由于移动网络应用与电子支付业务量不断增长, 信用卡欺诈的情况也呈现 快速增长的趋势, 由此给金融机构和运营商带来了巨大的挑战. 欺诈检测问题本质上是不平衡的序列二分类问题, 这类问题数据样本规模大, 计算复杂度高, 数据分布极 不平衡, 数据与数据之间会存在序列关系. 文章使用长短期记忆网络(long short term memory networks, LSTM)结合历史交易序列, 并针对交易数据不平衡的特性, 整合了合成少数类过采样技术(synthetic minority oversampling technique, Smote)算法和k最近邻(k-nearest neighbor, kNN)分类算法设计并构建了一个基于kNN-Smote-LSTM的信用卡欺诈检测网络模型, 它 可以通过kNN判别分类器来不断筛选出安全生成样本来提升模型的性能, 克服了Smote算法在生成 新样本时的盲目性和局限性, 实证表明通过kNN-Smote-LSTM模型结构化的融合可以大大改善模型组合时对多数类的误分类问题, 展现了优越的风险检测性能.

As the number of mobile web applications and electronic payment services continues to grow, the situation of credit card fraud has also shown a rapid growth trend, which has brought enormous challenges to financial institutions and operators. The fraud detection problem is essentially an unbalanced sequence two-class problem. The data samples of this type are large in scale, high in computational complexity, extremely uneven in data distribution, and there is a sequence relationship between data and data. This paper uses long short term memory networks (LSTM) combined with historical transaction sequences, and integrates the synthetic minority oversampling technique (Smote) algorithm and k nearest neighbors (kNN) for the characteristics of transaction data imbalance. The classification algorithm designed and constructed a credit card fraud detection network model based on kNN-Smote-LSTM, which can continuously filter out safe generated samples to improve the performance of the model through kNN discriminant classifier. The blindness and limitations of the Smote algorithm in generating new samples demonstrate that the structured fusion of the kNN-Smote-LSTM model can greatly improve the misclassification of most classes when model combination, and demonstrate superior risk detection performance.