How Graph Neural Networks Boost Anti‑Cheat in User Referral Activities

Operational activities such as referral invitations are essential for user growth, but they are frequently abused by fraudsters who create fake accounts to earn rewards, severely degrading campaign effectiveness.

Key Challenges

Lack of relational modeling: Existing tree, DNN, and traditional machine‑learning models focus on individual user features and ignore the strong connections among users in cheating groups.

Low sample purity: Randomly sampled "clean" users often contain undiscovered cheating accounts, limiting the performance of supervised models.

Graph‑Based Solution

Graph neural networks (GNNs) can jointly learn from graph topology and node attributes, and as semi‑supervised models they can leverage abundant unlabeled data, thereby enhancing recall of cheating accounts.

GCN Overview

GCN is a multi‑layer graph convolutional network that aggregates one‑hop neighbor information at each layer. The propagation rule can be expressed as:

H^{(l+1)} = \sigma\big(\hat{A}\,H^{(l)}\,W^{(l)}\big)

where \hat{A} is the normalized adjacency matrix (including self‑loops), H^{(l)} the node embeddings at layer l, and W^{(l)} the learnable weight matrix.

Referral (Master‑Apprentice) Graph Construction

In a referral scenario, each inviter ("master") and invitee ("apprentice") form a directed edge. Two edge types are built:

City+Device edges

IP+Device edges

Edges whose weight is below a threshold T are pruned to reduce noise.

Model Variants

edge_union: merges edges from both graphs into a single graph, treating all edge types uniformly.

scgcn‑split: uses the embedding learned from Graph A as input features for Graph B.

scgcn (serial fusion): concatenates the two graphs and trains them jointly, allowing parameters to be shared across graphs.

Experimental Results

GCN alone improves cheat‑sample recall by 42.97% . Multi‑graph approaches further increase recall, with scgcn achieving the highest absolute number of recovered cheating accounts, while edge_union performs worse than single‑graph GCN due to loss of edge‑type information.

Conclusion and Future Work

Graph models capture both node features and relational information, leading to a ~50% increase in cheat‑sample recall for referral activities. Future directions include learning edge weights, enriching node attributes, and exploring more advanced GNNs such as GAT and DeepGCN to handle increasingly sophisticated fraud patterns.

References

Kipf, T. N., & Welling, M. (2016). Semi‑supervised classification with graph convolutional networks. arXiv:1609.02907.

Veličković, P., et al. (2017). Graph attention networks. arXiv:1710.10903.

Li, G., et al. (2019). DeepGCNs: Can GCNs go as deep as CNNs? Proceedings of the IEEE/CVF International Conference on Computer Vision.