基于注意力机制的电信用户离网预测算法
首发时间:2019-08-19
摘要:随着电信市场用户增速的逐渐放缓,电信市场竞争趋于白热化,电信用户更换运营商的情况日益普遍。获取新用户的费用远远超过挽回潜在离网用户。因此利用大数据和深度学习方法,提前定位与挽留潜在离网用户,成为运营商维持用户数量,保证营业收入的一个重要途径。本文基于运营商呼叫详单记录(Call Detail Record,CDR)数据进行基于深度神经网络模型的用户离网预测。基于CDR数据,本文首先提取用户的基本通信行为特征、用户通信的社交对象相关特征,以及用户位置相关特征,并分别利用卷积神经网络(Convolution Neural Network,CNN)和循环神经网络(Recurrent Neural Network,RNN)对用户各类特征间的依赖关系及时序关系进行学习和表征。通过引入注意力机制,利用用户位置相关特征对各类其他特征影响的不同,本文学习各类特征的权重,提升最终的用户离网预测性能。最后,利用运营商CDR数据进行了实验评估,验证了基于注意力机制的这两种预测模型的性能均优于传统的机器学习算法。
For information in English, please click here
Churn Prediction Algorithms for Telecom Industry Based on Attention Mechanism
Abstract:With the slowdown of user growth in the telecommunication market, the competition in the telecommunication market tends to be white-hot, and it is increasingly common for telecom users to change operators. The cost of acquiring new users outweighs the cost of retrieving potential churn users. Therefore, with the help of large data and deep learning methods to locate and retain potential churn users in advance has become an important way for operators to maintain the number of users and ensure business income. In this paper, a churn prediction method based on deep neural network model with call detail record (CDR) data is presented. This paper first extracts basic communication behavior features, the social features of tele communication and the location-based features. Convolution Neural Network (CNN) and Recurrent Neural Network (RNN) are used to learn the dependencies and temporal relatioships among user features. By introducing attention mechanism and utilizing the different influence of user location-based features on other features, this paper improves the final churn prediction performance by learning the weights of various features. Finally experiments on operator CDR data show that the performance of two prediction models based on attention mechanism is better than traditional machine learning algorithms. .
Keywords: Artificial Intelligence Operator Data Churn Prediction Deep Learning Attention
基金:
引用
No.****
同行评议
勘误表
基于注意力机制的电信用户离网预测算法
评论
全部评论0/1000