Naive Bayes advantages and disadvantages

Advantages

• Simple, fast and easy to implement; training needs a single scan of the data (counting), and classification is quick.
• Works well with small training sets and is efficient on large and high-dimensional data (e.g. text, spam filtering).
• Gives probabilistic output (class membership probability).
• Handles categorical and continuous attributes, and multi-class problems.
• Tolerates noise and missing values (a missing attribute is simply left out of the calculation).
• Its accuracy is often comparable to that of decision trees and neural networks, and the model is incrementally updatable (new data only update the counts).

Disadvantages

• The class-conditional independence assumption is often false, which can reduce accuracy when attributes are strongly correlated.
• Zero-probability problem when a value is never seen with a class; it needs Laplacian correction.
• Needs prior probabilities; wrong estimates reduce accuracy.
• Continuous attributes are assumed to follow a (normal) distribution, which may not hold.
• Probability estimates are often unreliable, though the ranking of classes can still be right.
• Cannot learn interactions between attributes.