Naive Bayes for text in NLP - Model Metrics & Evaluation

Metrics & Evaluation - Naive Bayes for text

Which metric matters for Naive Bayes text classification and WHY

For text classification using Naive Bayes, accuracy, precision, and recall are important. Accuracy shows overall correct predictions. Precision tells us how many predicted positive texts are truly positive. Recall shows how many actual positive texts were found. We choose metrics based on the task. For spam detection, high precision avoids marking good emails as spam. For detecting harmful content, high recall avoids missing bad texts.

Confusion matrix example

      Actual \ Predicted | Positive | Negative
      -------------------|----------|---------
      Positive           |    80    |   20
      Negative           |    10    |   90
    

Here, TP=80, FN=20, FP=10, TN=90. Total samples = 200.

Precision = 80 / (80 + 10) = 0.89

Recall = 80 / (80 + 20) = 0.80

Accuracy = (80 + 90) / 200 = 0.85

Precision vs Recall tradeoff with examples

Imagine a spam filter:

• High precision means most emails marked as spam really are spam. This avoids losing good emails.
• High recall means catching most spam emails, but might mark some good emails wrongly.

For harmful content detection:

• High recall is key to catch all harmful texts, even if some good texts are flagged.
• High precision reduces false alarms but might miss some harmful texts.

Naive Bayes can be tuned to balance these by adjusting thresholds.

What good vs bad metric values look like

Good metrics for Naive Bayes text classification:

• Accuracy above 80% on balanced data
• Precision and recall both above 75%
• F1 score (balance of precision and recall) above 0.75

Bad metrics:

• Accuracy near random guess (e.g., 50% for two classes)
• Precision very low (e.g., 30%) means many false positives
• Recall very low (e.g., 20%) means many missed positives

Common pitfalls in metrics for Naive Bayes text classification

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced (e.g., 95% accuracy but model always predicts majority class).
• Data leakage: If test data leaks into training, metrics look unrealistically good.
• Overfitting: Very high training accuracy but low test accuracy means model memorizes training data, not generalizing.
• Ignoring class imbalance: Metrics like accuracy alone don't show if minority class is well detected.

Self-check question

Your Naive Bayes text classifier has 98% accuracy but only 12% recall on the positive class (e.g., spam). Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means it misses most positive cases (spam). Even though accuracy is high, the model mostly predicts negative class. This is bad if catching positives is important.

Key Result

Precision and recall are key to evaluate Naive Bayes text classifiers, balancing false alarms and missed detections.

Practice

1. What is the main assumption behind the Naive Bayes algorithm when used for text classification?

easy

A. Words always appear in a fixed order

B. Words in a document are independent of each other given the class label

C. All documents have the same length

D. The frequency of words does not affect classification

Solution

1. Step 1: Understand Naive Bayes assumption

   Naive Bayes assumes that each feature (word) is independent of others given the class label.

2. Step 2: Relate assumption to text classification

   This means the presence or absence of one word does not affect another word's probability in the same document for classification.

3. Final Answer:

   Words in a document are independent of each other given the class label -> Option B

4. Quick Check:

   Naive Bayes = word independence assumption [OK]

Hint: Naive Bayes treats words as independent features [OK]

Common Mistakes:

• Thinking word order matters
• Assuming word frequency is ignored
• Believing documents must be same length

2. Which of the following is the correct way to calculate the probability of a document belonging to a class using Naive Bayes?

easy

A. P(class) / \sum_{word} P(word|class)

B. P(class) + \sum_{word} P(word|class)

C. P(class) * \prod_{word} P(word|class)

D. P(class) - \prod_{word} P(word|class)

Solution

1. Step 1: Recall Naive Bayes formula for text

   The probability of a class given a document is proportional to the prior probability of the class times the product of the conditional probabilities of each word given the class.

2. Step 2: Match formula to options

   P(class) * \prod_{word} P(word|class) correctly shows multiplication (product) of P(word|class) terms with P(class).

3. Final Answer:

   P(class) * \prod_{word} P(word|class) -> Option C

4. Quick Check:

   Naive Bayes uses product of word probabilities [OK]

Hint: Multiply class prior by product of word likelihoods [OK]

Common Mistakes:

• Adding probabilities instead of multiplying
• Dividing probabilities incorrectly
• Subtracting probabilities

3. Given the following code snippet using sklearn's MultinomialNB for text classification, what will be the predicted class for the input text ['love this movie']?

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB

texts = ['I love this movie', 'I hate this movie']
labels = ['positive', 'negative']

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

model = MultinomialNB()
model.fit(X, labels)

new_text = vectorizer.transform(['love this movie'])
prediction = model.predict(new_text)
print(prediction[0])

medium

A. movie

B. negative

C. hate

D. positive

Solution

1. Step 1: Understand training data and labels

   The model is trained on two texts: one labeled 'positive' and one 'negative'. The words 'love' and 'hate' are key indicators.

2. Step 2: Analyze prediction input

   The input text 'love this movie' contains the word 'love' which appeared in the positive example, so the model predicts 'positive'.

3. Final Answer:

   positive -> Option D

4. Quick Check:

   Word 'love' matches positive class [OK]

Hint: Check which class words in input appeared during training [OK]

Common Mistakes:

• Confusing label names with words
• Ignoring vectorizer transformation
• Predicting word instead of class

4. Consider this code snippet using Naive Bayes for text classification:

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB

texts = ['spam spam spam', 'ham ham ham']
labels = ['spam', 'ham']

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

model = MultinomialNB()
model.fit(X, labels)

new_text = vectorizer.transform(['spam ham spam'])
prediction = model.predict(new_text)
print(prediction[0])

The output is unexpected. What is the likely cause?

medium

A. The input text contains words from both classes causing confusion

B. The vectorizer did not fit on the training data

C. MultinomialNB requires numeric labels, not strings

D. The model cannot handle words not seen in training

Solution

1. Step 1: Analyze training and input data

   The training data has clear spam and ham texts. The input text mixes words from both classes.

2. Step 2: Understand Naive Bayes behavior with mixed words

   Naive Bayes calculates probabilities for each class. Mixed words can cause the model to be uncertain or pick the class with higher prior or likelihood.

3. Final Answer:

   The input text contains words from both classes causing confusion -> Option A

4. Quick Check:

   Mixed class words confuse Naive Bayes prediction [OK]

Hint: Mixed class words can confuse Naive Bayes predictions [OK]

Common Mistakes:

• Assuming unseen words cause error
• Thinking vectorizer was not fitted
• Believing labels must be numeric

5. You want to improve a Naive Bayes text classifier that often misclassifies short texts with rare words. Which approach is best to reduce this problem?

hard

A. Use Laplace smoothing to handle rare or unseen words

B. Remove all stop words from the training data

C. Increase the number of classes to make classification finer

D. Use raw word counts without normalization

Solution

1. Step 1: Identify problem with rare words

   Rare or unseen words can cause zero probabilities, making Naive Bayes assign zero probability to classes incorrectly.

2. Step 2: Apply Laplace smoothing

   Laplace smoothing adds a small count to all words, preventing zero probabilities and improving classification on rare words.

3. Final Answer:

   Use Laplace smoothing to handle rare or unseen words -> Option A

4. Quick Check:

   Laplace smoothing fixes zero probability issues [OK]

Hint: Add smoothing to avoid zero probabilities for rare words [OK]

Common Mistakes:

• Thinking removing stop words fixes rare word issue
• Believing more classes always improve accuracy
• Ignoring smoothing effects on probabilities