CS 662 Assignment 6: Naive Bayes Spam Classificaion

Due: 4/22/2013

Naive Bayes Spam Classification

In this problem, you will implement a Naive Bayes classifier in Python that can distinguish between spam and non-spam, or "ham".

Your program should be able to train on a set of spam and a set of "ham." This training should include counting the frequency of each token in both spam and ham corpora.

Your program should then be able to classify an unseen email as either spam or ham by computing the MAP hypothesis:

P(spam | t1, t2, ...,tn) = alpha * P(t1,t2,tn | spam)P(spam)

Which we'll estimate as: alpha * P(t1 | spam)P(t2 | spam) ... P(spam)

You may build your program however you like, although we will still expect you to use the good programming and design practices you have learned throughout the semester. The only requirement is that we be able to run your program exactly like this:

python ./nb.py --hamtrain=dir1 --spamtrain=dir2 --hamtest=dir3 --spamtest=dir4

Where dir1-4 are directories containing ham and spam emails used for training and testing.

Your program must print out its results in the following format:

Size of ham training set: 500 emails
Size of spam training set: 700 emails:
Percentage of ham classified correctly: 98.2
Percentage of spam classified correctly: 97.0
Total accuracy: 97.5
False Positives: 1.8

There are several public repositories of email that can be used for training and testing. We will be using the SpamAssassin public corpus to evaluate your classifier.

The SpamAssassin corpus contains spam, easy ham, and hard ham. We'll use the hard ham to test your classifier.

Details

You will find that there are a lot of decisions and tweaks you can make to influence the performance of your classifier. For example, should you look at all terms, or just English words? Should you treat headers differently? When classifying an email, should you use all of its words, or just the most significant? What are reasonable priors for spam and non-spam? What about trying to parse the email and only use some chunks? These decisions are up to you: you are encouraged to experiment as much as possible.

Naive Bayes is a popular approach to spam filtering; you will find lots of resources on the Web, including:

A Plan for Spam by Paul Graham.
Spambayes - an open source spam filtering tool.
Machine Learning for Spam Detection resources - Lots of good stuff here.

There are many other resources as well. You are welcome to use the ideas (but NOT the code) from any outside source, with the following caveat:

You MUST give appropriate credit to any ideas you discover elsewhere. For example, if you read Graham's article and notice that he only uses the 15 most significant words in classifying an unseen email and decide to take this approach, you should indicate in your report (see below) that this idea is from Graham's article. Include author and URL whenever possible. Students who use other people's ideas without proper attribution will receive an automatic zero.

Grading

This part of your assignment will be graded as follows:

60 points: correctness, completeness, and style. The usual sorts of things.
30 points: performance. To compute your score on this, we will run each student's classifier on a dataset of our choosing and compute the following metric: accuracy - percentage of false positives. We will then score your performance as follows:
- More than 1 standard devation above class average: 30 points. (or above average, if >1 stdev is not possible)
- Within one standard deviation of class average: 28 points.
- More than 1 standard deviation below class average: 20 points.
10 points: Evaluation. You should check into subversion a short (approx. 1 page) document that describes the details of your approach and the accuracy of your classifier.