R tm package used for predictive analytics. How one classifies a new document?

Question

This is a general question about the procedures concerning text mining. Suppose one has a Corpus of documents classified as Spam/No_Spam. As standard procedure one pre-process the data, removing punctuation, stops words etc. After converting it into a DocumentTermMatrix one can build some models to predict spam/No_Spam. Here is my problem. Now I want to use the model built for new documents arrive. In order to check a single document I would have to build a DocumentTerm*Vector*? so it can be used to predict Spam/No_Spam. In the documentation of tm I found one converts the full Corpus into a Matrix using for example tfidf weights. How can I then convert a single vector using the idf from the Corpus? do i have to change my corpus and build a new DocumentTermMatrix every time? I processed my corpus, converted it into a matrix and then split it into a Training and Testing sets. But here the test set was built in the same line as the document matrix of the full set. I can check precision etc, but do not know whats the best procedure for new text classification.

Ben, Imagine I have a preprocessed DocumentTextMatrix, I convert it into a data.frame.

dtm <- DocumentTermMatrix(CorpusProc,control = list(weighting =function(x) weightTfIdf(x, normalize =FALSE),stopwords = TRUE, wordLengths=c(3, Inf), bounds = list(global = c(4,Inf))))

dtmDataFrame <- as.data.frame(inspect(dtm))

Added a Factor Variable and built a model.

Corpus.svm<-svm(Risk_Category~.,data=dtmDataFrame)

Now imagine I give you a new document d (was not in your Corpus before) and you want to know the model prediction spam/No_Spam. How you do that?

Ok lets create an example based on the code used here.

examp1 <- "When discussing performance with colleagues, teaching, sending a bug report or searching for guidance on mailing lists and here on SO, a reproducible example is often asked and always helpful. What are your tips for creating an excellent example? How do you paste data structures from r in a text format? What other information should you include? Are there other tricks in addition to using dput(), dump() or structure()? When should you include library() or require() statements? Which reserved words should one avoid, in addition to c, df, data, etc? How does one make a great r reproducible example?"
examp2 <- "Sometimes the problem really isn't reproducible with a smaller piece of data, no matter how hard you try, and doesn't happen with synthetic data (although it's useful to show how you produced synthetic data sets that did not reproduce the problem, because it rules out some hypotheses). Posting the data to the web somewhere and providing a URL may be necessary. If the data can't be released to the public at large but could be shared at all, then you may be able to offer to e-mail it to interested parties (although this will cut down the number of people who will bother to work on it). I haven't actually seen this done, because people who can't release their data are sensitive about releasing it any form, but it would seem plausible that in some cases one could still post data if it were sufficiently anonymized/scrambled/corrupted slightly in some way. If you can't do either of these then you probably need to hire a consultant to solve your problem" 
examp3 <- "You are most likely to get good help with your R problem if you provide a reproducible example. A reproducible example allows someone else to recreate your problem by just copying and pasting R code. There are four things you need to include to make your example reproducible: required packages, data, code, and a description of your R environment. Packages should be loaded at the top of the script, so it's easy to see which ones the example needs. The easiest way to include data in an email is to use dput() to generate the R code to recreate it. For example, to recreate the mtcars dataset in R, I'd perform the following steps: Run dput(mtcars) in R Copy the output In my reproducible script, type mtcars <- then paste. Spend a little bit of time ensuring that your code is easy for others to read: make sure you've used spaces and your variable names are concise, but informative, use comments to indicate where your problem lies, do your best to remove everything that is not related to the problem. The shorter your code is, the easier it is to understand. Include the output of sessionInfo() as a comment. This summarises your R environment and makes it easy to check if you're using an out-of-date package. You can check you have actually made a reproducible example by starting up a fresh R session and pasting your script in. Before putting all of your code in an email, consider putting it on http://gist.github.com/. It will give your code nice syntax highlighting, and you don't have to worry about anything getting mangled by the email system."
examp4 <- "Do your homework before posting: If it is clear that you have done basic background research, you are far more likely to get an informative response. See also Further Resources further down this page. Do help.search(keyword) and apropos(keyword) with different keywords (type this at the R prompt). Do RSiteSearch(keyword) with different keywords (at the R prompt) to search R functions, contributed packages and R-Help postings. See ?RSiteSearch for further options and to restrict searches. Read the online help for relevant functions (type ?functionname, e.g., ?prod, at the R prompt) If something seems to have changed in R, look in the latest NEWS file on CRAN for information about it. Search the R-faq and the R-windows-faq if it might be relevant (http://cran.r-project.org/faqs.html) Read at least the relevant section in An Introduction to R If the function is from a package accompanying a book, e.g., the MASS package, consult the book before posting. The R Wiki has a section on finding functions and documentation"
examp5 <- "Before asking a technical question by e-mail, or in a newsgroup, or on a website chat board, do the following:  Try to find an answer by searching the archives of the forum you plan to post to. Try to find an answer by searching the Web. Try to find an answer by reading the manual. Try to find an answer by reading a FAQ. Try to find an answer by inspection or experimentation. Try to find an answer by asking a skilled friend. If you're a programmer, try to find an answer by reading the source code. When you ask your question, display the fact that you have done these things first; this will help establish that you're not being a lazy sponge and wasting people's time. Better yet, display what you have learned from doing these things. We like answering questions for people who have demonstrated they can learn from the answers. Use tactics like doing a Google search on the text of whatever error message you get (searching Google groups as well as Web pages). This might well take you straight to fix documentation or a mailing list thread answering your question. Even if it doesn't, saying “I googled on the following phrase but didn't get anything that looked promising” is a good thing to do in e-mail or news postings requesting help, if only because it records what searches won't help. It will also help to direct other people with similar problems to your thread by linking the search terms to what will hopefully be your problem and resolution thread. Take your time. Do not expect to be able to solve a complicated problem with a few seconds of Googling. Read and understand the FAQs, sit back, relax and give the problem some thought before approaching experts. Trust us, they will be able to tell from your questions how much reading and thinking you did, and will be more willing to help if you come prepared. Don't instantly fire your whole arsenal of questions just because your first search turned up no answers (or too many). Prepare your question. Think it through. Hasty-sounding questions get hasty answers, or none at all. The more you do to demonstrate that having put thought and effort into solving your problem before seeking help, the more likely you are to actually get help. Beware of asking the wrong question. If you ask one that is based on faulty assumptions, J. Random Hacker is quite likely to reply with a uselessly literal answer while thinking Stupid question..., and hoping the experience of getting what you asked for rather than what you needed will teach you a lesson."



corpus2 <- Corpus(VectorSource(c(examp1, examp2, examp3, examp4)))

Note I took out example 5

skipWords <- function(x) removeWords(x, stopwords("english"))
funcs <- list(tolower, removePunctuation, removeNumbers, stripWhitespace, skipWords)
corpus2.proc <- tm_map(corpus2, FUN = tm_reduce, tmFuns = funcs)

corpus2a.dtm <- DocumentTermMatrix(corpus2.proc, control = list(wordLengths = c(3,10)))
dtmDataFrame <- as.data.frame(inspect(corpus2a.dtm)) 

Added a factor variable Spam_Classification 2 levels spam/No_Spam

dtmFinal<-cbind(dtmDataFrame,Spam_Classification)

I build a model SVM Corpus.svm<-svm(Spam_Category~.,data=dtmFinal)

Now imagine I have example 5 as a new document (email) How I generate a Spam/No_Spam value???

Answer 1

Thanks for this interesting question. I have been thinking over it for some time. Too keep things short, the quintessence of my findings: For weighting-methods except tf there is no way around laborious work or recalculating the whole DTM (and probably rerunning your svm).

Only for tf-weighting I could find an easy process for classifying new content. You have to transform the new document (for sure) to a DTM. During the transformation you have to add a dictionary containing all the terms you have used to train your classifier on the old corpus. Then you can use predict() as usually. For the tf part, here a very minimal sample and a method for classifying a new document:

### I) Data

texts <- c("foo bar spam",
           "bar baz ham",
           "baz qux spam",
           "qux quux ham")

categories <- c("Spam", "Ham", "Spam", "Ham")

new <- "quux quuux ham"

### II) Building Model on Existing Documents „texts“

library(tm)  # text mining package for R
library(e1071)  # package with various machine-learning libraries

## creating DTM for texts
dtm <- DocumentTermMatrix(Corpus(VectorSource(texts)))

## making DTM a data.frame and adding variable categories
df <- data.frame(categories, as.data.frame(inspect(dtm)))

model <- svm(categories~., data=df)

### III) Predicting class of new

## creating dtm for new
dtm_n <- DocumentTermMatrix(Corpus(VectorSource(new)),
                            ## without this line predict won't work
                            control=list(dictionary=names(df)))
## creating data.frame for new
df_n <- as.data.frame(inspect(dtm_n))

predict(model, df_n)

## > 1 
## > Ham 
## > Levels: Ham Spam