Interpretability Introduction
Feature effects
Inspect how a prediction responds to a changing input.
Before this
The intuition
One way to inspect a model is to change an input feature while holding other features fixed, then compare predictions. This exposes how the model behaves for those chosen inputs.
A concrete example
For a model f(area, rooms) = 2 × area + 5 × rooms, increasing area by 10 while holding rooms fixed raises the prediction by 20.
What to remember
A model response is not a causal effect in the real world. Changing correlated features independently may also create unrealistic inputs.
Check your understanding
Does a larger prediction after changing a feature prove that the feature causes the real outcome?
Reveal explanation
No. It reveals a relationship inside the model, not a demonstrated causal relationship.