基于注意力机制和图卷积的电信客户流失预测  

Prediction of telecom customer churn based on attention mechanism and graph convolution

作　　者：黄子璇 夏壬焕 张雄涛 HUANG Zi-xuan;XIA Ren-huan;ZHANG Xiong-tao(Computer and Software School,Hangzhou Dianzi University,Hangzhou 310000,China;School of Information Engineering,Huzhou University,Huzhou 313000,China)

机构地区：[1]杭州电子科技大学计算机学院,浙江杭州310000 [2]湖州师范学院信息工程学院,浙江湖州313000

出　　处：《计算机工程与设计》2023年第6期1685-1691,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61771193)。

摘　　要：针对客户流失预测任务中,离散特征进行one-hot编码后特征空间过大,特征表示向量过于稀疏,引发维度灾难的问题,提出一种流失预测模型。基于特征嵌入和Transformer将高维离散数据转换为基于上下文的嵌入,降低编码后的数据维度,增强特征值间的联系。结合基于自适应邻接矩阵定义的图卷积操作,自动学习特征间潜在的关联关系,提高对正类样本的识别精度。使用交叉验证的方式在两份公开数据集上进行对比实验和消融实验,实验结果表明,改进模型能够有效增强多层感知机对于样本的拟合能力,提高分类预测准确率。To address the problems that the feature space is too large after one-hot encoding of discrete features and the encoded features are too sparse,which lead to dimensional disasters in the customer churn prediction task,a churn prediction model was proposed.Based on feature embedding and Transformer,this model was used to convert high-dimensional discrete data into context-based embedding,reduce the dimensionality of encoded data,and enhance the association between discrete feature eigenva-lues.The adaptive adjacency matrix-based graph convolution was combined to automatically learn the potential association relationships of features to improve the recognition accuracy of positive class samples.Comparison experiments and ablation experiments were conducted on two publicly available datasets using cross-validation.Results of the experiments reveal that the modified model can successfully increase the multi-layer perceptron’s fitting ability for the samples and improve classification prediction precision.

关 键 词：客户流失 多层感知机 离散数据 分布不均衡 注意力机制 自适应邻接矩阵 表示学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]