基于特征加权的反事实解释方法:以信贷风控场景为例

Feature-weighted Counterfactual Explanation Method:A Case Study in Credit Risk Control Scenarios

机器学习技术在金融领域的应用越来越多,为用户提供可解释的机器学习方法已成为一个重要的研究课题。近年来,反事实解释引起了广泛关注,它通过提供扰动向量来改变分类器得到的预测结果,从而提高机器学习模型的可解释性。但现有方法存在生成的反事实用例缺乏可行性和可操作性的问题。文中提出了一种新的反事实解释框架,通过引入特征变量代价权重矩阵的概念,考虑不同特征变量改变的难易程度,使得反事实结果更符合实际情况并更具可行性。同时,通过专家预定义特征变量代价权重矩阵的方式,提出了一种计算特征变量代价权重的可行方法,并允许用户根据实际情况进行个性化调整。定义的目标函数综合考虑了特征加权距离、稀疏性和接近性3个指标,确保了反事实结果的可行性、简洁性和接近原始样本集的性质。采用遗传算法来求解问题,进而生成最佳的行动方案。通过对真实数据集进行实验,证实了所提方法相比现有的反事实方法能够生成可行性和可操作性更强的反事实用例。

The application of machine learning technology in the financial field is becoming more and more prevalent,and providing interpretable machine learning methods to users has become an important research topic.In recent years,counterfactual explanation has attracted widespread attention,which improves the interpretability of machine learning models by providing perturbation vectors to change the predicted results obtained by classifiers.However,existing methods face feasibility and operability issues in generating counterfactual instances.This paper proposes a new counterfactual explanation framework that introduces the concept of feature-variable cost weight matrix,considering the ease of changing different feature variables to make the counterfactual results more realistic and feasible.At the same time,by predefining the feature-variable cost weight matrix by experts,a feasible method for calculating the cost weight of feature variables is proposed,allowing users to make personalized adjustments according to actual situations.The defined objective function comprehensively considers three indicators:feature-weighted distance,sparsity,and proximity,ensuring the feasibility,simplicity,and closeness to the original sample set of counterfactual results.Genetic algorithms are used to solve the problem and generate the optimal action plan.Through experiments on real datasets,it is confirmed that our method can generate feasible and actionable counterfactual ins tances compared to existing counterfactual me-thods.