群智能算法优化XGBoost的信贷风险预测

Optimization of XGBoost Credit Risk Prediction Using Swarm Intelligence Algorithm

为改善极端梯度提升(extreme gradient boosting,XGBoost)集成算法的信贷风险预测准确率,提出了一种改进的麻雀算法(improved sparrow search algorithm based on golden sine search,Cauchy mutation and oppositionbased learning,GCOSSA)来优化XGBoost参数。采用黄金正弦搜索策略来更新发现者位置,既增强全局搜索能力又增强局部搜索能力;在算法中引入反向学习策略和柯西变异进行扰动来扩大搜索领域改善陷入局部最优,同时使用贪婪规则确定最优解;将改进的算法用6个基准函数进行测试,并对SSA和GCOSSA进行对比,评估GCOSSA寻优性能;用GCOSSA优化XGBoost参数。在数据集上测试,并与网格搜索寻优、SSA及其混合正余弦改进算法(improved sparrow search algorithm based on sine and cosine,ISSA)方法进行对比。结果表明改进后的GCOSSA优化XGBoost参数,在信贷风险预测中准确率更高。

In order to improve the accuracy of integrated algorithm XGBoost(extreme gradient boosting)in credit risk prediction,an improved sparrow search algorithm(sparrow search algorithm based on golden sine search,Cauchy mutation and opposition-based learning,GCOSSA)is proposed to optimize the parameters.Firstly,the golden sine search strategy is adopted to update the discoverer's position,which enhances both global and local search ability.Secondly,the opposition-based learning and Cauchy mutation are introduced into the algorithm to expand the search field to find the optimal solution,so as to avoid falling into local optimum,using greedy rules to determine the optimal.Then,the GCOSSA algorithm is tested with 6 benchmark functions,and compared with SSA to evaluate the optimization performance.Finally,the GCOSSA algorithm is used to optimize the parameters of XGBoost.The results show that the GCOSSA algorithm has higher accuracy than the grid search algorithm,the SSA and the ISSA(improved sparrow search algorithm based on sine and cosine)to optimize the parameters of XGBoost in credit risk prediction.