PaperlayerView exact revision on arXiv

Source-linked AI summary

Semi-supervised Credit Card Fraud Detection via Attribute-Driven Graph Representation

Sheng Xiang, Mingzhi Zhu, Dawei Cheng, Enxia Li, Ruihui Zhao, Yi Ouyang, Ling Chen, Yefeng Zheng

arXiv:2412.18287v1cs.LGcs.AIcs.SI

TL;DR

Fraud-detection methods may underuse unlabeled transactions, which are abundant when only a tiny proportion of real-life transactions are labeled. The paper uses an attribute-driven temporal graph network with risk propagation, and reports that GTAN outperforms baselines across three datasets, including with little labeled data.

  • Problem

    Existing fraud-detection techniques may ignore unlabeled data and categorical attributes or require extensive categorical feature engineering, while only a tiny proportion of real-life transactions are labeled.

    Source: Introduction · p. 2

  • Method

    GTAN learns transaction representations on a temporal transaction graph using gated temporal attention, attribute learning, and risk embedding and propagation.

    Source: Introduction · p. 2, Gated Temporal Attention Networks · p. 3, Risk Embedding and Propagation · p. 4

  • Results

    GTAN outperforms baselines on three fraud-detection datasets, and semi-supervised experiments report more detected fraud transactions using unlabeled data with a bit of labeled data.

    Source: Introduction · p. 2, Conclusion · p. 7

  • Takeaways & Limitations

    The authors report that GTAN performs well with a small portion of labeled data, including training ratios as low as 10% on YelpChi and Amazon.

    Source: Semi-supervised Experiment · p. 6

Abstract

from arXiv · show

Credit card fraud incurs a considerable cost for both cardholders and issuing banks. Contemporary methods apply machine learning-based classifiers to detect fraudulent behavior from labeled transaction records. But labeled data are usually a small proportion of billions of real transactions due to expensive labeling costs, which implies that they do not well exploit many natural features from unlabeled data. Therefore, we propose a semi-supervised graph neural network for fraud detection. Specifically, we leverage transaction records to construct a temporal transaction graph, which is composed of temporal transactions (nodes) and interactions (edges) among them. Then we pass messages among the nodes through a Gated Temporal Attention Network (GTAN) to learn the transaction representation. We further model the fraud patterns through risk propagation among transactions. The extensive experiments are conducted on a real-world transaction dataset and two publicly available fraud detection datasets. The result shows that our proposed method, namely GTAN, outperforms other state-of-the-art baselines on three fraud detection datasets. Semi-supervised experiments demonstrate the excellent fraud detection performance of our model with only a tiny proportion of labeled data.

Introduction

Existing fraud-detection methods can leave unlabeled transactions and categorical attributes underused. The paper proposes a semi-supervised temporal-graph model that learns from attributes and risk information across transactions.

Introduction

  • Prior methods can ignore unlabeled data and categorical attributes, and may require extensive feature engineering for categorical features.The paper identifies these as limitations of recent temporal or graph-based fraud-detection techniques.Source: Introduction · p. 2
  • The proposed approach combines a Gated Temporal Attention Network with attribute learning and risk embedding and propagation to model fraud patterns.GTAN extracts temporal and attribute information from the transaction graph.Source: Introduction · p. 2
  • GTAN models transactions as a temporal graph and formulates fraud detection as semi-supervised node classification using both labeled and unlabeled data.The model passes attributes and risk information across the graph.Source: Introduction · p. 2
  • Across three datasets, the authors report that GTAN outperforms baselines and detects more fraud transactions using unlabeled data and a bit of labeled data.The paper describes the semi-supervised setting as using rich unlabeled information alongside limited labels.Source: Introduction · p. 2

Related Works

Related work covers credit-card fraud classifiers and graph-based semi-supervised learning. The paper positions its approach against limitations in exploiting unlabeled data, categorical attributes, and large graphs.

Credit Card Fraud Detection

  • Prior credit-card fraud work spans classical classifiers and graph methods; Cheng et al. (2020a,b) model one transaction/cardholder and do not exploit unlabeled real-life transactions.The cited literature includes Bayesian networks, neural networks, decision trees, support vector machines, convolutional methods, and graph-based methods.Source: Credit Card Fraud Detection · p. 2

Graph-Based Semi-supervised Learning

  • Prior graph approaches are described as facing scalability or categorical-embedding limits; this paper instead message-passes categorical attributes, including risk information from neighboring nodes.The passage notes that vanilla graph attention networks have O(N^2) space complexity, which is unaffordable for graphs with millions of nodes.Source: Graph-Based Semi-supervised Learning · p. 2

Proposed Method

GTAN represents transactions on a temporal graph, combines transaction attributes with propagated risk information, and predicts fraud risk from the resulting representations.

Model Architecture

Attribute Embedding and Feature Learning

  • The feature-learning layer embeds card, transaction, and merchant categories and add-pools their embeddings into each transaction’s categorical representation.Preprocessing retains full transaction records rather than filtering cards or merchants with few authorized records.Source: Attribute Embedding and Feature Learning · p. 3

Gated Temporal Attention Networks

  • Directed temporal edges connect past transactions to current transactions, and attention aggregates messages from sampled past transactions of the same cardholder.Neighbor sampling and truncation control the number of temporal edges per node.Source: Gated Temporal Attention Networks · pp. 3–4

Risk Embedding and Propagation

  • Manually annotated labels are embedded as transaction risk attributes, while unlabeled transactions receive zero risk embeddings.Adding risk embeddings to node features lets the graph network model temporal fraud patterns and risk propagation together.Source: Risk Embedding and Propagation · p. 4

Fraud Risk Prediction

  • A two-layer MLP produces risk predictions for all transactions, and binary cross-entropy supplies the training objective.The model is optimized using standard SGD-based algorithms.Source: Fraud Risk Prediction · p. 4

Masking to Avoid Label Leakage

  • Training masks the sampled center nodes’ risk embeddings to prevent self-loop label leakage, while inference uses all observed labels as input attributes.The training objective predicts labels from node attributes, masked risk information, and the temporal graph.Source: Masking to Avoid Label Leakage · pp. 4–5

Experiments

The experiments compare GTAN with fraud-detection baselines on three datasets, then assess performance under varying label ratios, component ablations, and parameter settings. Across these evaluations, GTAN performs strongly, including with limited labeled data.

Experiment Settings

  • Experiments use FFSD, a partially labeled partner dataset, plus the public YelpChi and Amazon review-graph datasets.For YelpChi and Amazon, the training-to-testing ratio is 2:3; FFSD trains on the first seven months and tests on August–October 2021.Source: Datasets · p. 5, Fraud Detection Performance · p. 5

Fraud Detection Performance

  • GTAN significantly outperforms all baselines, with at least 10%, 0.5%, and 6% AUC improvements and at least 27%, 2.9%, and 12.8% AP improvements across the three datasets.Table 2 reports AUC, F1-macro, and AP; the passage gives the improvements in dataset order without naming the dataset-to-value pairings.Source: Fraud Detection Performance · p. 6, Table 2 · p. 6

Semi-supervised Experiment

  • GTAN is best across 10%–80% training ratios on YelpChi and Amazon, and remains effective with only 10% labeled training data.On Amazon, AUC varies by no more than 2% as the training ratio changes.Source: Semi-supervised Experiment · p. 6, Figure 3 · p. 6

Ablation Study

  • Removing either temporal graph attention or risk embedding lowers GTAN’s performance; removing attention has the greatest impact.The authors attribute the attention component’s effect to reweighting temporal transaction neighbors and the risk component’s effect to modeling fraud patterns through risk propagation.Source: Ablation Study · pp. 6–7, Figure 4 · p. 7

Parameter Sensitivity

  • Performance peaks with two GNN layers and six temporal edges per node; at least two neighbors are needed, and deeper layers can degrade performance.Performance remains stable up to eight layers; beyond that, the authors suggest oversmoothing may explain degradation.Source: Parameter Sensitivity · p. 7, Figure 5 · p. 7
  • Hidden dimensions and batch sizes are generally insensitive across values from 16 to 512, with less than 3% AUC variation.Performance peaks relatively at hidden dimension 256 and is best with batch sizes 64 or 256; the authors choose 256 for training efficiency.Source: Parameter Sensitivity · p. 7, Figure 5 · p. 7

Conclusion

The paper presents a semi-supervised fraud detection method built on temporal transaction graphs and attribute-driven gated temporal attention. Experiments report superiority over other baselines on three datasets, including with only a tiny proportion of manually annotated data.

Conclusion

  • The proposed methods outperform other baselines on three fraud detection datasets.Semi-supervised experiments also show strong performance with only a tiny proportion of manually annotated data.Source: Conclusion · p. 7
  • Temporal transaction graphs and attribute-driven gated temporal attention form the proposed semi-supervised fraud detection method.Attribute representation and risk propagation model fraud patterns using categorical attributes and human-annotated labels.Source: Conclusion · p. 7
  • The approach has been deployed in a transaction fraud analysis system.Source: Conclusion · p. 7
  • Future work will study temporal and risk-propagation fraud patterns more effectively and efficiently.Source: Conclusion · p. 7

Related papers

Loading 2412.18287v1…