Satisfaction prediction based on XGBoost classification and decision tree model

With the rapid development of mobile communication technology and the continuous improvement of network coverage, how to improve the online business has become an important demand for major mobile operators to operate in the market. In this paper, the influencing factors of customer satisfaction with online service are evaluated and analyzed, and a prediction model based on relevant influencing factors is established to predict customer scores. This paper mainly establishes the prediction model of customer scoring based on relevant influencing factors and predicts the customer scoring. Firstly, this paper extracts the main factors that affect the user rating based on Pearson correlation coefficient. The speed of mobile Internet, network coverage and signal strength, and the stability of mobile Internet are the three most important factors that affect the overall satisfaction of customers with mobile Internet. The Pearson correlation coefficient r also reaches the lowest value of 0.89, and the speed of mobile Internet is the most important influencing factor. The correlation coefficient r reaches 0.91, indicating a strong correlation. Then, through the analysis and observation of the data and according to the results obtained from multiple machine learning model tests, a fusion model based on three single models of XGBoost, LightGBM and decision tree is finally established as the final prediction model, where XGBoost is the main model, LightGBM and decision tree are the auxiliary models. At the same time, the hyperparameter tuning and classification prediction of the machine learning model are performed many times. Based on the analysis results of the main influencing factors and the results obtained from the training set, the corresponding data preprocessing is carried out, and the fusion algorithm model is used for classification training and prediction based on the selected influencing factors. The data was segmented, 70% of which was the training set, and the remaining 30% was used as the test set. The training process graphs such as the dendrogram of the decision tree were trained, and the accuracy of the training set was about 97%. The average number of features with great influence on the predicted attribute was 5-10, and the importance of the features was up to 41%. The data less than 0.001 in the prediction result is regarded as 0, and the probabilities of various prediction results are summed to judge the validity. After calculation, the predicted data are all valid.