TL;DR
This paper introduces DCL-GFD, a novel graph neural network (GNN) approach for fraud detection that tackles the challenges of class imbalance and homophily-heterophily mixtures in fraud graphs. By dynamically modeling neighborhoods based on estimated node abnormality and employing node-subgraph contrastive learning, DCL-GFD significantly improves fraud detection performance.
What Problem Does This Paper Address?
Fraud detection is a critical area, especially with the rise of online activities. Graph Neural Networks (GNNs) have shown promise by representing entities and their relationships as graphs. However, existing GNNs struggle with two major issues in fraud detection:
- Class Imbalance: Fraudulent activities are rare compared to legitimate ones, making it hard for models to learn the patterns of the minority class (fraud).
- Homophily-Heterophily Mixture: While GNNs often assume homophily (nodes with similar features connect to each other), fraud graphs exhibit a mix. Fraudulent nodes might disguise themselves within benign communities (heterophily) to avoid detection, while benign nodes might still exhibit homophily. This mixture can lead to the loss of crucial abnormal signals during information propagation.
These challenges compromise the ability of GNNs to accurately identify fraudulent activities.
Methods
The proposed method, DCL-GFD, introduces two core components:
- Node Abnormality Estimation: This module estimates the likelihood of a node being fraudulent or benign based on its feature similarity with its corresponding subgraph. The intuition is that benign nodes will have higher feature similarity with their neighbors (subgraph) than fraudulent nodes, which might be camouflaged. This is achieved using a node-subgraph contrastive learning approach. Partially observed node labels are used to train a contrastive loss function that encourages high similarity between benign nodes and their subgraphs, and low similarity between fraudulent nodes and their subgraphs.
- Dynamic Neighborhood Modeling: Guided by the estimated abnormal probabilities from the first module, this mechanism adaptively aggregates neighbor information. It models "benign-like" and "fraud-like" neighbor embeddings separately. By combining these, the model can dynamically adjust the influence of neighbors, effectively capturing the uncertain and mixed nature of the neighborhood, and thus mitigating the impact of homophily-heterophily and class imbalance.
The final node embedding is a concatenation of the ego (node's own) and the dynamically aggregated neighbor embeddings, which is then fed into an MLP for fraud probability prediction. The overall model is trained end-to-end by jointly optimizing a fraud detection loss (weighted cross-entropy) and the contrastive loss.
Datasets
The paper evaluates DCL-GFD on four real-world datasets:
- Weibo: Identifies anomalous accounts on social media.
- Amazon: Detects compensated users for deceptive product reviews.
- YelpChi: Focuses on identifying unfair reviews for products or businesses.
- T-Finance: Analyzes transaction graphs for anomalous accounts.
These datasets represent various domains where graph-based fraud detection is relevant.
Key Insights and Ideas
- Node-Subgraph Contrast is Key: The paper highlights that analyzing the feature similarity between a node and its local subgraph is a powerful way to estimate its abnormality, especially in the context of fraud.
- Dynamic Neighborhoods for Heterophily: Instead of a static aggregation, dynamically adjusting how neighbor information is incorporated based on estimated node properties is crucial for handling complex graph structures like those found in fraud detection.
- Preserving Minority Class Signals: The proposed method explicitly aims to preserve the characteristics of rare fraud nodes, which is often a bottleneck for traditional GNNs.
- Leveraging Partial Labels: Using contrastive learning with partially observed labels is an effective way to guide the learning process without requiring complete supervision.
Key Results
- Superior Performance: DCL-GFD outperforms state-of-the-art baselines on three out of four datasets in terms of AUC and Macro-F1 scores. For example, it achieved up to 2.03% better AUC on Weibo and 2.62% better Macro-F1 on Weibo compared to the best baseline.
- Robustness to Low Supervision: The model demonstrates strong performance even with limited training data (e.g., 10% training ratio), outperforming other methods like PMP and CONSISGAD.
- Effective Abnormality Estimation: Qualitative analysis shows a clear distinction in feature similarity between benign nodes and their subgraphs versus fraudulent nodes and their subgraphs, validating the node abnormality estimation module.
- Parameter Sensitivity: The study analyzes the sensitivity to subgraph sampling parameters (step size and restart probability), finding optimal ranges that balance local and global information capture.
Limitations and Future Work
- Dataset Specificity: While performing well on the tested datasets, the performance might vary on graphs with different characteristics.
- Computational Cost: Subgraph sampling and contrastive learning can add computational overhead, especially for very large graphs.
- Future Work: The authors suggest further exploration of dynamic neighborhood modeling and contrastive learning for graph-based fraud detection. They also implicitly point towards investigating more sophisticated subgraph sampling strategies or alternative contrastive objectives.
Summary updated