融合时间特征增强与定向GAN过采样的XGBoost信用卡欺诈检测研究

Study on XGBoost-Based Credit Card Fraud Detection Integrating Temporal Feature Enhancement and Targeted GAN Oversampling

摘要: 针对信用卡欺诈检测中样本极度不平衡导致模型偏向正常样本、传统特征缺乏时间关联性导致区分度不足、过采样生成样本无针对性导致误判率高等核心问题，本文提出一套融合时间特征增强与定向GAN过采样的XGBoost检测方案。首先，基于公开信用卡欺诈数据集构建“交易小时–临时用户标识–交易间隔–间隔统计”四层时间特征体系，补充欺诈行为的时序关联信息；其次，针对小额漏检与大额欺诈两类核心漏检模式，设计定向GAN过采样模型，通过特征约束生成目标样本，并以余弦相似度筛选高质量样本；最后，优化XGBoost关键参数构建高精度分类模型。实验结果表明，该方案在测试集上欺诈类精确率达98.81%、召回率达84.69%、F1值达91.21%；消融实验验证，时间特征增强可使F1值提升4.25个百分点，定向GAN过采样可提升0.99个百分点。

Abstract: Aiming at the core issues in credit card fraud detection, such as extreme class imbalance leading models to favor normal samples, insufficient discriminative power of traditional features due to the lack of temporal correlation, and high misjudgment rate caused by the lack of targeting in oversampled generated samples, this paper proposes a set of XGBoost detection schemes integrating temporal feature enhancement and targeted GAN oversampling. Firstly, based on the public credit card fraud dataset, a four-layer temporal feature system of “transaction hour—temporary user ID—transaction interval—interval statistics” is constructed to supplement the temporal correlation information of fraudulent behaviors. Secondly, targeting the two core missed detection patterns (small-amount missed detection and large-amount fraud), a targeted GAN oversampling model is designed. Target samples are generated through feature constraints, and high-quality samples are screened using cosine similarity. Finally, the key parameters of XGBoost are optimized to construct a high-precision classification model. Experimental results show that the scheme achieves a fraud class precision of 98.81%, recall of 84.69%, and F1-score of 91.21% on the test set. Ablation experiments verify that temporal feature enhancement can increase the F1-score by 4.25 percentage points, and targeted GAN oversampling can increase it by 0.99 percentage points.