Skip to content

lorenanicole/python-naive-bayes-spam-classifier

Repository files navigation

Basic Naive Bayes Classifier in Python

This approach makes use of pre-labeled data provided by the Kaggle Classroom spam detection challenge.

naive-bayes Python 2 Classifier

Python project code in naive-bayes is written with Python 2.7.

For setup create a virtualenv with the requirements:

virtualenv nbenv
source nbenv/bin/activate
pip install -r pathway/to/naive-baves/requirements.txt

To run the Naive Bayes classifier:

cd naive-bayes
python spam_detector.py

Python 3 Jupyter Notebook

The Python 2.7 project has been ported to Python 3 and can be run in the Jupyter notebook.

First you will want to create a Python3 virtualenv:

pyenv-3.5 python3env  # Update 3.5 with your version of Python 3
source python3env/bin/activate  # Name your env whatever you like!
pip3 install -r requirements.txt 

Then start the notebook!

jupyter notebook

Notes on Python Naive Bayes Implementation

You can have the detector either train and evaluate itself against the training data (using 90% of the pre-labeled data as training and 10% to label) with:

detector.train_and_evaluate()

Or you can train against the entire labeled data set (2500 emails) and classify on the unlabeled data (1827 emails).

detector.train()
detector.classify(1827)  # Number of emails to classify

Ham has a label of 1 while Spam has a label of 0.

How Naive Bayes Implemented

This solution makes use of Python's 2.7 Decimal module, which is used for floating point arithmetic. (Prevents floating point underflow!)

Inside the NaiveBayes#train method each document has common stop words removed using [NLTK](install http://www.nltk.org/install.html). The words have not yet been stemmed as this is a forthcoming feature.

Only the corpus of words are used as selectors to determine if an email is spam or ham.

To prevent words with 0 frequency from miscontruing the results, Laplace smoothing is applied to increment each 0 frequency word to 1.

Releases

No releases published

Packages

No packages published