Research and Practice of Telecommunication User Rating Method Based on Machine Learning

Abstract

The machine learning model has advantages in multi-category credit rating classification. It can replace discriminant analysis based on statistical methods, greatly helping credit rating reduce human interference and improve rating efficiency. Therefore, we use a variety of machine learning algorithms to study the credit rating of telecom users. This paper conducts data understanding and preprocessing on Operator Telecom user data, and matches the user’s characteristics and tags based on the time sliding window method. In order to deal with the deviation caused by the imbalance of multi-category data, the SMOTE oversampling method is used to balance the data. Using the Removing features with low variance method and packaging method for feature selection, then the basic models are established. The empirical results of the model show that the Random Forest and XGBOOST ensemble models are better than the single models such as Bayes, SVM, KNN, and Decision Tree. The performance of Decision Tree in single models is better. Therefore, Random Forest, XGBOOST and Decision Tree models were selected to debug the hyper parameters to achieve model optimization. Based on the optimized model, the accuracy, recall, precision, confusion matrix and other indicators are evaluated, and it is concluded that low-level recognition is more accurate than high-level recognition and fewer misjudgments. Comparing the evaluation indicators of each level of different models, it is found that the integrated model performs better, indicating that Random Forest and XGBOOST are more suitable for solving the problem of telecommunications user rating. For this reason, this article proposes an implementation plan based on Random Forest and XGBOOST algorithm and model for the problem of telecommunications user rating.