Credit Scoring Based On Self Adaptive Ensemble Tree

In this paper,an adaptive integrated learning algorithm based on decision tree is proposed,which uses self-organizing neural network as framework,decision tree as neuron to construct competitive subgroup,iterative method is used for competitive training among neurons,and Bayesian method is used to dynamically give weight to each decision tree in data output.Based on this method,this paper models the artificial data,German credit data,Australian credit data,domestic P2 P network loan platform data and Lending Club network loan platform data,and compares the performance with other traditional integrated learning algorithms based on decision tree.The comparison shows that when the data sample has a more obvious sub sample boundary and the data sample can be divided,for example,when the artificial data and the domestic P2 P network loan platform data,the adaptive integration decision tree can obtain higher prediction performance under the lower model complexity;when the data sample does not have a more obvious boundary,for example,the German credit data and the domestic P2 P network loan platform data.For Australian credit data,the prediction performance of the adaptive integrated decision tree is similar to that of the common decision tree,which is lower than that of the random forest and the gradient lifting tree.For the data set without sub sample boundary,this paper adopts the same voting method as random forest to improve the adaptive integration decision tree,and the improved prediction performance has a significant improvement;for the data set with sub sample boundary,the dynamic weight integration method based on Bayesian in this paper has a good effect.In the case of manual data and domestic P2 P network loan platform data,a better prediction performance can be obtained without integration,and the complexity of the model is far lower than other integrated learning models;in the case of German credit set and Australian credit set,the effect of not adopting integration method is slightly worse,but after joining the integration,the effect is significantly improved,which can be achieved high level of various models.