Welcome to Introduction to R for Data Science Session 4! The course is co-organized by Data Science Serbia and Startit. You will find all course material (R scripts, data sets, SlideShare presentations, readings) on these pages.

Welcome to the fourth session of Introduction to R for Data Science! Check out the Course Overview to acess the learning material presented thus far.

Data Science Serbia Course Pages [in Serbian]

Startit Course Pages [in Serbian]

Lecturers

• dipl. ing Branko Kovač, Data Analyst at CUBE, Data Science Mentor at Springboard, Data Science Serbia
• Goran S. Milovanović, Phd, DataScientist@DiploFoundation, Data Science Serbia

Summary of Session 4, 19. may 2016 :: Introduction to R: Control Flow + Data Structures = R Programs

Control Flow + Data Structures = R Programs; and then vectorize! Loops: for, while, repeat + break, next etc. Branching with if, else, the vectorized ifelse() construct, and switch(). When and why do loops slow down the R interpreter code execution. Introducing vectorization and some data science terminology: features in a classification toy model. The power of vectorization: why does it matter.

• Session 4 SlideShare
• Session 4 R Script
• Further Readings
• Session Photos

Intro to R for Data Science SlideShare :: Session 4

R script :: Session 4

########################################################
# Introduction to R for Data Science
# SESSION 4 :: 19 May, 2016
# Data Science Community Serbia + Startit
# :: Branko Kovač and Goran S. Milovanović ::
########################################################
 
# Starting with simple 'if'
num <- 2 # some value to test with
if (num > 0) print("num is positive") # if condition num > 0 stands than print()
                                      # is executed
 
# Sometimes 'if' has its 'else'
if (num > 0) { # test to see if it's positive
  print("num is positive") # print in case of positive number
} else {
  print("num is negative") # it's negative if not positive
}
 
# Multiple 'else's are also possible
if (!is.numeric(num)) { # check if it's numeric
  print("this is not a number") # print this if it isn't numeric
} else if (num > 0) { # check if it's greater than 0
  print("num is positive") # print if it's positive
} else if (num < 0) { # maybe it is negative
  print("num is negative") # print it if it's < 0
} else print("ZERO!") # who knows, maybe it is zero
 
# R is vectorized so there's vectorized if-else
simple_vect <- c(1, 3, 12, NA, 2, NA, 4) # just another num vector with NAs
ifelse(is.na(simple_vect), "nothing here", "some number") # nothing here if it's
                                                          # NA or it's a number
 
# For loop is always working same way
for (i in simple_vect) print(i) # iterate in set of values
 
# Be aware that loops can be slow if
vec  <-  numeric()
system.time(
  for(i in seq_len(50000-1)) {
    some_calc <- sqrt(i/10)
    vec <- c(vec, some_calc) # this is what makes it slow
  }  
)
 
# This solution is slightly faster
iter <- 50000
vec <- numeric(length=iter) # this makes it faster...
system.time(
  for(i in seq_len(iter-1)) {
    some_calc <- sqrt(i/10)
    vec[i] <- some_calc # ...not this!
  }
)
 
# This solution is even more faster
iter <- 50000
vec <- numeric(length=iter) # not because of this...
system.time(
  for(i in seq_len(iter-1)) {
    vec[i] <- sqrt(i/10) # ...but this!
  }
)
 
# Another example how loops can be slow (loop vs vectorized function)
iter <- 50000
 
system.time(
  for (i in 1:iter) {
    vec[i] <- rnorm(n=1, mean=0, sd=1) # approach from previous example
  }
)
 
system.time(y <- rnorm(iter, 0, 1)) # but this is much much faster
 
# R also knows about while loop
r <- 1 # initializing some variable
while (r < 5) { # while r < 5
  print(r) # print r
  r <- r + 1 # increase r by 1
}
 
# Loops can be nested
for(i in 1:5) { # outer loop
  for(j in 1:5) { # inner loop
    print(paste0(i,j)) # some code
  }
}
 
# Loops can be altered using break and next
for(i in 1:5) {
  if (i == 4) break # jump out of loop if condition is true
  print(i)
}
 
for(i in 1:5) {
  if (i == 4) next # just skip current iteration if condition is true
  print(i)
}
 
# Nope, we didn't forget 'repeat' loop
i <- 1
repeat { # there is no condition...
  print(i)
  i <- i + 1
  if (i == 10) break # ...so we have to break it if we don't want infinite loop
}
 
# And there's something called 'switch' :)
switch(2, "data", "science", "serbia") # choose one option based on value
 
# More on switch:
switchIndicator <- "A"
switchIndicator <- "switchIndicator"
switchIndicator <- "AvAvAv"
# rare situations where you do not need to enclose strings: ' ', or " "
switch(switchIndicator,
       'A' = {print(switchIndicator)},
       'switchIndicator' = {unlist(strsplit(switchIndicator,"h"))},
       'AvAvAv' = {print(nchar(switchIndicator))}
)
# is the same as:
switch(switchIndicator,
       A = {print(switchIndicator)},
       switchIndicator = {unlist(strsplit(switchIndicator,"h"))},
       AvAvAv = {print(nchar(switchIndicator))}
)
 
# now:
type = 2
cc <- c("A", "B", "C");
switch(type,
       c1 = {print(cc[1])},
       c2 = {print(cc[2])},
       c3 = {print(cc[3])},
       {print("Beyond C...")} # default choice
       )
 
# BUT if you do this, R will miss the default choice, so be careful w. switch:
type = 4
cc <- c("A", "B", "C");
switch(type,
       print(cc[1]),
       print(cc[2]),
       print(cc[3]),
       {print("Beyond C...")} # the unnamed default choice works only if previous choices are named!
)
 
# Switch and if-else are similar, but switch is faster (believe us!)
i <- 2
if(i == 1) {
  print("data")
} else if(i == 2) {
  print("science")
} else print("serbia")
 
 
#########################################################
### Exercise, exercise...
 
library(datasets)
head(iris)
is.data.frame(iris)
 
# suming up Sepal.Length + Sepal.Width
for (i in 1:length(iris$Sepal.Length)) {
  iris$Sepal.LW[i] <- iris$Sepal.Length[i] + iris$Sepal.Width[i]
}
iris$Sepal.LW <- NULL;
 
# remember: this is a vector programming language:
iris$Sepal.LW <- iris$Sepal.Length + iris$Sepal.Width # avoid loops anytime when possible!
iris$Sepal.LW <- NULL;
 
out <- length(iris$Sepal.Length);
i <- 0;
repeat
  {
    i <- i+1
    iris$Sepal.LW[i] <- iris$Sepal.Length[i] + iris$Sepal.Width[i];
    if(i==out) break;
  }
# there are many more stupid ways to do this, all in order to avoid the simple and fast:
iris$Sepal.LW <- iris$Sepal.Length + iris$Sepal.Width
iris$Sepal.LW <- NULL;
 
# for example, by avoiding loops altogether...
iris$Sepal.LW <- apply(data.frame(iris$Sepal.Length, iris$Sepal.Width), 1, function(x) {sum(x)});
iris$Sepal.LW <- NULL;
 
# ifelse in lapply in unlist
sepalLengthMean <- mean(iris$Sepal.Length);
iris$Sepal.LW <- unlist(lapply(iris$Sepal.Length, function(x){ifelse(x>=sepalLengthMean,T,F)}));
iris$Sepal.LW <- NULL;
# however...
iris$Sepal.LW <- ifelse(iris$Sepal.Length>=mean(iris$Sepal.Length),T,F);
# now leave the iris data set alone...
iris$Sepal.LW <- NULL;
 
#########################################################
### vectorization in R
 
dataSet <- USArrests;
head(dataSet)
 
# data$Murder, data$Assault, data$Rape: columns of data
 
# in behavioral sciences (psychology or biomedical sciences, for example) we would call them:
# variables (or factors, even more often)
# in data science and machine learning, we usually call them: FEATURES
# in psychology and behavioral sciences, the usage of the term "feature" is usually constrained
# to theories of categorization and concept learning
 
# Task: classify the US states according to some global indicator of violent crime
# Two categories (simplification): more dangerous and less dangerous (F)
# We have three features: Murder, Rape, Assault, all per 100,000 inhabitants
# The idea is to combine the three available features.
 
# Let's assume that we arbitrarily assign the following preference order over the features:
# Murder > Rape > Assault
# in terms of the severity of the consequences of the associated criminal acts
 
# Let's first isolate the features from the data.frame
featureMatrix <- as.matrix(dataSet[, c(1,4,2)]);
 
# Let's WEIGHT the features in accordance with the imposed preference order:
weigthsVector <- c(3,2,1); # mind the order of the columns in featureMatrix
 
# Essentially, we want our global indicator to be a linear combination of all three selected features
# Where each feature is weighted by the corresponding element of the weigthsVector:
 
featureMatrix <- cbind(featureMatrix,numeric(length(featureMatrix[,1])));
for (i in 1:length(featureMatrix[,1])) {
  featureMatrix[i,4] <- sum(weigthsVector*featureMatrix[i,1:3]);
  # don't forget: this "*" multiplication in R is vectorized and operates element-wise
  # we have a 1x3 weightsVector and a 1x3 featureMatrix[i,1:3], Ok
  # sum() then produces the desired linear combination
}
 
# Classification; in the simplest case, let's simply take a look at the distribution of our global indicator:
hist(featureMatrix[,4],20); # it's multimodal and not too symmetric; go for median
criterion <- median(featureMatrix[,4]);
# And classify:
dataSet$Dangerous <- ifelse(featureMatrix[,4]>=criterion,T,F);
 
# Ok. You will never do this before you have a model that has actually *learned* the
# most adequate feature weights. This is an exercise only.
 
# ***Important***: have you seen the for loop above? Well...
# N e v e r  d o  t h a t.
dataSet$Dangerous <- NULL;
 
# In Data Science, you will be working with huge amounts of quantitative data.
# For loops are slow. But in vector programming languages like R...
# matrix computations are seriously fast.
 
# What you ***want to do*** is the following:
 
# Let's first isolate the features from the data.frame
featureMatrix <- as.matrix(dataSet[, c(1,4,2)]);
# Let's WEIGHT the features in accordance with the imposed preference order:
weigthsVector <- c(3,2,1); # mind the order of the columns in featureMatrix
 
# Feature weighting:
wF <- weigthsVector %*% t(featureMatrix);
# In R, t() is for: transpose
# In R, %*% is matrix multiplication
 
# oh yes: R knows about row and column vectors - and you want to put this one
# as a COLUMN in your dataSet data.frame, while wF is currently a ROW vector, look:
wF
length(wF);
wF <- t(wF); 
 
# and classify:
dataSet$Dangerous <- ifelse(wF>=median(wF),T,F);

Readings :: Session 5 [26. May, 2016, @Startit.rs, 19h CET]

R for Data Science, Garrett Grolemund & Hadley Wickham: Chapter 10: Strings

Session 4 Photos