Prediction of Customers’ Behaviors Based on XGBoost Model

Useing XGBoost model to predict customer behavior is the main contribution of this paper. The data set describing customer behavior includes 37 characteristic variables and one response variable. We divided the 37 characteristic variables into numerical variables and classification variables, conducted data preprocessing respectively. The maximum and minimum normalization was applied to standardize the numerical variables, and the digital coding processing technology was used to digitize the classification variables. In the feature selection part, numerical variables are selected by variance filtering, and the chi-square test was used to choose categorical variables. In addition, Fisher score was used to determine the merging characteristics, and 14 variables were retained in the scoring results. According to the ratio of 7:3, the data set corresponding to the features filtered by Fisher’s score is divided into training set and test set, and XGBoost model was established. Four significant variables that significantly affect the prediction results were found, and the accuracy of our model was verified to be 0.63, which indicated that the prediction has certain reliability and was very helpful for the company to predict customer behavior. This shows that the prediction has a certain reliability, which is very helpful for the company to predict customer behavior.