How does Naïve Bayes compare to probabilistic graphical models with latent structure?

Naïve Bayes is a fully observed graphical model, while latent graphical models introduce hidden variables to capture structure.

Naïve Bayes assumes all dependencies are mediated through the class variable. Latent variable models (e.g., LDA) introduce hidden structure between features and classes, enabling richer representation of data dependencies. This improves expressiveness at the cost of computational complexity and inference difficulty.