library(RTextTools)
library(tm)
library(SnowballC)
library(plyr)
library(stringr)
library(readr)
library(wordcloud)
library(knitr)

Assignment

Perform text processing to figure out if

Train the Algo

Read in known spam and known ham

#Load Spam and Ham into a Corpus
spam_corpus <- Corpus(DirSource("spam/spam"))
  #Downloaded from: http://spamassassin.apache.org/publiccorpus/20021010_spam.tar.bz2

ham_corpus <- Corpus(DirSource("easy_ham/easy_ham"))
  #Downloaded from: http://spamassassin.apache.org/publiccorpus/20021010_easy_ham.tar.bz2

meta(spam_corpus, tag = "type") <- "spam"
meta(ham_corpus, tag = "type") <- "ham"

#Combine Corpi to create a training corupus

training_courpus <- sample(c(ham_corpus, spam_corpus, recursive=T))

is_spam <- ifelse( unlist(meta(training_courpus, "type")[,1]) == "spam",1,0)

Clean training corpus

training_courpus <- tm_map( training_courpus, str_replace_all, pattern = "[[:punct:]]", replacement = " ")
training_courpus <- tm_map( training_courpus, removeWords, words = stopwords("en"))

training_courpus <- tm_map( training_courpus, tolower)
training_courpus <- tm_map( training_courpus, stemDocument)
training_courpus <- tm_map( training_courpus, PlainTextDocument)

training_tdm <- DocumentTermMatrix(training_courpus)

Train the SVM algo on

t_corpus_length <- length(is_spam)

training_container <- create_container(
  training_tdm,
  labels = is_spam,
  trainSize = 1:floor(t_corpus_length*.2),
  testSize = ceiling(t_corpus_length*.8) : t_corpus_length,
  virgin = FALSE
)

svm_model <- train_model(training_container, "SVM")

Score the results

svm_out <- classify_model(training_container, svm_model)
svm_summary <- create_analytics(training_container, svm_out)
svm_summary@algorithm_summary

##   SVM_PRECISION SVM_RECALL SVM_FSCORE
## 0          1.00       0.99       0.99
## 1          0.95       0.98       0.96

Run the alog on new data

Note: This new corpus is 100% spam

Load the unkown corpus

unkown_corpus <-    Corpus(DirSource("spam_2/spam_2")) #https://spamassassin.apache.org/publiccorpus/20050311_spam_2.tar.bz2

meta(unkown_corpus, tag = "type") <- "unkown"

Clean the unkwown corpus

unkown_corpus <- tm_map( unkown_corpus, str_replace_all, pattern = "[[:punct:]]", replacement = " ")
unkown_corpus <- tm_map( unkown_corpus, removeWords, words = stopwords("en"))

unkown_corpus <- tm_map( unkown_corpus, tolower)
unkown_corpus <- tm_map( unkown_corpus, stemDocument)

unkown_corpus <- tm_map( unkown_corpus, PlainTextDocument)

Run the classify the unkown corpus

training_tdm <- DocumentTermMatrix(training_courpus) 

unkown_tdm <- DocumentTermMatrix(unkown_corpus, list(dictionary=findFreqTerms(training_tdm)) )


unkown_container <- create_container(
  unkown_tdm,
  labels = vector(),
  testSize = 1:nrow(unkown_tdm),
  virgin = TRUE
)


unkown_results_summary <- classify_model(unkown_container, svm_model)

Get a summary of the results of the classification

kable(count(unkown_results_summary, 'SVM_LABEL'))

SVM_LABEL	freq
0	259
1	1138

Note: remember when viewing results 1 = spam and 0 = ham

Word Clouds

Word Cloud of Unkown Corpus

wordcloud(unkown_corpus, max.words = 100)

Word Cloud of Known Spam Corpus

wordcloud(spam_corpus, max.words = 100)

Word Cloud of Known Ham Corpus

wordcloud(ham_corpus, max.words = 100)

Conclusions

An SVM is pretty good about classifing emails as spam or ham, but I should have probably scored other models. However, my computer is apparently too slow for text processing.

CUNY-DATA-607-Week10-Rnash

Raphael Nash

10/31/2016

Assignment

Train the Algo

Run the alog on new data

Word Clouds

Conclusions