Spam Email Detection with Gaussian Naive Bayes

This project implements a fully custom Gaussian Naive Bayes classifier to detect spam emails using the UCI Spambase dataset. It incorporates advanced preprocessing, feature selection, parameter tuning, and ensemble learning to achieve over 91% accuracy on the test set.

---

Project Highlights

• Custom Naive Bayes Implementation

  • Built from scratch using only NumPy.
  • Stable handling of zero variance using a minimum standard deviation.
  • Inference based on log-probabilities to prevent underflow.

• Advanced Preprocessing

  • Log1p transformation to reduce skew.
  • Robust IQR-based scaling.
  • Outlier clipping to ±3 IQR.

• Feature Selection

  • Uses a signal-to-noise ratio to rank features by discriminative power.
  • Enables top-k feature selection for optimized performance.

• Hyperparameter Optimization

  • Grid search over combinations of min_std, prior_smoothing, and top_k features.
  • Evaluation based on validation accuracy.

• Ensemble Learning

  • Combines 7 Naive Bayes models trained on different feature subsets.
  • Final predictions are made by majority voting.

• Evaluation

  • Achieves:
    • Accuracy: 0.9124
    • Precision: 0.8628
    • Recall: 0.9246
    • F1 Score: 0.8926
  • Includes confusion matrix visualizations.

---

File Structure

---

How to Run

1. Clone this repository.
2. Open notebook/SpamEmailDetection.ipynb in Google Colab or Jupyter.
3. Ensure the following libraries are installed:
   • numpy
   • pandas
   • scikit-learn
   • matplotlib
4. Run all cells to train and evaluate the models.

---

Dataset

UCI Spambase Dataset:
https://archive.ics.uci.edu/dataset/94/spambase

---

Course Information

This project was developed for CS 445/545: Machine Learning (Spring 2025) at Portland State University.