Fraud detection on multi-relation graphs aims to identify fraudsters in graphs. Graph Neural Network (GNN) models leverage graph structures to pass messages from neighbors to the target nodes, thereby enriching the representations of those target nodes. However, feature and structural inconsistency in the graph, owing to fraudsters' camouflage behaviors, diminish the suspiciousness of fraud nodes which hinders the effectiveness of GNN-based models. In this work, we propose DiG-In-GNN, Discriminative Feature Guided GNN against Inconsistency, to dig into graphs for fraudsters. Specifically, we use multi-scale contrastive learning from the perspective of the neighborhood subgraph where the target node is located to generate guidance nodes to cope with the feature inconsistency. Then, guided by the guidance nodes, we conduct fine-grained neighbor selection through reinforcement learning for each neighbor node to precisely filter nodes that can enhance the message passing and therefore alleviate structural inconsistency. Finally, the two modules are integrated together to obtain discriminable representations of the nodes. Experiments on three fraud detection datasets demonstrate the superiority of the proposed method DiG-In-GNN, which obtains up to 20.73% improvement over previous state-of-the-art methods. Our code can be found at https://github.com/GraphBerry/DiG-In-GNN.

As a new paradigm in training intelligent models, federated learning is widely used to train a global model without requiring local data to be uploaded from end devices. However, there are often mislabeled samples (i.e., noisy samples) in the dataset, which will cause the model update to deviate from the correct direction during the training process, thus reducing the convergence accuracy of the global model. Existing works employ noisy label correction techniques to reduce the impact of noisy samples on model updates by correcting labels; however, such methods necessitate the use of prior knowledge and additional communication costs, which cannot be directly applied to federated learning due to data privacy concerns and limited communication resources. Therefore, this paper proposes a noise-aware local model training method that corrects the noisy labels directly at the end device under the constraints of federated learning. By constructing a label correction model, a joint optimization problem is formally defined for optimizing both the label correction model and the client-side local training model (e.g., classification model). As a solution to this optimization problem, we propose a robustness training algorithm using label correction, along with a cross-validation data sampling algorithm that updates both models simultaneously. It is verified through experiments that the mechanism can effectively improve the model convergence accuracy on noisy datasets in federated learning scenarios.