Towards Designing Data-driven Credit Risk Control And Operation System For On-line Education

Since the rapid development of English Online Education(EOE),more and more companies strengthen their competiveness using big data.This research aims at designing data-driven operation system and credit risk management to strengthen the core capability of the EOE company.As an important area of EOE,user operation attracts more and more academic and industrial research attention during the recent years.Currently,the principle problems,such as lack of transaction data,systematical feature exploration and trial of prediction model,are to be researched.This part of research aims at a)analyzing the user features of EOE,b)mining the association rules among different features,and c)building course preference prediction model based on the analysis of a)and b).In order to prove the usability of the conclusions,the practical data of Jian Xiao Chi is adopted in this research.Based on the insights of feature analysis and association mining on the user set of 79483 users,a course preference prediction model is built.The effectiveness and efficiency of this prediction model is experimentally verified via comparing with manual selection and other classical machine learning models.The experimental results of execution in practical environment reveal that this model is remarkably better in both improving conversion ratio and saving marketing cost.Furthermore,another core domain of credit risk management for high price course(HPC)in EOE is researched by designing probabilistic model enhanced by machine learning technique.A total of more than 30 thousands practical users containing both transaction and credit data are adopted.This part of research includes the following three parts: a)analyzing all of 112 risk features and completing the feature section.As a result,top 19 features are selected for model building;b)building the credit risk management model using Logistic regression(LR),extreme gradient boosting(XGB)and deep neural network-logistic regression(DNN-LR).XGB performs the best amongst these three methods,and is implemented in the on-line environment;c)The on-line experiment reveals that the XGB based credit risk management model is significantly effective and robust.Moreover,the interception rate is investigated to balance the loss of operation and the loss of fraud.