Abstract: With the development of computer technology,using machine learning algorithms to build automated evaluation models has become an important tool to for the financial institutions to conduct credit evaluation.However,currently,the credit evaluation model is still facing challenges:credit data is class-imbalanced and high-dimensional,meanwhile,the behavior of customers can be influenced by the changeable external environment,namely,the concept drift will occur.As a result,this paper proposes a dynamic credit evaluation model,which can achieve the flexible model update by using ensemble learning algorithm to continuously add base classifiers trained on new incremental data,and dynamically adjusting the weight of each base classifier to adapt to concept drift.When concept drift occurs,according to the detection results of concept drift,the model is able to use different forms of equalization and feature selection on credit data.In particular,for feature selection,this paper proposes an incremental feature selection algorithm combining the choice of representative samples that makes the feature selection efficient and accurate,enabling the model to simultaneously process the high-dimensional imbalanced data and adapt the concept drift of incremental credit data.Finally,this paper manages to demonstrate that the proposed dynamic model is more efficient and accurate than other prevailing algorithms on real incremental high-dimensional credit datasets.

Key words: Credit evaluation, Feature selection, Concept drift, Sliding window, Dynamic model