[论文解读] Interpretable Credit Application Predictions With Counterfactual Explanations

Rory Mc Grath, Luca Costabello+5 more|arXiv (Cornell University)|Nov 13, 2018

Financial Distress and Bankruptcy Prediction被引用 88

一句话总结

本文将反事实解释应用于信贷决策，使其更具可解释性，提出对已批准贷款的正向反事实，并引入两种加权策略（全局特征重要性和最近邻）以在黑箱信用模型中产生更小、可操作性更强的反事实。

ABSTRACT

We predict credit applications with off-the-shelf, interchangeable black-box classifiers and we explain single predictions with counterfactual explanations. Counterfactual explanations expose the minimal changes required on the input data to obtain a different result e.g., approved vs rejected application. Despite their effectiveness, counterfactuals are mainly designed for changing an undesired outcome of a prediction i.e. loan rejected. Counterfactuals, however, can be difficult to interpret, especially when a high number of features are involved in the explanation. Our contribution is two-fold: i) we propose positive counterfactuals, i.e. we adapt counterfactual explanations to also explain accepted loan applications, and ii) we propose two weighting strategies to generate more interpretable counterfactuals. Experiments on the HELOC loan applications dataset show that our contribution outperforms the baseline counterfactual generation strategy, by leading to smaller and hence more interpretable counterfactuals.

研究动机与目标

• Explain why a loan was accepted as well as why it was rejected using counterfactuals (positive and negative explanations).
• Improve interpretability by reducing the complexity of counterfactuals through feature weighting.
• Demonstrate the effectiveness of weighting strategies in producing smaller, more actionable counterfactuals on a real credit dataset.
• Evaluate the predictive power of off-the-shelf black-box classifiers used in the pipeline.

提出的方法

• Treat the predictive model as a black box and optimize a counterfactual loss to find minimal input changes that flip the outcome.
• Use Manhattan distance weighted by inverse MAD to measure input changes and promote sparse counterfactuals.
• Iteratively optimize with Nelder-Mead while adjusting a lambda parameter to satisfy a target proximity to the desired outcome.
• Introduce positive counterfactuals by setting the target outcome to the decision boundary (e.g., P(y=1)=0.5) to explain how much the applicant was accepted.
• Propose two weighting schemes (global feature importance via ANOVA F-values and a K-Nearest Neighbors approach) to weight feature changes and generate more compact counterfactuals.

实验结果

研究问题

• RQ1 How can counterfactual explanations be adapted to explain positive credit decisions (accepted loans) as safety margins?
• RQ2 Do weighting strategies (global feature importance and nearest neighbors) yield smaller and more interpretable counterfactuals compared to baseline counterfactuals?
• RQ3 How do off-the-shelf black-box classifiers perform in credit prediction and how do counterfactual explanations relate to their decisions?

主要发现

• Weighted counterfactuals produce smaller explanations on average, improving interpretability.
• Global feature importance weighting yields counterfactuals 11.2% smaller on average than baseline for the tested models.
• KNN-based weighting also improves interpretability and never performs worse than baseline in the reported results.
• Among evaluated models, GradBoost and Logistic Regression achieved comparable F1 and accuracy, with F1 up to 0.72 and accuracy up to 0.74 on HELOC data.
• Positive counterfactuals provide a tangible tolerance from the decision boundary, indicating how much was the applicant accepted.