Personal Credit Evaluation Method Study Based On Hybrid Data Mining Model

With the popularity of unsecured loan,personal credit has become the focus of various financial institutions.Personal credit assessments rely on artificial judgment in the past,which is not only inefficient but also with low accuracy rate.As the development of artificial intelligence,various classification algorithms began to be active in financial markets.However,there are many kinds of classification algorithms,and the use of each algorithm differs from each other.It is the key point for researchers to select the appropriate algorithm according to the need.The purpose of this thesis is to establish a personal credit evaluation system using machine learning algorithm,the main work can be summarized as follows:In the first part,the BP neural network is used to process the data set.First of all,this thesis establishes a suitable network structure according to the data distribution characteristics of the data set,and then verifies the rationality of the parameter settings of BP neural network through a lot of experiments to select appropriate parameters training network and give the test results.The second part is to optimize the BP neural network.GA-BP network optimizes the BP neural network using genetic algorithm.An appropriate fitness function is set to select the best network initial weights and thresholds through the operation of selection,crossover and mutation.At the other side,the PSO-BP network looks for the optimal initial weights and thresholds through particle optimization.Both optimization algorithms improve the classification performance of the network.The third part is to integrate the BP neural network.BP-Adaboost trains the network by changing the sample weights and then integrates the trained network.GASEN algorithm uses genetic algorithms to evolve the network and set the threshold to screen the network.Only partial classifiers are selected for integration in the end.The fourth part is to improve the GASEN algorithm.the algorithm is improved on the basis of GASEN and DCGASEN algorithm is proposed.DCGASEN algorithm uses the method called roulette wheel for random sampling of data characteristics in order to increase the discrepancy of the trained classifier.Then genetic algorithm is used to evolve the weight.The base classifiers which are selected will be integrated to get the classification results in the final.In order to make the results more convincing,this thesis also compares the commonly used classification algorithms with that in this thesis.The results show that the optimized algorithm in this thesis is better than the commonly used;GA-BP algorithm and PSO-BP algorithm improves the accuracy of BP neural network;The integration algorithm is superior to the single classifier;GASEN algorithm has an advantage over the BP-Adaboost algorithm and the selective integration algorithm can integrate less network but achieve higher accuracy;DCGASEN algorithm has obtained better classification results by increasing the discrepancy of the base classifier,which has been greatly improved on the basis of GASEN.