Credit Evaluation Ensemble Model with Self-Contained Shunt

In the actual loan scenario, the proportion of “good” and “bad” samples is usually extremely imbalanced. Therefore, this paper constructs an ensemble model with a pre-judging mechanism for the imbalanced datasets of small enterprise. It can learn the structure of data set independently and adjust the structure spontaneously through the “data shunt”, which can improve the efficiency of model, simplify the complexity of model and optimize the indicators to achieve the goal of reducing bank operating costs and bad debt rates. (1) Supervised data splitter: In this paper, we select the density-based outlier test algorithm Local Outlier Factor (LOF) as the main algorithm of the shunt and the raw data will be reduced and clustered by Principal Component Analysis (PCA) and K-Means++ clustering algorithm, and the outlier test pipeline will be constructed for each cluster data. Finally, the samples identified as abnormal in each cluster will be sent to the “strict channel” to simplify the whole model, speed up the calculation and improve the recall rate and other indicators, whereas the samples identified as normal will be sent to the “loose channel” to speed up the approval of loans to normal customers. (2) Construction of the two-channel model: It has been verified that the “strict channel” has the best performance when using the Bagging model as the main model based on eXtreme Gradient Boosting (XGBoost), and the “loose channel” has the best performance when using the Gradient Boosting Decision Tree (GBDT) single model. In general, the empirical results show that the dual-channel credit evaluation model with shunt could achieve the effect of simplifying the whole model, speeding up the calculation, and optimizing the model indicators. In particular, it is possible to control the loan pass rate under different loan policies (loose or tighten) automatically by setting multiple sets of parameters for the “shunt”.