Personal Credit Scoring Model Based On Improved PSO-BP Neural Network

With the advent of the era of big data,credit data presents the characteristics of complex structure,the features of multi-dimension and high noise,resulting in the low prediction accuracy of the traditional machine learning model currently used by financial institutions.On the other hand,due to the privacy of financial data,the training data of a single financial institution may have the problems of shortage and poor quality of data.To solve the above problems,the contents of this study are organized as follows:1.In view of the problems of high dimension,high noise and non-linearity of credit data existing in financial institutions in recent years,several evaluation methods of personal credit evaluation models are compared.Among them,BP neural network model has more ability to fit high-dimensional and non-linear data,so BP network is chosen as the basic model in this thesis.2.The gradient descent function of the optimization parameters in the BP network is easy to fall into the local optimum and the model accuracy is not high.An improved PSO(CS-GA-PSO)optimization algorithm is proposed to optimize the parameters of BP.Firstly,the principle of PSO algorithm is analyzed,and then the common improvement methods of PSO are analyzed and introduced.On this basis,an improved PSO method is proposed,particles are divided into dominant and disadvantage particles according to different fitness by using the cross-mutation operation of adaptive genetic algorithm.Advantage particles focus on searching solution space for better accuracy,while disadvantage particles focus on the whole spatial search,which expands the solution space of search and improves the convergence speed.Finally,the perturbation method of chaotic search tent mapping is adopted for particles trapped in the local optimum,and the random and ergodic characteristics of the chaotic mapping can be used to traverse the solution space without repetition on the basis of the optimal solution to ensure that the particles can quickly get rid of Local optimal solution,thereby improving the accuracy of the algorithm model.Meanwhile,the improved algorithm is compared with other PSO improvement methods and standard PSO in the algorithm simulation.The results show that the CS-GA-PSO method in this study is superior to other algorithms in convergence speed and accuracy.3.The study builds the CS-GA-PSO-BP model on the framework of horizontal federated learning.By deploying the BP network model and local dataset to the local client,the CS-GA-PSO optimization algorithm is deployed to the central server.The client will upload the fitness of each particle to the central server through the model and local data calculation,and the central server will calculate the particle fitness through aggregation.The CS-GA-PSO optimization process is completed according to the fitness,and the parameters of each particle are sent out for the next round of fitness calculation.Through the transfer of parameters,the process of joint training model can be completed for each client without sending local data outside the client.It solves the problem that multiple financial institutions with less local data or poor data quality cannot jointly model due to the sensitivity of financial credit data..