Source-linked AI summary
Interpretable Credit Application Predictions With Counterfactual Explanations
Rory Mc Grath, Luca Costabello, Chan Le Van, Paul Sweeney, Farbod Kamiab, Zhao Shen, Freddy Lecue
TL;DR
Counterfactual explanations can be hard to interpret when they involve many features, and they are mainly designed to change undesired outcomes. This paper adapts them to explain accepted loans and proposes two weighting strategies; on HELOC applications, weighted strategies yield smaller explanations, with global feature importance reducing size by 11.2% on average versus baseline.
Problem
Counterfactuals mainly address undesired outcomes and can be difficult to interpret when explanations involve many features.
Source: §1 · p. 2Method
The paper applies counterfactual explanations to black-box credit classifiers, treating accepted outcomes as safety margins and weighting generation by global feature importance or nearest neighbours.
Source: §1 · p. 2Results
11.2% smaller on average using global feature importance than the baseline; both weighted strategies had smaller mean and standard deviation counterfactual sizes across 5,000 loan applications.
Source: §5.2 · pp. 6–7Takeaways & Limitations
Positive counterfactuals express how much an accepted applicant was accepted by, while smaller counterfactuals offer more compact and intelligible explanations.
Source: §1 · p. 2, §6 · p. 7The effectiveness of these explanations remains to be validated using human-grounded and application-grounded evaluation protocols.
Source: §6 · p. 7
Abstract
from arXiv · showhide
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.
1 Introduction
The paper addresses counterfactuals that poorly explain satisfactory outcomes and can change too many features to remain intelligible. It proposes positive counterfactuals and feature-weighted strategies, reporting smaller explanations on a credit-application dataset.
1 Introduction
- Counterfactuals are intended to explain black-box decisions by describing input changes that would flip a prediction.The example describes a rejected loan becoming approved after changes to income and current balance, with other values held constant.Source: §1 · p. 1
- Positive counterfactuals frame an accepted prediction as a safety margin from the decision boundary, answering how much an applicant was accepted by.Source: §1 · p. 2
- Two weighting strategies use global feature importance or nearest neighbours to generate more compact counterfactuals.The paper treats explanation size as a proxy for interpretability.Source: §1 · p. 2
- Weighted counterfactuals produced smaller explanations, which the paper presents as more interpretable.The introduction reports this result from experiments on a credit application dataset.Source: §1 · p. 2
2 Related Work
Related work spans local explanations of individual predictions, example-based methods, and credit-risk prediction. The paper identifies local interpretability for black-box credit-risk models as an unaddressed area in the literature it reviews.
2 Related Work
- Local interpretability concerns explaining individual predictions rather than a model as a whole.Approaches include transparent models, surrogate models, and model-agnostic explainers such as LIME and SHAP.Source: §2 · pp. 2–3
- Example-based methods include prototypes and criticisms, while counterfactuals describe feature changes that can alter a prediction.Unlike adversarial examples, counterfactual explanations inform how a record’s features must change to influence the predicted outcome.Source: §2 · p. 3
- The reviewed literature lacks work focused on local interpretability for black-box models applied to credit-risk prediction.The review notes that some black-box credit-risk studies evaluate predictive power without explaining predictions locally or globally.Source: §2 · p. 3
3 Preliminaries: Counterfactual Explanations
A counterfactual seeks a nearby input that makes a black-box model return a desired outcome. The method optimizes a loss balancing outcome fit against input change, then iteratively adjusts its settings subject to a tolerance and manual feature constraints.
3 Preliminaries: Counterfactual Explanations
- A counterfactual finds an input x′ close to x for which the black-box model returns the desired output y′.The model’s internal structure is not required; the task is to find an input near x whose prediction is sufficiently close to y′.Source: §3 · p. 3
- The loss balances how closely the model’s output for x′ matches y′ against the distance between x and x′.The method uses Manhattan distance weighted feature-wise by inverse median absolute deviation, yielding sparse solutions in which most entries are zero.Source: §3 · pp. 3–4
- The desired counterfactual preserves most input features while changing only a small number.Source: §3 · p. 4
- An iterative Nelder–Mead procedure adjusts λ until the output error is within tolerance ε, with features constrained manually.Once an acceptable λ is found, the procedure can repeat to obtain a set of counterfactuals.Source: §3 · p. 4
4 Contribution
The paper extends counterfactual explanations to accepted applications and weights them to produce more compact, individualized recommendations.
4.1 Counterfactuals for positive predictions
- Positive counterfactuals treat accepted applications as safety margins from the decision boundary, showing how much the applicant was accepted by.In the binary case, the target is set to P(y = 1) = 0.5; the resulting tolerance indicates how far features may move while approval remains likely, assuming other features stay constant.Source: §4.1 · pp. 4–5
4.2 Weighted Counterfactuals
- Equal feature weighting can ignore individual changeability and immutable features, so the weighted distance promotes discriminative features for more actionable, compact explanations.The paper gives historical delinquencies as a fixed feature and contrasts individuals who may find it easier to increase savings or reduce expenses.Source: §4.2 · p. 5
- Global feature importance weights use ANOVA F-values between each feature and the target to promote discriminative features and reduce the number of features in recommendations.Source: §4.2 · p. 6
- The K-Nearest Neighbors strategy weights locally important features based on nearby successful cases and their relative changes, aiming for more actionable recommendations.The authors say these weights can be learned automatically when applied to new problem spaces.Source: §4.2 · p. 6
5 Experiments
The experiments evaluate black-box credit classifiers and compare weighted counterfactuals on HELOC applications, finding smaller explanations with both weighting strategies.
5.1 Experimental Settings
- The HELOC dataset contained 9,870 records and 22 features after preprocessing, including 5,000 accepted applications.Highly correlated features and duplicate records were removed.Source: §5.1 · p. 6
5.2 Results
- The predictive-power experiment compared logistic regression, gradient boosting, linear-kernel SVC, and MLP using 3-fold cross-validated grid-search model selection.Table 1 presents predictive-power results for the adopted classifiers.Source: §5.2 · p. 6, Table 1 · p. 7
- 11.2% smaller on average than the baseline: explanations using global feature importance had lower mean and standard deviation in general.Source: §5.2 · p. 7, Table 2 · p. 7
- The KNN weighting strategy never performed worse than the baseline, while average counterfactual size could vary dramatically across models for the same data.Source: §5.2 · pp. 6–7
6 Conclusion
The conclusion presents positive counterfactuals for accepted applications and two weighting strategies, reporting more compact explanations with feature-importance weights on HELOC.
6 Conclusion
- The paper uses positive counterfactuals to explain accepted credit applications and proposes weighted strategies based on feature importance and nearest neighbours.Source: §6 · p. 7
- On HELOC applications, feature-importance weights produced more compact counterfactuals, which the paper presents as more intelligible for end users.Source: §6 · p. 7
- Future work will validate counterfactual effectiveness with human-grounded and application-grounded protocols, including whether smaller explanations are more interpretable.Source: §6 · p. 7