Multinomial Naïve Bayes Classifier - the YouTube Spam Collection Data Set

Project Objective

Welcome, here I created a YouTube Spam Comments Classifier with the Naïve Bayes Classifier. I created this project in order to:

• Practice with the Naïve Bayes Classifier, that's why I only use this method.
• Practice Text Classification
• Share with fellow learners as myself how to work with it
• Go through a simple DataScience project

Resources used:

• Python
• Jupyter Notebooks

Packages:

• numpy
• seaborn
• matplotlib
• scikit-learn
• pandas

Anaconda env creation conda env create -f environment.yaml

Data

The data was downloaded from the kaggle Data Set YouTube "Spam Collection Data Set".

The data are 5 different Datasets containing comments for 5 different Youtube Videos, and are already classified as:

• 1: Spam Comment
• 2: Ham Comment

EDA

The data was pretty clean and did not contain Null Values, so the only left thing to do was to extract the columns we needed: The Comment and the Class.

Next, the data was divided into train and test sets

Data Transformation

The data was passed through a CountVectorizer sklearn's class which turned each comment into a vector, in which each position represent how many times a specific word appeard in the comment.

The lenght of the vector ended beeing of 3814 different words (Words longer than 1 character).

Model Building

For the classifier I created 3 Naïve Bayes Models:

• The Default MultinomialNB
• The MultinomialNB with the priors adjusted on [0.6, 0.4] for Ham and Spam Comments respectively.
• The MultinomialNB with the priors adjusted on [0.7, 0.3] for Ham and Spam Comments respectively
• The ComplementNB with the priors adjusted on [0.7, 0.3] for Ham and Spam Comments respectively

Metrics

I chose to use the Confussion Matrix, focusing on Precision and Recall Metrics to Evaluate how the models changed their performances. I chose these Metrics in order to see how the model correctly classifies the Spam class, and because they are easy to understand.

Model Performance

Model Performance for Spam Classification

Model	Precision	Recall
Default MultinomialNB	89%	95%
MultinomialNB with [0.6, 0.4] priors	96%	95%
MultinomialNB with [0.7, 0.3] priors	97%	94%
ComplementNB with [0.7, 0.3] priors	93%	95%