| With the continuous development of my country’s credit economy,the proportion of credit loans in financial institutions has gradually increased.The consumption mode of credit loan has been greatly developed in people’s life.For financial institutions,how to accurately evaluate personal credit,reduce bad debt rate and reduce approval time has become an important content.In order to improve the accuracy and real-time performance of the personal credit evaluation model,this paper makes the following research:(1)In order to improve the accuracy of personal credit evaluation by financial institutions,this paper uses XGBoost,GBDT,random forest and logistic regression as the base learners from the perspective of model fusion,and uses the Stacking method for model fusion.A random downsampling method based on Stacking is proposed for the unbalanced characteristics of the personal credit evaluation data set,which effectively improves the processing ability of the Stacking integrated model for unbalanced data sets.Moreover,for the problem that a single poor base learner affects the overall effect in the Stacking method,a weighted Stacking integrated model is proposed,which reduces the impact of a poor base model on the integrated model and effectively improves the overall performance of the integrated model..(2)In order to solve the problem of long training time of the Stacking integration method,this paper proposes a Flink-based Stacking integration model parallel optimization method,which uses the resources of the Flink cluster for parallel computing,effectively reducing the training time of the integration model.(3)In order to solve the problem of poor real-time traditional personal credit evaluation,this paper proposes a real-time evaluation model based on Flink.Take advantage of Flink’s stream processing advantages to generate prediction results in real time.Compared with the traditional non-real-time model,this real-time evaluation model can make full use of the advantages of the cluster and meet the needs of real-time decision-making.(4)The experimental data set in this paper comes from the open source data set of the Give Me Some Credit competition on the Kaggle platform,and the experiment is carried out in the mode of Flink on Yarn.Through comparative experiments with XGboost,GBDT,random forest and logistic regression,the results show that the Flink-based Stacking integrated model has obvious advantages over other models in terms of AUC,recall rate and F1 value.And a distributed acceleration experiment was carried out,which verified that the Flink-based Stacking integrated model takes full advantage of the advantages of the cluster in a distributed computing environment and effectively reduces the training time of the integrated model.Finally,under different scales of data sets,the Flink real-time evaluation model is compared with the non-real-time model,which shows that the Flink-based realtime model has great advantages in real-time prediction. |
