Federated Learning And SHAP And Its Application In Big Data Risk Control

At present,data has become an important factor that restricts the development of machine learning.How to securely and conveniently combine multiple corporate data to establish models to obtain a model with better performance becoming an important issue.Federal learning focuses on the encryption parameters when exchanging certain model training and uses multi-party data to establish a model to break the data of the data islands to improve the model performance.As the amount of data to be processed by machine learning models is increasing,it is difficult to use simple machine learning models to give full play to the great advantage of data.With the widespread application of complex machine learning models in the field of big data,people are becoming more and more difficult to understand the process of machine learning model prediction.The interpretation of machine learning models has become an important issue.The interpretation of machine learning models under federal learning can not only help people understand the prediction process of the model,but also to a certain extent reflect the data value of enterprises from all parties.This article proposes the model interpretation method of the XGBOOST algorithm under the federal learning environment,which can make a certain degree of explanation of the prediction of each sample,to facilitate understanding the role of each feature in the model prediction.In the empirical part,the pre-processing of the data using the credit loan data of Lending Club has been processed by the data.Ten models are obtained.Ten features are allocated to both parties to simulate the modeling of the two parties in the federal learning environment,and the federal learning to learn from the federal learning environment and learn the federal learning to learn from the federal learning environment.The models constructed and only using the data and label construction of only one of them and the model of the label and the model that directly combined the data of the two parties.Finally,it was found that the model built by the federal learning and the directly concentrated data of the two parties were constructed together.The model is almost no different and better than the model built by only one side of the data and labels and obtained the most important feature of the model prediction results in this empirical analysis through the interpretation of the Federal Learning model.