Spam Classifier — NLP Project

A machine learning project that classifies SMS messages as spam or ham using text preprocessing and Logistic Regression.

Overview

This project was built as part of learning Natural Language Processing. It covers the full pipeline from raw text to a trained classifier — including tokenization, lemmatization, TF-IDF feature extraction, and model evaluation.

Dataset

• File: spam.csv
• Size: 5575 messages
• Source: SMS Spam Collection Dataset (UCI ML Repository)
• Class distribution: ~87% ham, ~13% spam
• Encoding: latin-1

---

Project Structure

natural_language_processing/
├── dataset/
│   └── spam.csv
├── spam_classifier_model/
│   └── spam_classifier.ipynb
└── README.md

---

Pipeline

1. Data Loading and Cleaning

• Loaded CSV with pandas using latin-1 encoding
• Dropped NaN rows from message column
• Filtered only ham and spam labeled rows
• Converted all messages to string to avoid type errors

2. Text Preprocessing

Each message goes through the following steps:

• Lowercasing — convert all text to lowercase
• URL replacement — replace links with the token url
• Number replacement — replace digits with the token num
• Punctuation removal — strip all special characters
• Tokenization — split sentence into individual words using NLTK
• Stopword removal — remove common words like "the", "is", "a"
• Lemmatization — reduce words to base form, "prizes" becomes "prize"

3. Feature Engineering

• Used TfidfVectorizer from scikit-learn
• Top 5000 features selected
• sublinear_tf=True to handle repeated words in spam messages
• Vectorizer fitted only on training data to avoid data leakage

4. Model Training and Testing

• Algorithm used: Logistic Regression
• 80/20 train test split with stratify=y to preserve class ratio
• Evaluated using precision, recall, F1 score and confusion matrix

---

Libraries Used

pandas
numpy
scikit-learn
nltk
matplotlib
seaborn

Install all at once:

pip install pandas numpy scikit-learn nltk matplotlib seaborn

---

NLTK Downloads

Run this once before starting:

import nltk
nltk.download('punkt')
nltk.download('punkt_tab')
nltk.download('wordnet')
nltk.download('stopwords')
nltk.download('averaged_perceptron_tagger')

How to Run

Step 1 — Install dependencies

pip install -r requirements.txt

Step 2 — Open the notebook

jupyter notebook spam_classifier.ipynb

Step 3 — Run all cells from top to bottom

Results

              precision    recall    f1-score    support
ham              0.99       0.99       0.99        966
spam             0.95       0.96       0.96        149
accuracy                               0.98       1115
macro avg        0.97       0.97       0.97       1115

---

Key Learnings

• Data leakage is the most common mistake — always fit the vectorizer on training data only, never on the full dataset
• Accuracy is misleading on imbalanced data — F1 score is the right metric here since spam is only 13% of the data
• TF-IDF beats raw word counts — down-weighting common words improves classification noticeably
• Lemmatization reduces sparsity — fewer unique tokens means a cleaner feature space

---

What Can Be Improved

• Add bigrams using ngram_range=(1,2) for phrase-level signals like "free entry" and "click here"
• Try LinearSVC which often outperforms Logistic Regression on text
• Use cross-validation for more reliable evaluation
• Add POS-aware lemmatization for better verb handling

---

Author

Built as part of NLP learning journey covering tokenization, lemmatization, Bag of Words, TF-IDF and text classification.