What is the difference between soft voting and stacking in ensemble learning?

Soft voting averages predicted probabilities, while stacking learns how to combine model outputs using a meta-learner.

Soft voting is a fixed aggregation method where class probabilities from multiple models are averaged (or weighted averaged) and the class with highest probability is selected. It assumes all models contribute linearly and equally or proportionally. Stacking is more flexible: it trains a second-level model that learns nonlinear relationships between base model outputs. Stacking can capture when certain models are more reliable in specific regions of feature space, whereas soft voting cannot adapt dynamically.