Spam detection pipeline in NLP - Model Metrics & Evaluation

Metrics & Evaluation - Spam detection pipeline

Which metric matters for Spam Detection and WHY

In spam detection, precision and recall are the most important metrics.

Precision tells us how many emails marked as spam really are spam. High precision means fewer good emails wrongly blocked.

Recall tells us how many actual spam emails we catch. High recall means fewer spam emails slip through.

We want a balance, but often prioritize high precision to avoid blocking important emails by mistake.

Confusion Matrix Example

      | Predicted Spam | Predicted Not Spam |
      |----------------|--------------------|
      | True Positives (TP) = 80  | False Negatives (FN) = 20 |
      | False Positives (FP) = 10 | True Negatives (TN) = 890 |
    

Total emails = 80 + 20 + 10 + 890 = 1000

Precision = TP / (TP + FP) = 80 / (80 + 10) = 0.89

Recall = TP / (TP + FN) = 80 / (80 + 20) = 0.80

Precision vs Recall Tradeoff with Examples

If we set the spam filter very strict, we catch almost all spam (high recall), but may block many good emails (low precision).

If we set it very loose, we block fewer good emails (high precision), but more spam gets through (low recall).

Example:

• High precision, low recall: Only mark very obvious spam. Few false alarms, but some spam slips through.
• High recall, low precision: Mark many emails as spam. Catch almost all spam, but many good emails get blocked.

Spam filters usually aim for high precision to avoid annoying users by blocking good emails.

Good vs Bad Metric Values for Spam Detection

Good:

• Precision above 0.85 (most flagged emails are really spam)
• Recall above 0.75 (catch most spam emails)
• Accuracy high but not the main focus

Bad:

• Precision below 0.5 (many good emails wrongly blocked)
• Recall below 0.5 (many spam emails missed)
• High accuracy but low precision or recall (accuracy paradox)

Common Metric Pitfalls in Spam Detection

• Accuracy paradox: Because most emails are not spam, a model that always predicts "not spam" can have high accuracy but is useless.
• Data leakage: If spam keywords appear in test data from training, metrics look better but model won't generalize.
• Overfitting: Model performs well on training but poorly on new emails, causing misleading metrics.
• Ignoring class imbalance: Spam is usually a small part of emails, so metrics like accuracy can be misleading.

Self Check

Your spam detection model has 98% accuracy but only 12% recall on spam emails. Is it good for production?

Answer: No, it is not good. The model misses 88% of spam emails (low recall), so many spam messages will reach users. High accuracy is misleading because most emails are not spam. Improving recall is critical.

Key Result

Precision and recall are key; high precision avoids blocking good emails, high recall catches most spam.

Practice

1. What is the main purpose of a spam detection pipeline in NLP?

easy

A. To convert text messages into numbers and train a model to identify spam

B. To translate messages into different languages

C. To summarize long emails automatically

D. To generate new text messages based on spam examples

Solution

1. Step 1: Understand the role of a spam detection pipeline

   A spam detection pipeline processes text data to prepare it for a machine learning model that can classify messages as spam or not spam.

2. Step 2: Identify the key function

   The pipeline converts text into numbers (features) and trains a model to spot spam messages automatically.

3. Final Answer:

   To convert text messages into numbers and train a model to identify spam -> Option A

4. Quick Check:

   Spam detection pipeline = convert text + train model [OK]

Hint: Spam detection means turning text into numbers to train a model [OK]

Common Mistakes:

• Thinking it translates or summarizes text
• Confusing spam detection with text generation
• Ignoring the conversion of text to numbers

2. Which of the following code snippets correctly creates a simple spam detection pipeline using scikit-learn's Pipeline with a TfidfVectorizer and a LogisticRegression model?

easy

A. Pipeline([('vectorizer', TfidfVectorizer()), ('model', LogisticRegression())])

B. Pipeline(('vectorizer', TfidfVectorizer()), ('model', LogisticRegression()))

C. Pipeline({'vectorizer': TfidfVectorizer(), 'model': LogisticRegression()})

D. Pipeline(['vectorizer' = TfidfVectorizer(), 'model' = LogisticRegression()])

Solution

1. Step 1: Recall the correct syntax for scikit-learn Pipeline

   The Pipeline constructor expects a list of tuples, each tuple containing a name and a transformer or estimator.

2. Step 2: Check each option's syntax

   Pipeline([('vectorizer', TfidfVectorizer()), ('model', LogisticRegression())]) uses a list of tuples correctly. Other options use incorrect syntax like using '=' inside lists, passing tuples as separate arguments, or dictionary syntax.

3. Final Answer:

   Pipeline([('vectorizer', TfidfVectorizer()), ('model', LogisticRegression())]) -> Option A

4. Quick Check:

   Pipeline syntax = list of (name, step) tuples [OK]

Hint: Pipeline needs a list of (name, step) tuples inside brackets [OK]

Common Mistakes:

• Using parentheses instead of brackets for the list
• Using dictionary syntax inside Pipeline
• Assigning steps with '=' inside a list

3. Given the following code, what will be the output of print(predictions) if the input messages are ["Win a free prize now", "Meeting at noon"] and the model predicts 1 for spam and 0 for not spam?

from sklearn.pipeline import Pipeline
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression

pipeline = Pipeline([
    ('vectorizer', TfidfVectorizer()),
    ('model', LogisticRegression())
])

# Assume pipeline is already trained
messages = ["Win a free prize now", "Meeting at noon"]
predictions = pipeline.predict(messages)
print(predictions)

medium

A. [0 1]

B. [1 0]

C. [1 1]

D. [0 0]

Solution

1. Step 1: Understand the input and model output

   The input has one spam-like message "Win a free prize now" and one normal message "Meeting at noon". The model labels spam as 1 and not spam as 0.

2. Step 2: Predict expected labels

   The first message is likely spam, so prediction is 1. The second is normal, so prediction is 0.

3. Final Answer:

   [1 0] -> Option B

4. Quick Check:

   Spam message = 1, normal message = 0 [OK]

Hint: Spam message predicts 1, normal message predicts 0 [OK]

Common Mistakes:

• Swapping labels 0 and 1
• Assuming both messages are spam
• Confusing output format with list of strings

4. Identify the error in this spam detection pipeline code and choose the correct fix:

from sklearn.pipeline import Pipeline
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression

pipeline = Pipeline([
    ('vectorizer', CountVectorizer),
    ('model', LogisticRegression())
])

pipeline.fit(train_messages, train_labels)

medium

A. Add parentheses to pipeline.fit() call

B. Replace LogisticRegression() with LogisticRegression

C. Remove the pipeline and train model directly

D. Change CountVectorizer to CountVectorizer() to create an instance

Solution

1. Step 1: Check the pipeline steps for correct instantiation

   CountVectorizer is a class and must be instantiated with parentheses to create an object.

2. Step 2: Identify the error and fix

   The code uses CountVectorizer without parentheses, causing an error. Adding parentheses fixes it.

3. Final Answer:

   Change CountVectorizer to CountVectorizer() to create an instance -> Option D

4. Quick Check:

   Instantiate classes with () in pipeline steps [OK]

Hint: Always instantiate transformers with () in pipeline steps [OK]

Common Mistakes:

• Forgetting parentheses after class names
• Confusing model and vectorizer instantiation
• Trying to remove pipeline instead of fixing syntax

5. You want to improve your spam detection pipeline by adding a step to remove common stop words before vectorizing. Which pipeline modification correctly adds this step using CountVectorizer with stop words removal?

hard

A. Pipeline([('stopwords', StopWordsRemover()), ('vectorizer', CountVectorizer()), ('model', LogisticRegression())])

B. Pipeline([('vectorizer', CountVectorizer()), ('stopwords', StopWordsRemover()), ('model', LogisticRegression())])

C. Pipeline([('vectorizer', CountVectorizer(stop_words='english')), ('model', LogisticRegression())])

D. Pipeline([('vectorizer', CountVectorizer(stop_words=None)), ('model', LogisticRegression())])

Solution

1. Step 1: Understand how to remove stop words in CountVectorizer

   CountVectorizer has a parameter stop_words which can be set to 'english' to remove common English stop words automatically.

2. Step 2: Check pipeline options for correct usage

   Pipeline([('vectorizer', CountVectorizer(stop_words='english')), ('model', LogisticRegression())]) correctly sets stop_words='english' inside CountVectorizer. Other options either use a non-existent StopWordsRemover step or set stop_words=None, which disables removal.

3. Final Answer:

   Pipeline([('vectorizer', CountVectorizer(stop_words='english')), ('model', LogisticRegression())]) -> Option C

4. Quick Check:

   Use stop_words='english' in CountVectorizer to remove stop words [OK]

Hint: Use stop_words='english' inside CountVectorizer [OK]

Common Mistakes:

• Trying to add a separate stop words remover step
• Setting stop_words to None disables removal
• Misplacing stop words removal after vectorizing