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

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.