Applying Graph Machine Learning for Intelligent Anti‑Fraud: Models, Algorithms, and Real‑World Applications

01 Business Background

What is fraud? According to the Civil Code of the People’s Republic of China, fraud is a civil wrongdoing where a party intentionally provides false information or hides the truth to induce the other party to make an erroneous statement, emphasizing intentional deception.

Common fraud risks in finance and e‑commerce include account‑registration abuse, bulk login, marketing abuse (e.g., “wool‑pullers” and scalpers), malicious loan applications, cash‑out transactions, fake transactions, and account theft during payment.

On the B‑side, there are fake transactions, cash‑out, gambling, money‑laundering, etc., often supported by large black‑gray industry chains that provide resources (e.g., black‑card, proxy IP), develop automation scripts, and monetize the results.

The platform typically combats fraud via three methods: black/white lists (simple but slow), rule‑based strategies (extracting risk features), and machine‑learning models (both traditional and deep learning, though supervised learning needs many labels and deep models lack interpretability).

02 Common Models and Graph Algorithm Principles

1. Anti‑Fraud Model System

The system consists of three layers:

Risk‑early‑warning models based on anomaly detection and clustering to spot new risk patterns.

Risk‑identification models centered on graph and sequence algorithms for high‑precision detection of known risks.

Rule‑learning and frequent‑item mining models to translate model outputs into interpretable strategies.

2. Pain Points of Traditional Modeling and Advantages of Graph Machine Learning

Labels are hard to obtain; many risks remain undiscovered without expert input.

Traditional models treat samples in isolation, ignoring relational links crucial for group fraud.

Severe class imbalance (fraud samples often <0.1% of data).

High demand for model interpretability.

Graph machine learning addresses these issues by supporting unsupervised or semi‑supervised learning, leveraging node features and relational structure to mitigate imbalance, and offering natural visual interpretability.

3. Common Graph Algorithms in Anti‑Fraud

Community detection (e.g., Label Propagation, InfoMap) to discover tightly‑connected fraud groups.

Graph representation learning (DeepWalk, Metapath2Vec, DGI) to embed nodes into vectors.

Graph Neural Networks (GCN, RGCN) to combine node features with topology.

Probabilistic graphs (e.g., Hidden Markov Models) to model sequential behavior and compute generation probabilities.

03 Practical Business Applications

1. Community Detection in Marketing Scenarios

Fraudsters often operate in batches to harvest small profits; community detection on a user‑centric graph (users as nodes, any shared medium as edges) helps uncover hidden groups that bypass simple aggregation rules.

Challenges include converting heterogeneous graphs to homogeneous ones, assigning appropriate edge weights via supervised learning, and translating discovered groups into actionable rules using frequent‑item mining.

2. Graph Representation Learning for Cash‑Out Detection in Offline Transactions

Offline fraud users are rarely active online, resulting in sparse features. By constructing a bipartite user‑merchant graph and performing meta‑path random walks, node embeddings capture similar transaction behaviors, enabling clustering of cash‑out gangs.

3. Probabilistic Graphs in Marketing

Behavioral clustering combined with HMMs evaluates the generation probability of click‑stream sequences; low probability clusters are flagged as high‑risk.

04 Summary and Outlook

Graph machine learning, supporting both unsupervised and semi‑supervised paradigms, is indispensable for intelligent anti‑fraud systems. Future work includes mining implicit relational signals under stricter privacy, incorporating spatio‑temporal dynamics for real‑time confrontation, and enhancing interpretability via causal graphs.

05 Q&A

Q: Can probabilistic graphs use data other than sequential data? A: Currently they mainly rely on sequence data.

Q: Is converting graph models to strategies done manually? A: No, rule‑learning (e.g., decision trees) automatically translates model outputs into strategy rules.

Q: What time granularity is used in graphs? A: Day‑level granularity.

Q: How are edge weights determined? A: Initially via expert‑assigned weights, later refined by training a supervised logistic‑regression model on edge‑type features.

Q: Are heterogeneous‑to‑homogeneous conversions performed online? A: They are performed offline.

For the full slide deck and additional resources, please scan the QR code in the original article.