How does Naïve Bayes behave under feature sparsity vs feature density regimes?

Naïve Bayes performs best under sparse feature regimes and degrades as feature density increases with correlation.

In sparse regimes (e.g., text), most features are zero, reducing interaction effects and making independence assumption more reasonable. In dense regimes (e.g., images), feature dependencies increase, violating NB assumptions and reducing performance. This explains NB’s dominance in NLP but weaker performance in vision tasks.