| With the rapid development of information technology, the Internet and traditional industries have been combined to produce similar Internet banking and other emerging concept. It has completely different analytical methods such as machine learning and data mining. In this paper, one of the three classification models: logistic model, decision tree model, random forest model are used in personal credit assessment.Logistic model is the most widely used method and the reference ability in personal credit assessment. The data comes from UCL database. The classification process use variable selection before and after comparison and different methods of screening variables to derive the classification results.Decision tree is the most influential machine learning methods, with advantages of easy to interpret, high recognition rate, resulting in discrimination rules. To use C5.0 algorithm, introduced post-classification, tree pruning, misjudgment cost matrix, boosting algorithm to get the classification results.Random Forest is integrated by decision trees. The purpose is to compare the effect of classification between the decision tree and random forest. Through adjust various parameters, cost-sensitive learning, to get weighted random forest model, and then ranked the importance of each variable.Finally, to evaluate the above three categories model performance by the ROC curve, AUC values, Lift and other standard curve and generalization estimates of each model. It concluded: Random Forest model has the lowest overall error rate; C5.0 has the lowest error rate class A, but its high class B error rate; there is no one model for the various types of error rate were lower than those of other models.In this paper, the data split into training data, test data, validation data, each of the parameters are constantly tested in order to achieve optimal results, the first to analysis and comparison of each model, and then the three models were compared, when it comes to the model evaluation to use the accuracy and ROC curve and other indicators. In this way to ensure a greater extent that the actual data can be applied, it has some practical reference value to the actual classification requirements. |
