Researches About Ensemble Selection Learning Method Integrated With GRASP Algorithm

In recent years, credit assessment has received more and more attention of financial institutions because the accuracy of the assessment will seriously affect the financial institution's loss. At present, the scholars have proposed many methods to solve the problem of credit scoring.These algorithms are generally divided into two categories: based on statistical and non-statistical methods. The former mainly include linear discriminant analysis, logit analysis and probit analysis,etc. While the latter mainly include Support Vector Machine, Artificial Neural Network and Decision Tree, etc. Although the researches shows that the non-statistical has better classification performance, the single model is used to solve the problem of credit scoring has some limitations.So the scholars have proposed that solve this problem by ensemble learning method. However,ensemble learning requires a lot of classifiers, which increase the time and space complexity. And the classifiers of low generalization performance will affect the final classification results. To solve these problems, we can select a sub set from the original ensemble to construct a new ensemble system. This method is named as ensemble selection, which is also called ensemble pruning.In this paper, the ELMsGraspEnS algorithm is proposed to solve the credit assessment problem. The ELM is used as the base learner of this method to generate the ensemble system.The GraspEnS is used as the ensemble pruning method to select a subset of original ensemble system. Therefore, the ELMs GraspEnS inherit the advantages the ELM and GraspEnS. The learning speed of ELM algorithm is very fast, and the generalization performance is good, and the problems of local optima and overfitting could be effectively solved by the ELM method. The algorithm GraspEnS is an application of GRASP method in ensemble pruning, and it is a heuristic algorithm for combinatorial optimization. The GraspEnS not only inherit the features of the greedy ensemble pruning method but also avoid the problem of local optima of it. In addition, the GraspEnS also realizes the muti-start search. And experimental results show that the method ELMsGraspEnS has good classification performance.However, the method GRASP is a non memory algorithm, that is, the GRASP method can not use the information of previous iteration in every iterative process. Path-Relinking is a strengthening algorithm, and the problem exited in the GRASP method could be solved by incorporating GRASP with Path-Relinking. In view of this, this paper proposes a new PRlinGraspEnS algorithm to solve the credit risk assessment problem, which is also based on the ELM and it is used as the base learner. The difference of these two methods is that the Baggingtechnology is used into the latter method, which increases the diversity of the classifiers. The PRlinGraspEnS uses the GRASP with Path-Relinking as the ensemble selection method. The new method has the advantages of the GRASP and Path-Relinking algorithm, and the PRlinGraspEnS is also a memory method. The experimental results also show that the new proposed method PRlinGraspEnS has good generalization performance and can speed up the convergence rate.