基于集成学习的用户信用卡违约预测模型研究

用户信用卡违约预测任务有助于银行等金融机构平衡经济风险与经济利益，对于银行信用卡业务的风险管控具有重要作用。针对用户信用卡违约预测问题，提出了一种基于集成学习的预测模型，有异于传统集成学习中的弱学习器。该模型采用集成模型和神经网络模型作为基学习器，从而提升模型整体的预测效果。首先通过预处理提取用户信用卡数据集的相关特征，然后分别采用优化后的决策树、随机森林、GBDT、XGBoost、CatBoost和SPE六种机器学习模型与神经网络模型进行并行训练和预测，最后通过加权软投票法集成基学习器结果并输出最终预测结果。结果表明，相对于基学习器，该模型在各项评估指标上均有所提升，且拥有更好的模型泛化能力。

A CREDIT CARD DEFAULT PREDICTION MODEL BASED ON ENSEMBLE LEARNING

The user credit card default prediction can help banks and other financial institutions to balance economic risks and interests, and play an important role in risk control of bank credit card business. Aiming at the problem of credit card default prediction, a credit card default prediction model based on ensemble learning was proposed. Being different from the weak learner in the traditional ensemble learning, the ensemble model and the neural network were adopted as the base learners in this model, so as to improve the prediction effect of the ensemble model. Specifically, the relevant features of the user credit card data by pre-processing were extracted. Then the optimized decision tree, random forest, GBDT, XGBoost, CatBoost and SPE and neural network models were adopted to train data and predict results. Finally, the combined strategy (i.e., the weighted soft voting) was used to integrate the results of the base learners and output the final prediction results. It showed that compared with the base learners, the prediction model had improved in all evaluation indicators and had better model generalization ability.