Naive Bayes

---

🎯 Purpose

This QuickRef covers Naive Bayes Classifiers — fast, probabilistic models ideal for text classification, categorical data, and scenarios where independence assumptions are tolerable.

---

📦 1. When to Use

Condition	Use NB Classifier?
You need fast, interpretable results	✅ Yes
Text or count-based features	✅ Yes (use Multinomial NB)
Features are mostly categorical	✅ Yes
Strong feature interactions present	❌ Try trees or SVMs

---

🧮 2. Core Logic

• Based on Bayes' Theorem:

$$ P(y \mid x_1, x_2, ..., x_n) \propto P(y) \prod_{i=1}^n P(x_i \mid y) $$

• Assumes conditional independence between features given the class label

---

🧪 3. Naive Bayes Variants

Variant	Use Case
GaussianNB	Continuous data (assumes normal distribution)
MultinomialNB	Text classification, word counts, discrete data
BernoulliNB	Binary features (e.g., presence/absence)
CategoricalNB	Explicitly labeled categories (since sklearn v0.22)

---

🛠️ 4. Fitting in sklearn

from sklearn.naive_bayes import MultinomialNB
model = MultinomialNB()
model.fit(X_train, y_train)

# For continuous features:
from sklearn.naive_bayes import GaussianNB

---

📊 5. Output Interpretation

Output	Meaning
model.class_log_prior_	Log prior of each class
model.feature_log_prob_	Log likelihood of each feature per class
predict()	Predicted class label
predict_proba()	Class probability estimates

---

⚠️ 6. Limitations

• Assumes independence between features — rarely true, but often effective
• Can be overconfident with highly correlated inputs
• Works best when features map cleanly to likelihoods (e.g., word counts)

---

✅ Checklist

• [ ] Feature type matched to NB variant (Gaussian, Multinomial, etc.)
• [ ] Categorical/binary features encoded cleanly
• [ ] Priors interpreted if class imbalance exists
• [ ] Assumptions of feature independence acknowledged
• [ ] Evaluation includes AUC or log loss if using probabilities

---

💡 Tip

“Naive Bayes assumes the worst — and still often performs surprisingly well.”