Email Spam Classifier

Classify message text as spam or ham (not spam) using TF-IDF features and either Multinomial Naive Bayes or a linear-kernel SVM. A Streamlit interface lets you enter text and choose the classifier.

Pipeline

1. Lowercase text and remove punctuation and non-alphabetic characters.
2. Remove English stopwords and lemmatize words using NLTK.
3. Split messages into 80% training and 20% test data with random_state=42.
4. Fit the TF-IDF vectorizer on training messages and transform the test messages.
5. Apply random undersampling to the training features only.
6. Train both classifiers, print their test metrics, and export the model artifacts.

The application operates on message text; it does not analyze email headers or attachments.

Run locally

git clone https://github.com/mtaimoor1407/Email-Spam-Classifier.git
cd Email-Spam-Classifier
python -m venv .venv

Activate with .venv\Scripts\activate on Windows or source .venv/bin/activate on macOS/Linux.

pip install -r requirements.txt

The preprocessing module downloads the NLTK stopwords, wordnet, and omw-1.4 resources. Internet access is required if they are not already installed.

Train and evaluate

The training script expects data/email.csv, read with Latin-1 encoding, with these columns:

Column	Expected value
Category	ham or spam
Message	Message text

Run from the repository root so imports and relative paths resolve correctly:

python -m src.classifier

This prints accuracy and a classification report for both models, then creates:

• models/nb_model.pkl
• models/svm_model.pkl
• models/tfidf_vectorizer.pkl

Start the interface

streamlit run app.py

Choose Naive Bayes or SVM in the sidebar, enter a message, and select Classify. If the saved models are missing, run the training command first.

Evaluation

The training script reports held-out accuracy, precision, recall, and F1 through scikit-learn's classification report. Numerical results are not added here without a preserved run output.

For a stronger comparison, preserve the dataset version and class counts, check duplicates across splits, use a stratified split, and report spam-class precision/recall/F1 alongside a confusion matrix. A false positive marks a legitimate message as spam; a false negative lets spam through.

Repository guide

Path	Purpose
app.py	Streamlit interface and model selection
src/preprocess.py	Text cleaning, stopword removal, lemmatization
src/classifier.py	Training, evaluation, undersampling, model export
data/	Input and cleaned CSV files
models/	Saved classifiers and TF-IDF vectorizer
Email Spam Classifier Report.docx	Project report

Limitations

English stopwords and the alphabetic-only cleaning rule limit the system's usefulness for multilingual text, code, URLs, and other meaningful symbols. Performance can change when message sources differ from the training data. This is a learning project rather than a complete email-filtering service.

Tools: Python, pandas, scikit-learn, imbalanced-learn, NLTK, joblib, Streamlit.