Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

I want to build a document classifier in R, using the Naive Bayes approach.

Here are steps, that I've done so far:

  • I have corpus with about 30 documents from 2 authors (Classes are: "target author" and "other author").
  • "Vocabulary" (training set) has been pre-processed (removed numbers, removed punctuation, words to lower case, removed stop words, stem documents, strip whitespace), and I am considering only frequent words (top 700).
  • Now I have matrix which looks like:

    enter image description here

Then I trained my classifier using Bayes using some existing R library, e1071.

Here are my questions:

I want to test my classifier on other documents that were not part of the training set.

  • How to prepare my data matrix? What if those other documents don't contain all the words (attributes) from my training set? Should I put dummy columns there (e.g., with value=0)?
  • Does the position of the words (columns order) matter?

Here is an example:

Training attributes:

"wild"  "wind"  "woman"

Testing attributes:

"woman" "wind" "wild"  

Is this ok, or should columns be in the same order as in training matrix?

share|improve this question

migrated from stackoverflow.com Jan 11 '12 at 21:22

This question came from our site for professional and enthusiast programmers.

7  
"Classifying documents with Bayes" sounds like you'd like to resurrect the poor reverend and make him help you sort a pile of papers... –  mbq Jan 13 '12 at 13:53
1  
This question is off-topic because it is about necromancy. –  Marc Claesen Apr 24 at 7:44

3 Answers 3

order of variables is not an issue.I guess you are using the actual tokens as variables then randomforest or svm or any other model can understand that using variable names .THe issue can be when you dont have certain tokens in test data you might need to introduce dummy values

share|improve this answer
    
what are tokens? –  Glen Nov 17 '13 at 19:01
    
"token" is NLP jargon for what we think of as unique words –  ssdecontrol Oct 2 at 1:31

You should construct your features (in this case, the words you're including as descriptors of each document) based only on your training set. This will calculate the probability of having a certain word given that it belongs to a particular class: $P(w_i|c_k)$. In case you're wondering, this probability is needed when calculating the probability of a document belonging to some class: $P(c_{k}|\text{document})$

When you want to predict the class for a new document in the test set, ignore the words that are not included in the training set. The reason is that you can't use the test set for anything other than testing your predictions. Furthermore, the training set must be representative of the test set. Otherwise, you won't get a good classifier. Therefore, it is to be expected that the majority of the words in the test set are also included in the training set.

Some people add an extra column for unknown words and try to calculate a probability of such words given a certain class: $P(\text{unknown} | c_{i})$. I don't think this is necessary or even appropriate because in order to obtain this probability, you need to peek at the test set. That's something you must never do.

share|improve this answer

Why did you choose top 700 words? In many cases much less would be appropriate. So if you first filter off stopwords and other meaningless frequent words, then you could try some smaller amount and check how does it work. Generally, if you use big amount of words in your set most of them are pure noise and do not give much information. Make few tries and check what rate is enough, but with predicting only two categories I imagine that you could use much smaller amount of them.

What to do with missing words? They do not occur so they have frequency of zero. On the other hand, Naive Bayes uses products heavily and if you multiply anything by zero you get zero. And in most (probably all) rows you will have some words that did not occurred, so your matrix will become a collections of zeros. Because of that it is better to choose some arbitrary small number and add it to all the values in your matrix so there is no zeros (and most ready made algorithms do this for you).

Position of the words in the matrix does not matter. However, position of the words in text could matter, so you can include such a variable in your analysis (however that is beyond scope of simply using Naive Bayes algorithm).

Final general remark: pay very much attention on cleaning and preprocessing the data since it is crucial in NLP, remember: garbage in, garbage out. Also deciding on which words to include in your training set is important step - taking "top $n$ words" could be not enough in many cases.

share|improve this answer

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.