Why Generative Models Handle Missingness More Naturally

You have millions of labeled examples and care only about predictive accuracy on the deployed label, not about simulating x. Which side usually deserves the first try?