Naive Bayes

# ===== The math (in one line) =====
# P(class | features) ∝ P(class) * Π P(feature_i | class)
# The 'naive' part: assume features are conditionally independent given the class.
# Wrong in practice; works astonishingly well anyway, especially on text.

# ===== Variants =====
# GaussianNB     features are continuous (assumed Gaussian per class)
# MultinomialNB  feature counts / frequencies (e.g. word counts)
# BernoulliNB    binary features (word presence)
# ComplementNB   imbalanced text; often best for short docs

# ===== Quick text classification =====
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import Pipeline
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

texts = [
    'great service highly recommended',
    'rude staff bad food',
    'wonderful experience kind people',
    'never coming back terrible',
    'love this place amazing',
    'awful worst meal',
]
labels = [1, 0, 1, 0, 1, 0]

Xtr, Xte, ytr, yte = train_test_split(texts, labels, test_size=0.33, random_state=42)

pipe = Pipeline([
    ('tfidf', TfidfVectorizer(ngram_range=(1, 2), min_df=1)),
    ('nb', MultinomialNB(alpha=1.0)),    # alpha = additive smoothing
])
pipe.fit(Xtr, ytr)
print(classification_report(yte, pipe.predict(Xte), zero_division=0))

# ===== Continuous features =====
from sklearn.naive_bayes import GaussianNB
from sklearn.datasets import load_iris
from sklearn.model_selection import cross_val_score
X, y = load_iris(return_X_y=True)
print(cross_val_score(GaussianNB(), X, y, cv=5).mean())

# ===== When NB wins =====
# - Spam / not-spam, sentiment (text classification)
# - Tiny + medium datasets (works well with little data)
# - Real-time + low-memory inference (great for edge / mobile)
# - Multi-class with clear class structure
# - A sane baseline for ANY classification task

# ===== When NB hurts =====
# - Features that are STRONGLY correlated (the independence assumption hurts)
# - High-dimensional regression (it does classification, not regression)
# - When calibrated probabilities matter (NB probabilities are notoriously poorly calibrated)
# - Image / structured-data classification (use trees / nets)

# ===== Calibration =====
from sklearn.calibration import CalibratedClassifierCV
clf = CalibratedClassifierCV(MultinomialNB(), method='isotonic', cv=5)

# ===== Why so fast =====
# Training and prediction are both O(n*d) — linear in data size and feature count.
# No iterative optimisation; closed-form parameter estimation from counts.

# ===== Smoothing (alpha) =====
# alpha = 1.0       Laplace smoothing (default)
# Higher alpha      smoother / less overconfident
# alpha = 0         pure MLE; zero counts -> zero probability -> always wrong on unseen
# Tune via cross-validation.

# ===== Patterns to internalise =====
# - Start with MultinomialNB + TF-IDF for any text classification baseline
# - Pipeline it: vectoriser + classifier, fit once, score on held-out
# - Cross-validate to pick alpha + ngram_range
# - Treat NB as a baseline; only move to more complex models when it underperforms

# ===== Pitfalls =====
# - Negative-valued features with MultinomialNB (it expects non-negative)
# - Trusting raw NB probabilities for thresholding (calibrate first)
# - Ignoring class imbalance (use class_prior or ComplementNB)
# - Over-cleaning text removes useful signal (lowercasing is fine; stripping all punctuation can hurt)

from sklearn.naive_bayes import MultinomialNB
clf = MultinomialNB().fit(X_train_counts, ytr)  # great for text classification