Instance-Level Explanation

An instance-level explanation answers the question: Why did the model make this specific prediction for this specific input? It does not describe how the model works overall. The explanation is generated post-hoc for a single row in a dataset, a single image, or a single text document. This contrasts sharply with global explanations, which summarize feature importance across all predictions. The output is a set of feature attribution scores showing which inputs pushed the prediction toward or away from a particular class.

The core trade-off in instance-level methods is local fidelity—how accurately the explanation mirrors the black-box model's behavior in the immediate vicinity of the instance. A high-fidelity local explanation may be a poor global representation of the model. Key concepts:

• Local decision boundary: The complex prediction surface near the instance
• Locality constraint: Explanations are only valid within a tight neighborhood
• Fidelity-interpretability trade-off: Simpler explanations are easier to understand but may lose precision even locally

Most instance-level methods produce an additive explanation, decomposing the prediction into a sum of individual feature contributions plus a baseline value. This linear structure makes explanations intuitive:

• Each feature receives a weight indicating its contribution magnitude and direction
• Positive weights push the prediction higher; negative weights push it lower
• The sum of all feature attributions plus the baseline equals the model's output
• This framework underlies LIME, SHAP, and Integrated Gradients

A defining characteristic of many instance-level explanation systems is model-agnosticism—the ability to explain any classifier or regressor without accessing its internal structure. The explanation algorithm requires only black-box access: the ability to query the model with inputs and receive predictions. This enables:

• Explaining proprietary or third-party models
• Consistent explanation workflows across different model types
• Auditing ensembles, neural networks, and gradient-boosted trees with the same tool LIME and Anchor Explanations are canonical model-agnostic approaches.

To understand the local decision boundary, instance-level methods generate a synthetic neighborhood around the instance by perturbing its features. The process:

1. Perturbation sampling: Randomly alter or mask features to create variants
2. Model querying: Obtain predictions for all perturbed samples
3. Distance weighting: Assign higher importance to samples closer to the original instance using an exponential kernel
4. Surrogate training: Fit an interpretable model on this weighted dataset The kernel width hyperparameter controls the effective neighborhood size.

Raw features are often not human-understandable. Instance-level methods transform inputs into an interpretable representation before building explanations:

• Images: Raw pixels become superpixels—contiguous, perceptually similar regions
• Text: Words become a bag-of-words or TF-IDF vector presence indicator
• Tabular data: Continuous values may undergo feature discretization into quantiles or bins This transformation ensures the final explanation uses concepts a human can reason about, not raw model features.