查询结果:   周捷,严建峰,杨璐,夏鹏,王猛.LSTM模型集成方法在客户流失预测中的应用[J].计算机应用与软件,2019,36(11):39 - 46.
中文标题
LSTM模型集成方法在客户流失预测中的应用
发表栏目
应用技术与研究
摘要点击数
316
英文标题
APPLICATION OF LSTM ENSEMBLE METHOD IN CUSTOMER CHURN PREDICTION
作 者
周捷 严建峰 杨璐 夏鹏 王猛 Zhou Jie Yan Jianfeng Yang Lu Xia Peng Wang Meng
作者单位
苏州大学计算机科学与技术学院 江苏 苏州 215006     
英文单位
School of Computer Science and Technology, Soochow University, Suzhou 215006, Jiangsu, China     
关键词
流失预测 长短期记忆网络 深度学习 集成学习 时序数据
Keywords
Churn prediction Long short-term memory Deep learning Ensemble learning Time series data
基金项目
国家自然科学基金项目(61572339)
作者资料
周捷,硕士生,主研领域:机器学习,数据挖掘。严建峰,副教授。杨璐,副教授。夏鹏,硕士生。王猛,硕士生。 。
文章摘要
目前客户流失预测任务中常用的模型集成方法采用传统机器学习模型作为基学习器。而传统机器学习模型相比于深度学习模型,存在无法对时序数据进行有效建模、特征工程对模型效果影响较大等缺点。针对这些问题,提出基于LSTM的模型集成方法。采用LSTM作为基学习器进行时序数据建模;改进snapshot模型集成方法,增加样本权重调整方法,在训练单个LSTM模型的过程中得到多个具有不同权值的模型;利用得到的多个模型构造新数据集,在新数据集上训练逻辑回归模型。实验结果表明,该方法相比于单模型LSTM,可以在仅花费其1.8倍训练时间的前提下,将查准率和PR-AUC分别提升4.67%和3.74%,显著提高了客户流失预测效果。
Abstract
At present, most of the ensemble learning methods used in customer churn prediction tasks use conventional machine learning models as a base learner. Compared with deep learning, the conventional machine learning models cannot efficiently model time series data and feature engineering has a great effect on the performance of models. To solve these problems, this paper proposed an ensemble learning method which was based on LSTM. We took LSTM as a base learner to model time series data, and used snapshot ensemble method and adjusted sample weights during the training process. Multi-models with different weights were saved in the process of training a single LSTM model. Finally, a new dataset was generated by using the multi-models, and a logistic regression model was trained on this dataset. Experimental results show that compared with LSTM, the proposed method improves the precision and PR-AUC by 4.67% and 3.74% respectively, and the training time is only 1.8 times as long as LSTM. This shows that the proposed method significantly improves the customer churn prediction effect.
下载PDF全文