ViCE: Visual Counterfactual Explanations for Machine Learning Models

1 Introduction

• ViCE targets end-users seeking to understand individual predictions from models used in complex, high-risk applications.The introduction names client-facing people such as doctors and admissions officers as potential users.Source: §1 · p. 1
• Its explanations suggest changes that could alter a model’s prediction, such as changes a loan officer might consider for a rejected application.The tool describes what changes are required to alter the current state.Source: §1 · p. 1

2 Background and Related Works

• ViCE’s model-independent explanation algorithm works without access to a model’s internal logic.Black-box models are described as hard to understand because of their complex structures.Source: §2 · p. 2
• Counterfactual explanations identify minimal feature-value changes that switch an instance’s prediction, complementing feature-contribution approaches such as LIME [18] and SHAP [11].The section also describes prior counterfactual-generation approaches by Wachter et al. and Ustun et al. [21].Source: §2 · p. 2
• Visual analytics systems support understanding, debugging, verification, and refinement, while ViCE focuses on how inputs can be changed to reach another prediction.The section contrasts ViCE’s question with WIT [23], which shows how predictions change after data are inferred.Source: §2 · p. 2
• Earlier black-box counterfactual tools addressed target-class changes, but the cited algorithm was designed for text and works only on binary inputs; ViCE extends it to continuous numerical data.The passage names Rivelo and Krause et al. as related tools.Source: §2 · p. 2

3 ViCE

• ViCE presents local counterfactuals for numeric features by showing which features must change and by how much to alter an instance’s prediction.Its iterative design process considered end-user questions about distributions, relevant features, possible changes, and actionable changes.Source: §3 · p. 2

3.1 Counterfactual Algorithm

• The algorithm discretizes each feature into Gaussian-based bins, then greedily moves feature values toward the opposite class until the prediction flips or constraints stop the search.The constraints limit the number of changed features and the number of bins crossed; the search chooses the move producing the largest output change at each iteration.Source: §3.1 · p. 2

3.2 Visual Interface

• The interface pairs prediction probability and correctness with feature values positioned against dataset densities, and displays suggested counterfactual changes and controls.Users can lock features, sort by standardized values, and select distributions for all, positive, or negative cases; the diabetes example visualizes a single local explanation.Source: §3.2 · pp. 2–3, Figure 1 · p. 3
• Counterfactuals are most often found for predictions near the 50% cutoff; predictions far above or below it may require changes beyond the algorithm’s limits.The interface can still support anomaly exploration when no counterfactual is available, and can accommodate other counterfactual-generation methods [9] [16] [21].Source: §3.2 · p. 3

3.3 Implementation

• ViCE accepts binary-classification datasets in CSV format, custom models with probability prediction, and real-time user inputs; its implementation uses 10 bins and sets w = l = 5.The default model is an SVM trained with scikit-learn; the application uses Flask, Python, D3, and JavaScript.Source: §3.3 · p. 3

4 Case Study

• For the selected HELOC application, the model predicted a negative outcome with 29% probability, matching the ground truth.Feature-density views and sorting helped identify unusual values and compare the instance with other negative samples.Source: §4 · p. 4, Figure 2 · p. 4
• The counterfactual suggested increasing Number of Satisfying Trades and Months Since Most Recent Delinquency substantially, and Number of Inquiries in the Last 6 Months excluding last 7 days slightly.The authors note that time-based changes might be infeasible for a client seeking quick approval.Source: §4 · p. 4
• After time-dependent features were locked, the algorithm could not find a decision-changing explanation within its constraints; unlocking External Risk Estimate suggested large changes to that feature.The authors interpret this as evidence that time variables weigh highly in this instance’s decision, and that External Risk Estimate is strongly correlated with the model decision.Source: §4 · p. 4

• Categorical features are not effectively handled; proposed options are presetting a search path or brute-force analysis of known categorical features.Source: §5 · p. 4
• The tool does not extend to multiclass classification or contexts such as image classification.Source: §5 · p. 4
• The visualization can display around 30 features, while larger datasets can be accommodated by sorting and displaying only the top k features.Source: §5 · p. 4

6 Conclusions and Future Work

• ViCE visualizes counterfactuals for non-binary data, showing each sample’s minimal changes needed to alter the model’s decision.Source: §6 · p. 4
• Users can customize explanations according to their requirements, and the tool was applied in a loan-dataset use case.Source: §6 · p. 4
• ViCE’s modular, black-box-based design allows integration of improved counterfactual-generation methods or multiple alternative explanations.Source: §6 · p. 4
• Future work includes letting users inspect custom changes, adding explanation methods and local-importance sorting, and aggregating instance explanations for model developers.Source: §6 · p. 5