Research On Credit Evaluation Method Based On Feature Selection And Deep Rotation Forest

The research and application of cloud computing,big data and artificial intelligence technologies are constantly accelerating the process of social informatization,and the financial distress prediction,credit risk warning and credit score of small and micro enterprises have become the new trend of current financial development.In recent years,the unsound credit evaluation system has seriously affected the credit construction of small and micro enterprises in China,resulting in credit financing difficulties for small and micro enterprises and seriously restricting the development of small and micro enterprises in China.Therefore,how to effectively solve the credit evaluation problems faced by small and micro enterprises has become an important research direction of relevant researchers at home and abroad.The paper will take the project of "small and micro enterprises credit loss prediction" as the engineering background,conduct in-depth research and analysis of the existing credit assessment methods,sort out the relevant knowledge of small and micro enterprises credit assessment,and summarize the research progress.With the development of Internet finance,the access to data of small and micro enterprises has changed.Based on the real data generated by Chinese small and micro enterprises in the industry,the paper conducts an empirical study,from the processing of credit data of small and micro enterprises to the development and design of evaluation model.By using advanced technology tools such as natural language processing machine learning and group intelligent optimization,the paper completes the data visualization feature engineering and the establishment of credit evaluation model,and realizes the credit evaluation system to predict the credit loss of small and micro enterprises.The main research work is as follows:(1)In order to overcome feature redundancy and interference in credit evaluation data of small and micro enterprises,the paper proposes a feature selection method based on fitness evaluation index.The commonly used search strategies,such as forward sequence search,backward sequence search and recursive feature elimination,need to specify the number of feature output when searching for features,which is a feature sorting method and cannot directly get a more appropriate feature subset.The algorithm adopts gray Wolf optimization algorithm in search strategy,and can select a group of better featuresthrough the evaluation function of feature subset.Based on the original evaluation function,the paper constructs a new evaluation index to avoid the impact of the accuracy or error rate of a single measurement model on the classification of unbalanced data.In order to optimize the accuracy and local stoning of the search strategy,the corresponding improvement of the cross-optimization method and the dynamic elite information cross method is used to make the corresponding improvement,and then the binary transformation is carried out to solve the characteristics combination optimization problem,search the feature space,and find the better subset of the characteristics.Finally,the effectiveness of the improved feature selection algorithm in the combination optimization of characteristics is verified by experiment.(2)For small micro enterprise in the construction of credit evaluation system of the data in the unbalance and the deep forest model in prediction of the effect not beautiful,the paper puts forward a few kinds of synthetic sampling method and depth revolving credit evaluation model of forest,to improve the data distribution imbalance and to strengthen the learning ability of the model.Firstly,in the preprocessing of the data,a small number of synthetic sampling methods with Shared neighbors are used to balance the original data set;Secondly,the rotation transformation strategy is introduced into the model structure of deep forest to improve the diversity of sample learning of each layer model;Thirdly,Characteristics of each layer model after studying again,according to each sample in the model into leaf node position,for hot code to generate new eigenvector alone,add to the original features of the follow-up study,make up the depth of the original forest was carried out in each layer of feature extraction of the abstract characteristic quantity is less,the data representation learning ability is insufficient;Finally,the effectiveness of the new model is verified by comparing the results of evaluation indexes.The experimental results show that the new model performs better in the credit evaluation data of small and micro enterprises than other models.