IntroToTextAnalytics_Part7.R 12.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438
#
# Copyright 2017 Data Science Dojo
#    
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# 
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# 


#
# This R source code file corresponds to video 7 of the Data Science
# Dojo YouTube series "Introduction to Text Analytics with R" located 
# at the following URL:
#     https://www.youtube.com/watch?v=Fza5szojsU8     
#


# Install all required packages.
install.packages(c("ggplot2", "e1071", "caret", "quanteda", 
                   "irlba", "randomForest"))




# Load up the .CSV data and explore in RStudio.
spam.raw <- read.csv("spam.csv", stringsAsFactors = FALSE, fileEncoding = "UTF-16")
View(spam.raw)


# Clean up the data frame and view our handiwork.
spam.raw <- spam.raw[, 1:2]
names(spam.raw) <- c("Label", "Text")
View(spam.raw)



# Check data to see if there are missing values.
length(which(!complete.cases(spam.raw)))



# Convert our class label into a factor.
spam.raw$Label <- as.factor(spam.raw$Label)



# The first step, as always, is to explore the data.
# First, let's take a look at distibution of the class labels (i.e., ham vs. spam).
prop.table(table(spam.raw$Label))



# Next up, let's get a feel for the distribution of text lengths of the SMS 
# messages by adding a new feature for the length of each message.
spam.raw$TextLength <- nchar(spam.raw$Text)
summary(spam.raw$TextLength)



# Visualize distribution with ggplot2, adding segmentation for ham/spam.
library(ggplot2)

ggplot(spam.raw, aes(x = TextLength, fill = Label)) +
  theme_bw() +
  geom_histogram(binwidth = 5) +
  labs(y = "Text Count", x = "Length of Text",
       title = "Distribution of Text Lengths with Class Labels")



# At a minimum we need to split our data into a training set and a
# test set. In a true project we would want to use a three-way split 
# of training, validation, and test.
#
# As we know that our data has non-trivial class imbalance, we'll 
# use the mighty caret package to create a randomg train/test split 
# that ensures the correct ham/spam class label proportions (i.e., 
# we'll use caret for a random stratified split).
library(caret)
help(package = "caret")


# Use caret to create a 70%/30% stratified split. Set the random
# seed for reproducibility.
set.seed(32984)
indexes <- createDataPartition(spam.raw$Label, times = 1,
                               p = 0.7, list = FALSE)

train <- spam.raw[indexes,]
test <- spam.raw[-indexes,]


# Verify proportions.
prop.table(table(train$Label))
prop.table(table(test$Label))



# Text analytics requires a lot of data exploration, data pre-processing
# and data wrangling. Let's explore some examples.

# HTML-escaped ampersand character.
train$Text[21]


# HTML-escaped '<' and '>' characters. Also note that Mallika Sherawat
# is an actual person, but we will ignore the implications of this for
# this introductory tutorial.
train$Text[38]


# A URL.
train$Text[357]



# There are many packages in the R ecosystem for performing text
# analytics. One of the newer packages in quanteda. The quanteda
# package has many useful functions for quickly and easily working
# with text data.
library(quanteda)
help(package = "quanteda")


# Tokenize SMS text messages.
train.tokens <- tokens(train$Text, what = "word", 
                       remove_numbers = TRUE, remove_punct = TRUE,
                       remove_symbols = TRUE, remove_hyphens = TRUE)

# Take a look at a specific SMS message and see how it transforms.
train.tokens[[357]]


# Lower case the tokens.
train.tokens <- tokens_tolower(train.tokens)
train.tokens[[357]]


# Use quanteda's built-in stopword list for English.
# NOTE - You should always inspect stopword lists for applicability to
#        your problem/domain.
train.tokens <- tokens_select(train.tokens, stopwords(), 
                              selection = "remove")
train.tokens[[357]]


# Perform stemming on the tokens.
train.tokens <- tokens_wordstem(train.tokens, language = "english")
train.tokens[[357]]


# Create our first bag-of-words model.
train.tokens.dfm <- dfm(train.tokens, tolower = FALSE)


# Transform to a matrix and inspect.
train.tokens.matrix <- as.matrix(train.tokens.dfm)
View(train.tokens.matrix[1:20, 1:100])
dim(train.tokens.matrix)


# Investigate the effects of stemming.
colnames(train.tokens.matrix)[1:50]


# Per best practices, we will leverage cross validation (CV) as
# the basis of our modeling process. Using CV we can create 
# estimates of how well our model will do in Production on new,
# unseen data. CV is powerful, but the downside is that it
# requires more processing and therefore more time.
#
# If you are not familiar with CV, consult the following 
# Wikipedia article:
#
#   https://en.wikipedia.org/wiki/Cross-validation_(statistics)
#

# Setup a the feature data frame with labels.
train.tokens.df <- cbind(Label = train$Label, data.frame(train.tokens.dfm))


# Often, tokenization requires some additional pre-processing
names(train.tokens.df)[c(146, 148, 235, 238)]


# Cleanup column names.
names(train.tokens.df) <- make.names(names(train.tokens.df))


# Use caret to create stratified folds for 10-fold cross validation repeated 
# 3 times (i.e., create 30 random stratified samples)
set.seed(48743)
cv.folds <- createMultiFolds(train$Label, k = 10, times = 3)

cv.cntrl <- trainControl(method = "repeatedcv", number = 10,
                         repeats = 3, index = cv.folds)


# Our data frame is non-trivial in size. As such, CV runs will take 
# quite a long time to run. To cut down on total execution time, use
# the doSNOW package to allow for multi-core training in parallel.
#
# WARNING - The following code is configured to run on a workstation-
#           or server-class machine (i.e., 12 logical cores). Alter
#           code to suit your HW environment.
#
#install.packages("doSNOW")
library(doSNOW)


# Time the code execution
start.time <- Sys.time()


# Create a cluster to work on 10 logical cores.
cl <- makeCluster(10, type = "SOCK")
registerDoSNOW(cl)


# As our data is non-trivial in size at this point, use a single decision
# tree alogrithm as our first model. We will graduate to using more 
# powerful algorithms later when we perform feature extraction to shrink
# the size of our data.
rpart.cv.1 <- train(Label ~ ., data = train.tokens.df, method = "rpart", 
                    trControl = cv.cntrl, tuneLength = 7)


# Processing is done, stop cluster.
stopCluster(cl)


# Total time of execution on workstation was approximately 4 minutes. 
total.time <- Sys.time() - start.time
total.time


# Check out our results.
rpart.cv.1



# The use of Term Frequency-Inverse Document Frequency (TF-IDF) is a 
# powerful technique for enhancing the information/signal contained
# within our document-frequency matrix. Specifically, the mathematics
# behind TF-IDF accomplish the following goals:
#    1 - The TF calculation accounts for the fact that longer 
#        documents will have higher individual term counts. Applying
#        TF normalizes all documents in the corpus to be length 
#        independent.
#    2 - The IDF calculation accounts for the frequency of term
#        appearance in all documents in the corpus. The intuition 
#        being that a term that appears in every document has no
#        predictive power.
#    3 - The multiplication of TF by IDF for each cell in the matrix
#        allows for weighting of #1 and #2 for each cell in the matrix.


# Our function for calculating relative term frequency (TF)
term.frequency <- function(row) {
  row / sum(row)
}

# Our function for calculating inverse document frequency (IDF)
inverse.doc.freq <- function(col) {
  corpus.size <- length(col)
  doc.count <- length(which(col > 0))

  log10(corpus.size / doc.count)
}

# Our function for calculating TF-IDF.
tf.idf <- function(x, idf) {
  x * idf
}


# First step, normalize all documents via TF.
train.tokens.df <- apply(train.tokens.matrix, 1, term.frequency)
dim(train.tokens.df)
View(train.tokens.df[1:20, 1:100])


# Second step, calculate the IDF vector that we will use - both
# for training data and for test data!
train.tokens.idf <- apply(train.tokens.matrix, 2, inverse.doc.freq)
str(train.tokens.idf)


# Lastly, calculate TF-IDF for our training corpus.
train.tokens.tfidf <-  apply(train.tokens.df, 2, tf.idf, idf = train.tokens.idf)
dim(train.tokens.tfidf)
View(train.tokens.tfidf[1:25, 1:25])


# Transpose the matrix
train.tokens.tfidf <- t(train.tokens.tfidf)
dim(train.tokens.tfidf)
View(train.tokens.tfidf[1:25, 1:25])


# Check for incopmlete cases.
incomplete.cases <- which(!complete.cases(train.tokens.tfidf))
train$Text[incomplete.cases]


# Fix incomplete cases
train.tokens.tfidf[incomplete.cases,] <- rep(0.0, ncol(train.tokens.tfidf))
dim(train.tokens.tfidf)
sum(which(!complete.cases(train.tokens.tfidf)))


# Make a clean data frame using the same process as before.
train.tokens.tfidf.df <- cbind(Label = train$Label, data.frame(train.tokens.tfidf))
names(train.tokens.tfidf.df) <- make.names(names(train.tokens.tfidf.df))


# Time the code execution
start.time <- Sys.time()

# Create a cluster to work on 10 logical cores.
cl <- makeCluster(3, type = "SOCK")
registerDoSNOW(cl)

# As our data is non-trivial in size at this point, use a single decision
# tree alogrithm as our first model. We will graduate to using more 
# powerful algorithms later when we perform feature extraction to shrink
# the size of our data.
rpart.cv.2 <- train(Label ~ ., data = train.tokens.tfidf.df, method = "rpart", 
                    trControl = cv.cntrl, tuneLength = 7)

# Processing is done, stop cluster.
stopCluster(cl)

# Total time of execution on workstation was 
total.time <- Sys.time() - start.time
total.time

# Check out our results.
rpart.cv.2



# N-grams allow us to augment our document-term frequency matrices with
# word ordering. This often leads to increased performance (e.g., accuracy)
# for machine learning models trained with more than just unigrams (i.e.,
# single terms). Let's add bigrams to our training data and the TF-IDF 
# transform the expanded featre matrix to see if accuracy improves.

# Add bigrams to our feature matrix.
train.tokens <- tokens_ngrams(train.tokens, n = 1:2)
train.tokens[[357]]


# Transform to dfm and then a matrix.
train.tokens.dfm <- dfm(train.tokens, tolower = FALSE)
train.tokens.matrix <- as.matrix(train.tokens.dfm)
train.tokens.dfm


# Normalize all documents via TF.
train.tokens.df <- apply(train.tokens.matrix, 1, term.frequency)

  
# Calculate the IDF vector that we will use for training and test data!
train.tokens.idf <- apply(train.tokens.matrix, 2, inverse.doc.freq)

  
# Calculate TF-IDF for our training corpus 
train.tokens.tfidf <-  apply(train.tokens.df, 2, tf.idf, 
                             idf = train.tokens.idf)

  
# Transpose the matrix
train.tokens.tfidf <- t(train.tokens.tfidf)

  
# Fix incomplete cases
incomplete.cases <- which(!complete.cases(train.tokens.tfidf))
train.tokens.tfidf[incomplete.cases,] <- rep(0.0, ncol(train.tokens.tfidf))


# Make a clean data frame.
train.tokens.tfidf.df <- cbind(Label = train$Label, data.frame(train.tokens.tfidf))
names(train.tokens.tfidf.df) <- make.names(names(train.tokens.tfidf.df))


# Clean up unused objects in memory.
gc()




#
# NOTE - The following code requires the use of command-line R to execute
#        due to the large number of features (i.e., columns) in the matrix.
#        Please consult the following link for more details if you wish
#        to run the code yourself:
#
#        https://stackoverflow.com/questions/28728774/how-to-set-max-ppsize-in-r
#
#        Also note that running the following code required approximately
#        38GB of RAM and more than 4.5 hours to execute on a 10-core 
#        workstation!
#


# Time the code execution
# start.time <- Sys.time()

# Leverage single decision trees to evaluate if adding bigrams improves the 
# the effectiveness of the model.
# rpart.cv.3 <- train(Label ~ ., data = train.tokens.tfidf.df, method = "rpart", 
#                     trControl = cv.cntrl, tuneLength = 7)

# Total time of execution on workstation was
# total.time <- Sys.time() - start.time
# total.time

# Check out our results.
# rpart.cv.3

#
# The results of the above processing show a slight decline in rpart 
# effectiveness with a 10-fold CV repeated 3 times accuracy of 0.9457.
# As we will discuss later, while the addition of bigrams appears to 
# negatively impact a single decision tree, it helps with the mighty
# random forest!
#