danielfm / judgr-spam-demo

Spam classifier demo application using Judgr, a Naive Bayes classifier library.

Judgr Spam Classification Demo

This is a sample application built on top of Judgr, a naïve Bayes classifier library written in Clojure.

Getting Started

First, clone this repository and start a REPL by running lein repl in its root directory.

Then, load the core namespace:

user=> (use 'judgr-spam-demo.core)
nil

Training The Classifier

This repository comes with a few thousand messages for training and testing the classifier. See License section below for further information.

The following command will train the classifier using the messages stored in data/training:

user=> (time (train!))
"Elapsed time: 20983.71 msecs"

This operation might take several seconds to finish.

Trying It Out

Choose a few messages from data/testing and see if the classifier got them right:

user=> (.classify classifier (slurp "data/testing/TEST_XXXXX.eml"))
:spam

If you are curious about how a specific feature are distributed between spam and ham messages:

user=> (.get-feature (.db classifier) "viagra")
{:feature "viagra", :total 42, :classes {:spam 41, :ham 1}}

Evaluating The Classifier

Use the judgr.cross-validation namespace to generate a Confusion Matrix and analyze the results.

The following example shows how to perform a 10-Fold Cross Validation:

user=> (use 'judgr.cross-validation)
nil

user=> (def conf-matrix (k-fold-crossval 10 classifier))
#'user/conf-matrix

user=> (float (accuracy conf-matrix))
0.9310345

This operation might take several minutes to finish.

License

Email Dataset

The repository includes a subset of the CSDMC2010 SPAM corpus, which is one of the datasets for the data mining competition associated with ICONIP 2010.

Copyright for the text in the messages remains with the original senders.

The complete dataset can be downloaded here.

Code

Copyright (C) Daniel Fernandes Martins

Distributed under the New BSD License. See COPYING for further details.