Global vs Local Explanations
You'll learn: The difference between explaining a model's overall behavior across all predictions versus explaining why it made one specific prediction.
Two Lenses on Model Behavior
Imagine you're a food critic. You could write about a restaurant's overall cuisine philosophy and signature ingredients (global), or you could review the specific dish you ate last night (local). Both perspectives are valuable, but they answer different questions.
In interpretability, this distinction is fundamental:
Global Explanations
Global explanations describe how your model behaves in general, across your entire dataset or input space. They answer: "What patterns does this model rely on overall?"
Examples include:
- Feature importance scores averaged across all predictions
- Decision boundaries visualized across the feature space
- Summary statistics about which features matter most globally
Use global explanations when you need to audit the model, understand its general strategy, or communicate to stakeholders what the model "learned."
Local Explanations
Local explanations focus on a single prediction. They answer: "Why did the model make THIS specific decision for THIS particular input?"
Examples include:
- Attribution scores for one instance's features
- Counterfactual reasoning ("if this patient's age were 5 years older...")
- Highlighting which words influenced a specific sentiment classification
Use local explanations when you need to debug individual errors, provide decision justifications to end-users, or understand edge cases.
When to Use Which
Global methods help you understand systematic biases and general model behavior. Local methods help you trust (or question) individual high-stakes predictions, like loan approvals or medical diagnoses.
Key Takeaway: Global explanations reveal overall model patterns and are useful for auditing; local explanations justify individual predictions and are essential for instance-level trust and debugging.