Spam detection pipeline in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Model Choice

intermediate

Choosing the best model for spam detection

You want to build a spam detection system that classifies emails as spam or not spam. Which model is best suited for this binary text classification task?

A A linear Support Vector Machine (SVM) with TF-IDF features

B K-Means clustering with raw email text

C A Convolutional Neural Network (CNN) designed for image recognition

D Principal Component Analysis (PCA) for dimensionality reduction

💡 Hint

Think about models that work well with text features and binary classification.

SVM with TF-IDF features is a strong, simple model for text classification. K-Means is unsupervised and not for classification. CNNs for images won't work well on raw text. PCA is for reducing features, not classification.

❓ Predict Output

intermediate

Output of text preprocessing step

Given this Python code that preprocesses email text, what is the output?

NLP

import re
text = "Hello!!! This is spam??? Visit http://spam.com now."
cleaned = re.sub(r'http\S+', '', text)
cleaned = re.sub(r'[^a-zA-Z ]', '', cleaned).lower().split()
print(cleaned)

A ['hello', 'this', 'is', 'spam', 'visit', 'now']

B ['hello!!!', 'this', 'is', 'spam???', 'visit', 'http://spam.com', 'now']

C ['hello', 'this', 'is', 'spam', 'visit', 'httpspamcom', 'now']

D ['hello', 'this', 'is', 'spam', 'visit', 'http', 'spamcom', 'now']

💡 Hint

Look at how URLs and punctuation are removed, and text is lowercased and split.

The regex removes URLs first, then removes all non-letter characters, then lowercases and splits into words.

❓ Hyperparameter

advanced

Choosing the best hyperparameter for TF-IDF vectorizer

In a spam detection pipeline, you use a TF-IDF vectorizer. Which max_features value is best to balance performance and speed on a large email dataset?

A max_features=10

B max_features=1000000

C max_features=None

D max_features=10000

💡 Hint

Too few features may miss important words; too many may slow training.

10 features is too small to capture text variety. 1,000,000 is too large and slow. None uses all features, which can be very large. 10,000 is a good balance.

❓ Metrics

advanced

Interpreting spam detection model metrics

Your spam detection model has these results on test data: 90% accuracy, 70% precision, 95% recall. What does this mean?

A The model correctly identifies most spam emails but also marks many non-spam as spam.

B The model rarely misses spam emails but sometimes wrongly flags non-spam as spam.

C The model is very precise but misses many spam emails.

D The model has balanced precision and recall, so it is perfect.

💡 Hint

Recall is about catching spam; precision is about avoiding false alarms.

High recall (95%) means most spam is caught. Lower precision (70%) means some non-spam is wrongly flagged. Accuracy is high but less informative if classes are imbalanced.

🔧 Debug

expert

Debugging model training failure in spam detection

Why does this spam detection training code raise a ValueError?

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
texts = ["Free money now", "Hello friend", "Win a prize"]
labels = [1, 0]
vectorizer = TfidfVectorizer()
X = vectorizer.fit_transform(texts)
model = LogisticRegression()
model.fit(X, labels)

A fit_transform returns a dense matrix, but LogisticRegression needs sparse

B LogisticRegression requires labels to be strings, not integers

C The labels list length does not match the number of texts

D TfidfVectorizer cannot process short texts

💡 Hint

Check if the number of labels matches the number of input samples.

The labels list has 2 items but texts has 3. The model expects one label per text sample.