📧 Email Spam Detection

Author: Giuseppe Pio La Tosa
Project Date: January 9, 2025

---

🎯 Objective

This project aims to develop an intelligent email classification system that distinguishes between Spam and Ham (non-spam) emails. Leveraging Natural Language Processing (NLP) and various Machine Learning (ML) techniques, the system analyzes the textual content of emails to detect unwanted or malicious messages.

---

📦 Dataset

• Source: [TREC 2006 Public Spam Corpus]
• Total Emails: 37,822
  • Spam: 24,802 (65.9%)
  • Ham: 12,910 (34.1%)

Features:

• label: Integer – 1 for spam, 0 for ham
• origin: String – raw email content

---

🧼 Data Preprocessing

Comprehensive preprocessing was applied to clean and standardize the text data:

• Renamed origin to emails for clarity
• Extracted email body using MIME parsing
• Created label_text: maps 0 to “Ham” and 1 to “Spam”
• Removed:
  • Punctuation and excessive whitespace
  • Stop words
  • Short and non-alphanumeric tokens
  • Duplicate or conflicting entries
  • Emails with empty or malformed content
• Converted all text to lowercase and added a clean text field

---

📊 Exploratory Data Analysis & Visualization

Key insights derived from EDA:

• WordClouds and Bar Charts illustrated the most frequent terms in spam/ham messages
• Spam emails: short, promotional, with more special characters
• Ham emails: longer, more structured and formal
• Distributions significantly changed after cleaning, revealing that raw data contained substantial noise

---

⚙️ Feature Engineering

Two types of text vectorization methods were used:

• TF-IDF (Term Frequency-Inverse Document Frequency)
  • Sparse representation based on word frequency
  • Ignores semantic context
• Word2Vec
  • Dense embeddings capturing word context and similarity
  • Trained using the Skip-Gram model

---

🤖 Models Evaluated

Implemented and compared the following classification models:

• Support Vector Machine (SVC)
• Random Forest
• Logistic Regression
• Decision Tree

Evaluation Metrics:

• Accuracy
• Precision
• Recall
• F1-Score
• Confusion Matrix
• ROC Curve & AUC

Tested with both TF-IDF and Word2Vec feature representations using hold-out (70/30) and 10-fold cross-validation.

---

🔁 10-Fold Stratified Cross-Validation

Used Stratified K-Fold (k=10) to maintain class balance during evaluation.

Model	Accuracy	Precision	Recall	F1 Score	AUC
SVC	0.9857	0.9761	0.9723	0.9742	0.9949
Random Forest	0.9682	0.9792	0.9050	0.9406	0.9940
Logistic Reg.	0.9798	0.9730	0.9538	0.9633	0.9947
Decision Tree	0.9470	0.8992	0.9123	0.9057	0.9219

➡️ Best model: SVC – high F1-score and lowest error rate

---

📈 Statistical Model Comparison – Wilcoxon Signed-Rank Test

Performed pairwise comparisons using the Wilcoxon test on F1-scores:

• Null hypothesis rejected (p-value < 0.05) → performance differences are statistically significant
• SVC showed the lowest error rate, confirming its superiority

---

🌐 Real-Time Spam Classifier Web App

Built a simple web interface for real-time predictions using Flask:

• 📌 Frontend: Form to input email text
• 🤖 Backend:
  • Loads pretrained SVC model (joblib) and TfidfVectorizer
  • Endpoint /classify receives input, vectorizes it, and predicts spam/ham
• 🟢 Instant classification results returned to the user

---

🚀 Future Work & Improvements

1. Hybrid Feature Extraction: Combine TF-IDF + Word2Vec
2. Deep Learning Models: Use Transformers like BERT
3. Data Augmentation: Expand dataset with synthetic samples
4. Imbalanced Learning: Implement SMOTE or class-weight tuning

---

🧠 Technologies Used

• Python 3.11
• Pandas, NumPy, Matplotlib, Seaborn
• Scikit-learn
• Gensim (Word2Vec)
• NLTK / SpaCy
• Flask
• Joblib

---

📚 References

1. Classification of Phishing Emails with Word Embedding
2. Spam Detection with Deep Learning – Elsevier
3. TF-IDF, Word2Vec, BERT – Comparative Study
4. Machine Learning vs Deep Learning for Spam

Note: Although some references focus on phishing, the methodologies are largely transferable to spam detection.

---

🙏 Acknowledgements

Thank you for your attention!
If you found this project interesting or useful, feel free to ⭐ the repository or open an issue for discussion or collaboration.