Accuracy is a problem. NBs are linear classifiers, and more so, the linear boundary is found by assuming features to be independent. So, in cases: (1) data ca

Accuracy is a problem. NBs are linear classifiers, and more so, the linear boundary is found by assuming features to be independent. So, in cases: (1) data cannot be separated linearly - NBs don't do well for the same reason as any other linear classifier wouldn't work well (2) for linearly separable data - if the feature independence assumption is incompatible with the data, linear classifiers that don't make this assumption will beat it. For ex you can prove that logistic regression beats (or is at least as good as) Naive Bayes asymptotically. [1] [1] See proposition 1 here: "On Discriminative vs Generative classifiers ... Andrew ng, Michael Jordan. https://ai.stanford.edu/~ang/papers/nips01-discriminativegenerative.pdf https://ai.stanford.edu/~ang/papers/nips01-discriminativegen... [PDF]