Application Of Improved Lightgbm Algorithm In Auto Sales Customer Churn Alert

With the prosperous development of China’s auto industry,the growth of car ownership and the rising age of cars,the profit chain of China’s auto industry tends to be oriented to the auto aftermarket.However,the development of auto sales and service stores in meeting the various needs of customers in the after-sales,as well as service marketing measures,is not perfect,leading to the problem of churned customers.Therefore,auto sales and service stores can use the information of customers to predict whether customers are lost or not,and can adopt precise marketing methods for customers identified as lost,so as to reduce the loss of customers and reduce loss.The distribution of customer data in auto sales service stores is unbalanced.The number of churned customers should be significantly lower than the number of unchurned customers.In this case,traditional algorithms will focus too much on the unchurned customers,leading to misclassification of the churned customers,and this misclassification would be extremely costly to the auto sales and service stores.Therefore,it is significant to study the problem of customer churn classification in auto sales and service stores.To address this phenomenon,this research takes over 60,000 customer data of a car sales and service store as the experimental object.A total of 18 features including customer base information,vehicle base information,vehicle loan information and vehicle insurance information are included.This research performs exploratory analysis,numerical data normalization,non-numerical data feature coding,feature dimensionality reduction and other pre-processing operations on the customer data.After literature review,most of the previous algorithms such as decision trees and logistic regression commonly used in customer churn problems have problems such as slow running rate and poor prediction of churned customers.Therefore,this research selects Light GBM algorithm,which is more suitable for handling large volume and unbalanced data sets,to model the problem.The results are compared with other classical machine learning algorithms(GBDT、RF、DT、SVM).The recall rate and the AUC value are used as evaluation metrics.Algorithms other than Light GBM are found to have lower AUC values,and the recall of all algorithms is extremely low.So,the IHT undersampling operation is used and the values of recall and AUC of the algorithm after undersampling reaches over 0.88 and 0.73 respectively.Then,in this research,the Focal Loss loss function in deep learning is introduced into Light GBM to obtain the FL-LGBM model.Modeling with the FL-LGBM model is found to have better prediction results.The recall rate reaches 0.9345,and the AUC value also reaches0.7608,and the time consumption is not significantly higher than other algorithms.Finally,the feature importance ranking is output by FL-LGBM model.Based on the results of the study,this research draws the following conclusions:(1)Light GBM model has better application effect in customer churn prediction,and the recall rate and AUC are significantly higher than other algorithms;(2)Applying the improved FL-LGBM model to the customer churn problem.The results are better than the Light GBM model after the resampling method and the original Light GBM model,and its interpretability is also stronger;(3)The age of the car,vehicle insurance information,vehicle sales price,whether the loan to buy a car and other factors have a greater impact on whether the customer churn.Auto sales and service stores can take precise marketing based on this information to reduce customer churn rate.