Personalized Recommendation For Existing Telecom Customers Based On Improved XGBoost Algorith

With the continuous development of China’s telecommunications industry,especially in the further spread of technologies such as 5G and gigabit broadband,it is more difficult for operators’ services to meet the needs of a wide range of users at the same time.Coupled with increasingly complex business rules and frequent product updates in recent years,users are unable to choose a product that suits them perfectly from the vast number of packages available.The ability to accurately analyse user needs will therefore have a significant impact on the outcome of user behaviour predictions.For the operators themselves,personalised package recommendations are becoming an extremely important research direction in the telecoms sector in order to promote healthy corporate development and iterative product management,enhance the quality of subscriber communication services and consolidate stock retention.Integrated algorithms in machine learning are better for processing data classification,especially the XGBoost algorithm which is more popular in various competitions.If a high dimensionality of data features is encountered,using the model to train the data can take up a lot of computational resources and time.Direct recommendations to users often do not produce good results if they still do not take into account the differences in their products.In response to the above problems,a model based on dual XGBoost is developed in this paper.By analysing a large amount of user behaviour data,multi-classifying telecom user data and pre-processing it using data mining techniques to create more valuable features according to the attributes of the package.The first use of XGBoost was based on RFE for feature selection,which is a combination of the XGBoost algorithm and recursive feature elimination.Remove invalid features by iterative creation and multiple iterations,greatly simplifying the model feature complexity and improving the efficiency of the model when formally processing data.The second use of XGBoost aimed at classification modelling,where users with large differences between the two business attributes were classified for policy application and separate moderation to obtain results,the final model evaluation method used a macro-averaged F1-Score and obtained a prediction of 85.4%.