基于特征相似度降采样的交易欺诈预测模型

TRANSACTION FRAUD PREDICTION MODEL BASED ON FEATURE SIMILARITY DOWN-SAMPLING

在交易欺诈评估场景下,正负样本比例极其悬殊,需要对样本进行采样来解决样本不平衡。传统采样由于在采样过程中会丢失样本信息,导致模型预测的准确率不是很高。针对这类情况,提出一种基于特征相似度降采样方式的模型构建方法。该方法主要包括三个部分。(1)依据样本数据,构建有效的与欺诈相关的特征集。(2)通过引入样本差异度函数,在降采样时尽可能多地保留样本信息。(3)通过多个分类器进行融合输出欺诈概率。将该方法与其他常见采样方式进行比对,实验结果表明,该方法具有更好的评估结果。

In the scenario of transaction fraud evaluation,the proportion of positive and negative samples is extremely different,so it is necessary to sample the samples to solve the sample imbalance.Due to the loss of sample information in the traditional sampling process,the accuracy of model prediction is not very high.Aimed at this kind of situation,a model construction method based on feature similarity down-sampling is proposed.This method mainly included three parts.(1)According to the sample data,an effective feature set related to fraud was constructed.(2)By introducing the sample difference function,as much sample information as possible was retained when down-sampling.(3)Multiple classifiers were fused to output the fraud probability.This method was compared with other common sampling methods.Experimental results show that this method has better evaluation results.