What are example-based explanations in Explainable AI?

Example-based explanations in Explainable AI (XAI) are methods that clarify how a machine learning model makes decisions by showing specific instances or data points. These explanations work by highlighting examples that the model considers similar (prototypes), cases where small changes would alter the outcome (counterfactuals), or influential training samples that shaped the model’s behavior. The goal is to make complex model behavior tangible by grounding explanations in concrete data, which developers can inspect and relate to real-world scenarios.

For instance, consider a loan approval model. A prototype-based explanation might show an average profile of applicants who were approved, like someone with a 700 credit score, $60k income, and no defaults. A counterfactual explanation could demonstrate that a rejected applicant with a 680 credit score and $58k income would have been approved if their income increased by $5k. Influential examples might reveal that the model heavily weights historical data from applicants in a specific region. These examples help developers verify if the model’s logic aligns with domain expectations—like whether income thresholds are reasonable—or uncover biases, such as over-reliance on geographic data.

Implementing example-based explanations requires techniques like clustering (to find prototypes) or optimization methods (to generate counterfactuals). For example, using k-nearest neighbors (KNN), a developer might retrieve the most similar cases to a prediction to show prototypes. To create counterfactuals, tools like adversarial perturbation or genetic algorithms can tweak input features until the model’s output changes. However, challenges include computational costs for large datasets and ensuring examples are meaningful (e.g., a “+$5k income” counterfactual must be actionable for a loan applicant). Despite these hurdles, example-based methods are practical because they align with how humans reason—using specific cases rather than abstract rules—making them particularly useful for debugging models or communicating results to stakeholders.