Abstract:  [Purpose/Significance] To ensure the health of the e-commerce environment, business reputation and market fairness, this paper aims to propose a new spammer detection algorithm from the perspective of graph neural network. [Design/Methodology] Combined with the multiple relational neural network, we introduce a new sampling strategy, and design a new method of multiple relationship graph neural network based on precise sampling and adaptive neighborhood.Then this new approach is used on the real world Yelp and Amazon datasets to test the effect. [Findings/Conclusion] Compared with the existing fraud detection methods, this new algorithm, has a significant effect in mitigating the impact of category imbalance. [Originality/Value] This method introduces a new sampling strategy to solve the category imbalance problem caused by a small number of fraudulent users among a large number of benign users in fraud detection research.

Key words: Fraud detection, Category imbalance, Precise sampling, Adaptive neighborhood , Multiple relationship graph