Detecting NFT Wash Trading with Temporal Graph Networks: A Comparison of Class Imbalance Handling Strategies

Abstract

The risk of market manipulation, especially wash trading—where colluding parties repeatedly trade the same asset to create false trading volume and misleading market signals—has increased due to the Non-Fungible Token (NFT) marketplaces’ rapid growth. The temporal nature of blockchain transactions and the stark class disparity between fraudulent and authorized activity make it difficult to detect wash trading. Previous research has mostly used graph-based models, rule-based heuristics, and conventional machine learning techniques; however, these methods frequently fail to capture changing transaction patterns over time. In order to close this gap, this paper models NFT transactions as a continuous-time dynamic graph and suggests a Temporal Graph Network (TGN) architecture for NFT wash trading detection. Wash trading labels are generated using heuristic rules adapted from prior literature, and five class imbalance handling strategies are evaluated: Binary Cross-Entropy, Weighted Binary Cross-Entropy, Class-Balanced Loss, Focal Loss, and Asymmetric Loss. Experiments are conducted on an Ethereum NFT transaction dataset using a chronological train-validation-test split. The results show that the baseline TGN model achieves the highest ROC-AUC of 0.930, PR-AUC of 0.706, and tuned F1 of 0.736, while class-balanced loss yields the best default-threshold F1 (0.738). The five strategies perform within a narrow band overall, indicating that once the temporal model is adequately trained, the choice of imbalance-handling strategy has only a secondary effect on detection performance. These findings demonstrate the effectiveness of temporal graph learning for NFT wash trading detection and provide insight into the impact of class imbalance handling strategies.