Grad: Guided Relation Diffusion Generation for Graph Augmentation in Graph Fraud Detection

AI summary

Overview

Research area: Graph machine learning for financial fraud detection, specifically graph augmentation using denoising diffusion and supervised contrastive learning.

Technical level: Intermediate to Advanced. Readers should be comfortable with graph neural networks, denoising diffusion probabilistic models, and contrastive learning objectives. The paper is empirically oriented and assumes familiarity with standard fraud-detection benchmarks.

Scope: The paper introduces Grad, a graph augmentation framework that generates new homophilic (same-class) edges via guided relation diffusion to expose fraudsters who hide by mimicking benign users, validated on three public benchmark graphs and two large-scale datasets from WeChat Pay.

What This Paper Is About

Fraudsters on modern payment platforms have learned to disguise themselves by imitating the normal transaction behavior of legitimate users, so their features and neighbor relationships look almost identical to benign accounts. This "Adaptive Camouflage" erases the statistical gaps that existing fraud detectors rely on, causing those models to smooth away the few remaining faint fraud signals. The paper's goal is to artificially construct new, camouflage-free connections between similar nodes so that those weak signals become strong enough for a detector to notice.

Key Contributions

1. First work to target "Adaptive Camouflage" directly. The authors identify and quantify a high fraud-benign similarity regime (83.2% on WeChat Pay versus roughly 41.8% on the classic Amazon dataset) and frame it as a distinct, unsolved problem for graph fraud detection.

2. A relation-generation approach instead of node-generation. Rather than synthesizing fake fraudulent nodes, which camouflage can mislead, Grad generates auxiliary homophilic relations from scratch, producing graph edges that contain no inter-class (fraud-to-benign) links.

3. A guided diffusion generator with dual guidance. The denoising diffusion process is steered by two signals: a supervised graph contrastive learning gradient (pushing connected nodes toward same-label similarity) and a degree penalty (keeping the generated graph's degree distribution consistent with real financial networks).

4. Strong empirical results across five datasets, including industrial-scale data. Grad outperforms state-of-the-art baselines by up to 11.10% in AUC and 43.95% in AP, and remains memory-efficient on large graphs where several competing models run out of memory.

Main Findings

• Large gains under heavy camouflage: On the WeChat Pay-Large dataset (441,640 nodes, 0.4% training ratio), Grad improves AUC by 7.47% and AP by 5.21% over the best baseline. On YelpChi, AUC rises 8.86% and AP rises 32.66%.

• Strongest improvements on the hardest, highest-dimensional data: On BlogCatalog, where 8,189-dimensional features cause most models to overfit, Grad achieves +11.10% AUC and +43.95% AP. The authors attribute this to the supervised contrastive module reducing effective feature dimensionality while preserving fraud-relevant information.

• Modest but consistent gains on easy datasets: On Amazon (+0.60% AUC, +1.55% AP) and WeChat Pay-Small (+1.17% AUC, +3.41% AP), improvements are smaller because benign and fraudulent nodes are already fairly separable.

• Memory and scalability advantage: Graph Transformer and generative baselines (GraphGPS, Graphormer, TGEditor, GGAD, CARE-GNN) frequently produced out-of-memory failures on the large datasets, while Grad scaled.

• Ablation confirms each component matters: Removing relation generation (Grad w/o Gen) causes the largest drops on camouflage-heavy datasets; removing diffusion guidance (Grad w/o Gui) degrades performance notably on YelpChi; removing the wavelet filter fusion (Grad w/o WFu) collapses AP on BlogCatalog from 92.15 to 67.15.

• Tensor/feature analysis supports the premise: Embedding visualizations and neighbor-similarity distributions show fraudsters separating more clearly from benign users after Grad's augmentation.

Methodology in Plain English

Grad works in five stages:

1. Node Group Sampler. Instead of taking K-hop neighborhoods (which would preserve the fraudster's disguised connections), the method randomly shuffles all nodes and partitions them into equal-sized, non-overlapping groups. This physically separates fraudsters from the benign accounts they have carefully imitated, weakening the camouflage before any learning happens.

2. Supervised Graph Contrastive Learning. Using the small set of labeled nodes, the model is trained so that same-label nodes get pulled together and different-label nodes get pushed apart. The encoder uses a high-pass filter formulation — each node's representation is contrasted against the mean of its neighbors — which emphasizes local differences rather than smoothing them. The resulting loss gradient is reused later as a guidance signal.

3. Guided Relation Diffusion Generator. The adjacency matrices of the node groups are treated like images in a denoising diffusion model. Noise is added over many timesteps, and a network is trained to predict that noise. During sampling, two extra gradient terms steer generation: a similarity guidance term encouraging edges only between representationally similar nodes, and a degree penalty keeping node degrees reasonable. This yields new edges that contain essentially no fraud-benign connections.

4. Relation Augmentation. Personalized PageRank graph diffusion is applied to the full merged network to attach global structural context to the generated relations, so the new edges are not purely local artifacts.

5. Multi-Relation Detector. Each original and generated relation is processed by β wavelet filters (band-pass, spatially and spectrally localized), and outputs are combined with learnable per-relation weights. This lets the model select whichever relations matter most for a given deployment scenario and avoids the low-pass filtering failure of earlier spectral methods.

Training combines an MSE loss on the injected noise with a standard binary cross-entropy loss for the final fraud classification.

Why This Matters

Impact on research: The paper reframes graph fraud detection as a problem of relation quality rather than node synthesis, and it provides a concrete benchmark for the under-studied high-similarity camouflage regime. The combination of supervised contrastive learning as a diffusion guidance signal is a transferable technique beyond fraud detection.

Real-world applications:

• Online payment platforms (the direct deployment context) screening transactions where fraud accounts mimic legitimate spending patterns.
• Banking and credit-card fraud detection, where organized fraud rings deliberately imitate normal account behavior.
• E-commerce and marketplace abuse detection, including fake seller or review networks that blend into genuine activity.
• Money-laundering and illicit-fund-flow detection in transaction graphs, where actors deliberately maintain low-profile transaction histories.
• Social network spam and bot detection, where automated accounts are tuned to look like ordinary users.

Industry relevance: The WeChat Pay datasets come from a production payment platform with billions of users and roughly 110 million two-hop transactions across 441,640 users, with a fraud ratio near 0.33%. Grad was also employed in an online industrial environment, and the paper emphasizes scalability and memory efficiency — practical constraints that eliminated several academic baselines. Code is publicly released.

Future Directions

• Adapting to evolving camouflage: The paper assumes a fixed camouflage strategy observed at training time. Future work could study adversarial settings where fraudsters react to the new detection method and shift their imitation behavior over time.

• Reducing label dependence: The supervised contrastive module depends on a small pool of expert-labeled nodes (4,371 in the WeChat Pay setup). Extending to semi-supervised or self-supervised guidance could improve applicability where labels are scarcer.

• Temporal and dynamic graphs: Financial networks are inherently time-evolving; the current formulation appears static. Incorporating temporal relation diffusion is a natural extension.

• Beyond binary classification: Real fraud ecosystems involve multiple fraud types and collusion structures. Generalizing the relation generation to multi-class or group-level detection remains open.

• Interpretability and auditability: Since relation generation is guided by contrastive similarity and degree constraints, future work could expose why a particular new edge was created, which matters for regulatory review in finance.

Target Audience

Researchers and practitioners in graph representation learning, financial fraud detection, and applied generative modeling. It is particularly valuable for industrial ML engineers at payment platforms, banks, and large marketplaces who face imbalanced, heavily camouflaged fraud data at scale, and for academics studying graph augmentation, diffusion models on graphs, or contrastive learning under weak supervision. Readers without background in GNNs or diffusion models will need to consult prior work on DDPM and graph contrastive learning to follow the methodological details.

Authors’ abstract

Nowadays, Graph Fraud Detection (GFD) in financial scenarios has become an urgent research topic to protect online payment security. However, as organized crime groups are becoming more professional in real-world scenarios, fraudsters are employing more sophisticated camouflage strategies. Specifically, fraudsters disguise themselves by mimicking the behavioral data collected by platforms, ensuring that their key characteristics are consistent with those of benign users to a high degree, which we call Adaptive Camouflage. Consequently, this narrows the differences in behavioral traits between them and benign users within the platform's database, thereby making current GFD models lose efficiency. To address this problem, we propose a relation diffusion-based graph augmentation model Grad. In detail, Grad leverages a supervised graph contrastive learning module to enhance the fraud-benign difference and employs a guided relation diffusion generator to generate auxiliary homophilic relations from scratch. Based on these, weak fraudulent signals would be enhanced during the aggregation process, thus being obvious enough to be captured. Extensive experiments have been conducted on two real-world datasets provided by WeChat Pay, one of the largest online payment platforms with billions of users, and three public datasets. The results show that our proposed model Grad outperforms SOTA methods in both various scenarios, achieving at most 11.10% and 43.95% increases in AUC and AP, respectively. Our code is released at https://github.com/AI4Risk/antifraud and https://github.com/Muyiiiii/WWW25-Grad.