Internet Financial Default Prediction System Based On Feature Engineering And Multiple Classifiers Fusion

In recent years,Internet finance has become the trend of social finance development.With the development of technologies,such as artificial intelligence and big data technology,relying on fintech to collect,analyze,and organize financial data to provide accurate risk control services for segmented populations,which has become an effective way to solve Internet financial risk control problems.Applying artificial intelligence and big data technology to predict default users of Internet finance,it provides financial businesses with a more comprehensive understanding and evaluation of users.Machine learning algorithms are the key to the success of default user prediction,there are many deficiencies in the current research of many risk control models.The main contents of this thesis are as follows:1.In order to improve the prediction effect and enrich the information of data,data preprocessing and feature construction are carried out.In this thesis,data partition is constructed after missing value,time and information redundancy processing.A large number of features are carried out,and they are grouped according to feature types.Through experimental analysis,the effectiveness of each feature group is verified and evaluated,a wide range of ideas for feature engineering are provided.A variety of feature selection schemes are compared and analyzed to find a feature selection scheme suitable for user default prediction in financial scenarios.2.In order to improve the prediction effect of Internet financial,parameter tuning and multi-model building and fusion are carried out.The parameter has a large disturbance to the model and the single model is not stable.The parameter tuning is used to reduce the parameter disturbance.Building multi-model of XGBoost,Cat Boost,GBDT,Light GBM,Random Forest.Multi-model fusion takes advantage of each single model to improve the generalization performance and prediction effect of user default prediction algorithms.3.Taking the above feature engineering and parameter tuning,multi-model fusion as the core to design and implement the Internet financial default prediction system to assist managers in scientific prediction.The dataset of this thesis is derived from the users' real loan consumption behavior data from April 2016 to April 2017 provided by 2018 AI Global Challenger Contest,and the evaluation standard is the AUC of the prediction of the user default(The AUC value is the area under the ROC curve.The more accurate the model,the higher the AUC value).In this thesis,AUC is improved to 0.8267 through data preprocessing,feature construction and feature selection and parameter tuning.Through multi-model fusion,the prediction effect is significantly improved.The AUC is improved to 0.8293.It ranks first in the 1,222 teams in 2018 AI Global Challenger Contest.Based on the above mentioned core methods,this thesis designs and implements Internet financial user default prediction system to provide more accurate risk control services for the financial field.