Skip to content

Latest commit

 

History

History
70 lines (44 loc) · 2.55 KB

README.md

File metadata and controls

70 lines (44 loc) · 2.55 KB
Raw

beiyesi

"beiyesi" is a bayesian text classifier in python.

Installation

execute

sudo python setup.py install

Initiate a python object

>>> import beiyesi
>>> clss = beiyesi.classifier.Classifier()

Usage

Data File Format

each line is a doc, like below

docid label1,label2 word1 word2 word3

  • docid is a string that uniquely identifies a doc. Multiple docs with the same docid are ignored except the first one.
  • A doc can have one or more labels, separated by comma. ( no spaces around the comma)
  • The rest of the line is a list of words separated by spaces

Training the classifier

The classifier should be fed with a list of docs that are already labeled

>>> clss.trainLine("1 ham aa bb cc")
>>> clss.trainLine("2 spam xx yy zz")

Classify an unknown doc

>>> words = ['aa','cc']
>>> print clss.classifyDoc(words) 
[('ham', 0.017492711370262388), ('spam', 0.0043731778425655969)]

As the classification result, the probability for each label is returned in a list. In the order of highest probability first. So x[0][0] is the name of the most possible label, x[0][1] is its probability

Explain how the probability is calculated

>>> clss.explain('ham', words)
(0.017492711370262388, 0.21428571428571427, [('aa', 0.2857142857142857), ('cc', 0.2857142857142857)])

It explains that for the given doc "words", how its probablity of being label "ham" is calculated.

  • The first float 0.017492711370262388 is the final result as returned by classifyDoc
  • The second float 0.21428571428571427 is the prior probablity of label "ham"
  • The last list [('aa', 0.2857142857142857), ('cc', 0.2857142857142857)] shows that for each word in the doc, the probablity that it contributes to the result. In the order of highest probablity word first.

Math

I am going to skip explaining the bayesian algorithm here. Naive Bayes classifier in 50 lines is a good article I refered to.

Smoothing

I implemented the "add-one-smoothing" as described by the article above.

Examples

I downloaded the test data set from SIAM2007. You can apply beiyesi to it following the next steps.

  1. Enter the test directory

    cd beiyesi/test

  2. Following script converts the data set to our format. Then runs the tests.

    ./run.sh

    total=3083, wrong=2398, score=0.222186

It doesn't do well on this data set. But it's still ok for my personal projects. So your miles may very depending on the data you need to process.