spammer detection model with multi-head graph channel attention network.

Area of Science:

• Computational Social Science
• Artificial Intelligence
• E-commerce Security

Background:

• Online shopping relies heavily on user reviews, but fraudulent reviews from spammers mislead consumers.
• Existing spammer detection methods often function as black boxes, lacking interpretability.
• This opacity hinders trust and the practical application of spam detection technologies.

Purpose of the Study:

• To develop a locally interpretable spammer detection model for online reviews.
• To address the lack of transparency in current black-box spam detection systems.
• To provide understandable explanations for why a review or user is flagged as spam.

Main Methods:

• A multi-head graph channel attention network was designed to capture high-order user interactions.
• Interpretability was achieved by combining the HSIC Lasso algorithm and a random walk with restart strategy.
• Key features influencing detection results were selected to provide final explanations.

Main Results:

• The proposed model demonstrated significant improvements in accuracy, precision, recall, and F1-measure across multiple benchmark datasets (Amazon, YelpChi, YelpNYC, YelpZip, Yelp_four).
• Specific performance gains over state-of-the-art methods ranged from 2.9% to 16.13% in various metrics.
• The interpretation method achieved superior performance in terms of frequency distributions without noise and fidelity under different sparsity levels.

Conclusions:

• The developed model effectively detects spammers while providing interpretable insights into the detection process.
• This interpretable approach enhances the reliability and trustworthiness of spam detection systems in e-commerce.
• The findings pave the way for more transparent and accountable AI applications in online review analysis.