How does Naïve Bayes behave under Bayesian posterior collapse in extremely high-confidence regimes?

Naïve Bayes can produce posterior collapse where one class dominates with near-1 probability due to multiplicative likelihood amplification.

Posterior collapse occurs when repeated multiplication of small likelihood terms results in extreme log-sum differences between classes. Even moderate feature evidence accumulates additively in log-space, producing highly confident predictions. While useful for classification, this leads to poor probability calibration and brittle decision thresholds in downstream systems.