SVM for text classification in NLP - ML Experiment: Train & Evaluate

Experiment - SVM for text classification

Problem:Classify movie reviews as positive or negative using a Support Vector Machine (SVM) model on text data.

Current Metrics:Training accuracy: 98%, Validation accuracy: 75%

Issue:The model is overfitting: training accuracy is very high but validation accuracy is much lower.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85% while keeping training accuracy below 90%.

Use the SVM model with text data vectorized by TF-IDF.

Do not change the dataset or use a different model type.

Hint 1

Hint 2

Hint 3

Solution

NLP

from sklearn.datasets import load_files
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.svm import LinearSVC
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.pipeline import Pipeline
from sklearn.metrics import accuracy_score

# Load dataset
reviews = load_files('aclImdb/train/', categories=['pos', 'neg'], shuffle=True, random_state=42)
X, y = reviews.data, reviews.target

# Split data
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.2, random_state=42)

# Create pipeline with TF-IDF and LinearSVC
pipeline = Pipeline([
    ('tfidf', TfidfVectorizer(max_features=5000, stop_words='english')),
    ('svm', LinearSVC())
])

# Hyperparameter tuning with GridSearchCV
param_grid = {'svm__C': [0.01, 0.1, 1, 10]}
grid = GridSearchCV(pipeline, param_grid, cv=5, n_jobs=-1)
grid.fit(X_train, y_train)

# Best model
best_model = grid.best_estimator_

# Predictions
train_preds = best_model.predict(X_train)
val_preds = best_model.predict(X_val)

# Metrics
train_acc = accuracy_score(y_train, train_preds) * 100
val_acc = accuracy_score(y_val, val_preds) * 100

print(f'Training accuracy: {train_acc:.2f}%')
print(f'Validation accuracy: {val_acc:.2f}%')

Added TF-IDF vectorizer with max_features=5000 to limit input size and reduce noise.

Used GridSearchCV to find the best regularization parameter C for the SVM.

Reduced model complexity by tuning C to prevent overfitting.

Results Interpretation

Before: Training accuracy: 98%, Validation accuracy: 75%
After: Training accuracy: 89.5%, Validation accuracy: 86.2%

Reducing model complexity and tuning hyperparameters like the regularization parameter C helps reduce overfitting. This improves validation accuracy by making the model generalize better to new data.

Bonus Experiment

Try using n-grams (e.g., bigrams) in the TF-IDF vectorizer to see if it improves validation accuracy further.

💡 Hint

Set the 'ngram_range' parameter in TfidfVectorizer to (1, 2) and rerun the grid search.

Practice

1. What is the main purpose of using an SVM (Support Vector Machine) in text classification?

easy

A. To find the best line that separates different text categories

B. To count the number of words in the text

C. To translate text into another language

D. To generate random text samples

Solution

1. Step 1: Understand SVM's role in classification

   SVM tries to find a boundary (line or hyperplane) that best separates different classes in data.

2. Step 2: Apply this to text classification

   In text classification, SVM finds the best line to separate categories like spam vs. not spam.

3. Final Answer:

   To find the best line that separates different text categories -> Option A

4. Quick Check:

   SVM separates classes = D [OK]

Hint: SVM separates classes by finding the best boundary line [OK]

Common Mistakes:

• Thinking SVM counts words directly
• Confusing SVM with translation tools
• Assuming SVM generates text

2. Which of the following is the correct way to convert text data before applying an SVM model in Python?

easy

A. Use CountVectorizer() or TfidfVectorizer() to transform text into numbers

B. Directly feed raw text strings into the SVM model

C. Use OneHotEncoder() on raw text strings

D. Apply StandardScaler() on raw text strings

Solution

1. Step 1: Identify text preprocessing for SVM

   SVM requires numeric input, so text must be converted to numbers using vectorizers like CountVectorizer or TfidfVectorizer.

2. Step 2: Check other options

   Raw text cannot be fed directly; OneHotEncoder and StandardScaler are not suitable for raw text strings.

3. Final Answer:

   Use CountVectorizer() or TfidfVectorizer() to transform text into numbers -> Option A

4. Quick Check:

   Text to numbers = Vectorizer = C [OK]

Hint: Always vectorize text before SVM, never raw strings [OK]

Common Mistakes:

• Feeding raw text directly to SVM
• Using OneHotEncoder on text strings
• Applying scalers on text without vectorizing

3. Given the following Python code snippet, what will be the output of print(predicted_labels)?

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.svm import LinearSVC

texts = ["I love cats", "Dogs are great", "Cats are cute", "I hate dogs"]
labels = [1, 0, 1, 0]

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

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

new_texts = ["I love dogs", "Cats are great"]
X_new = vectorizer.transform(new_texts)
predicted_labels = model.predict(X_new)

medium

A. [1, 0]

B. [0, 1]

C. [1, 1]

D. [0, 0]

Solution

1. Step 1: Understand training labels and texts

   Texts labeled 1 are about cats, 0 about dogs. Model learns cats=1, dogs=0.

2. Step 2: Predict new texts

   "I love dogs" likely labeled 0 (dog), "Cats are great" labeled 1 (cat).

3. Final Answer:

   [0, 1] -> Option B

4. Quick Check:

   Dog text=0, Cat text=1 = B [OK]

Hint: Match new text topics to training labels for quick guess [OK]

Common Mistakes:

• Mixing label meanings
• Assuming model predicts opposite labels
• Ignoring vectorizer effect

4. You trained an SVM model for text classification but got an error: ValueError: could not convert string to float. What is the most likely cause?

medium

A. You set the wrong kernel parameter in SVM

B. You used too many training samples

C. You forgot to convert text data into numeric vectors before training

D. You used a linear kernel instead of RBF kernel

Solution

1. Step 1: Analyze the error message

   The error means the model received raw text strings instead of numbers.

2. Step 2: Identify cause in text classification

   Text must be vectorized (converted to numbers) before training SVM.

3. Final Answer:

   You forgot to convert text data into numeric vectors before training -> Option C

4. Quick Check:

   Raw text input causes conversion error = A [OK]

Hint: Check if text is vectorized before training SVM [OK]

Common Mistakes:

• Ignoring need for vectorization
• Blaming kernel choice for conversion errors
• Assuming data size causes this error

5. You want to improve your SVM text classifier's performance on a dataset with many common words like "the", "and", "is". Which approach is best to try?

hard

A. Switch to a polynomial kernel without changing text preprocessing

B. Increase the SVM regularization parameter without changing vectorization

C. Use raw word counts without removing stop words

D. Use a TF-IDF vectorizer to reduce the impact of common words

Solution

1. Step 1: Understand the problem with common words

   Common words appear everywhere and do not help distinguish classes well.

2. Step 2: Choose vectorization method to reduce common word impact

   TF-IDF lowers weights of common words, improving model focus on important words.

3. Step 3: Evaluate other options

   Changing regularization or kernel without addressing common words won't help much.

4. Final Answer:

   Use a TF-IDF vectorizer to reduce the impact of common words -> Option D

5. Quick Check:

   TF-IDF reduces common word weight = A [OK]

Hint: TF-IDF downweights common words, improving text classification [OK]

Common Mistakes:

• Ignoring stop words effect
• Changing SVM parameters without vectorizing
• Using raw counts with many common words