Naive Bayes Classifiers Comparison

Explored alternatives for improved performance in ML course, uOttawa 2023. This repository contains Python code implementing a Spambase Dataset analysis comparing Naïve Bayes classifiers. Evaluated accuracy, confusion matrices on different splits in a Spambase dataset as part of a Machine Learning course project at my study in the University of Ottawa in 2023.

• Required libraries: scikit-learn, pandas, matplotlib.
• Execute cells in a Jupyter Notebook environment.
• The uploaded code has been executed and tested successfully within the Google Colab environment.

Binary-class classification problem

Task is to classify the email dataset into two classes: Spam / Not Spam.

Independent Variables:

• 57 Features related to word frequencies, character frequencies, and capital run lengths.

Target variable:

• 'Target' indicating the classification into two classes.

Key Tasks Undertaken

1. Dataset Splitting:

   • Divided the dataset into 80% training and 20% test samples, preserving the split for later analysis.

2. Classifier Evaluation (80/20 Split):

   • Computed confusion matrices and accuracy scores for Gaussian and Multinomial Naïve Bayes classifiers on test data.

     • Identified that both classifiers didn't predict any spam instances due to unbalanced test data.

3. Further Evaluation:

   • Employed train-test split function, noting dataset shuffling to avoid zero instances of 'spam' in test data.

     

4. Alternate Classifier Assessment:

   • Explored Bernoulli and Complement Naïve Bayes classifiers, comparing their performance metrics with Gaussian and Multinomial models.

     

5. Subset Evaluation:

   • Analyzed four subsets' accuracies, revealing varied performances due to biased training on specific class labels.

6. Visualization:

   • Presented subset accuracies via a bar chart, highlighting classifier performance variations.