Leave-One-Out
An attribution method in machine learning interpretability that assesses feature importance by removing features one at a time and observing changes in the prediction. It is simple to implement, but produces incorrect attributions in scenarios involving interacting features and redundant features.
Timeline (last 90 days)
In the interaction model where prediction is 10 only when the features appear together, leave-one-out attributes 10 points to each of the two features, totaling 20 points, which exceeds the true predicted value of 10 points and violates the efficiency axiom.
In the redundant feature scenario (two features carrying exactly the same information), leave-one-out gives both 0 points, while Shapley gives each copy 4 points.
On a purely additive model, the results of leave-one-out and Shapley are completely identical.
All Facts (3)
In the interaction model where prediction is 10 only when the features appear together, leave-one-out attributes 10 points to each of the two features, totaling 20 points, which exceeds the true predicted value of 10 points and violates the efficiency axiom.
50%UnverifiedIn the redundant feature scenario (two features carrying exactly the same information), leave-one-out gives both 0 points, while Shapley gives each copy 4 points.
50%UnverifiedOn a purely additive model, the results of leave-one-out and Shapley are completely identical.
50%