Telecom user churn prediction scheme based on large language model  

作　　者：Chen Hao Yang Liu Ma Chao Wei Yifei 

机构地区：[1]School of Electronic Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China [2]China Academy of Information and Communications Technology(CAICT),Beijing 100083,China

出　　处：《The Journal of China Universities of Posts and Telecommunications》2024年第6期57-65,94,共10页中国邮电高校学报(英文版)

摘　　要：With the fierce competition among major telecom operators for existing users, machine learning has made remarkable progress in customer churn prediction. However, most existing work and experiments rely on the data of local telecom operators, resulting in significant differences in algorithm selection and experimental results. Large language models(LLMs) have obtained surprising abilities in solving complex tasks through learning on large-scale unlabeled corpora and continuously scaling model size. By converting user data into a natural language format, LLM is employed to capture underlying patterns in the data and predict the probability of user churn in the following month. With extensive training on abundant text data from large-scale public corpora, the integration of LLMs in this task has the potential to serve as a general scheme and demonstrate unique advantages in few-shot scenarios and interpretability. A series of experiments are performed to test the capabilities and limitations of LLMs in customer churn prediction. The results indicate that with a certain number of prompts and instructions, LLM can perform well in this task. However, due to the limited rate of interaction with the model, it is necessary to initially employ some methods upstream of the model to screen potential churn users.

关 键 词：telecom user churn large language model(LLM) machine learning few-shot scenarios 

分 类 号：TN9[电子电信—信息与通信工程]