Data

preparation and feature engineering on Titanic data set

For this Lab, we will use the Titanic data set, available from Kaggle.com:
http://www.kaggle.com/c/titanic-gettingStarted/data
Load the data (training and test sets)
titanic.train <- read.csv("data/titanic/train.csv", stringsAsFactors = F)
titanic.test <- read.csv("data/titanic/test.csv", stringsAsFactors = F)

Let’s start by examining the structure of the data sets Note: description of all the varibles is
available at the Kaggle website
str(titanic.train)

## 'data.frame': 891 obs. of 12 variables:

## $ PassengerId: int 1 2 3 4 5 6 7 8 9 10 ...
## $ Survived : int 0 1 1 1 0 0 0 0 1 1 ...
## $ Pclass : int 3 1 3 1 3 3 1 3 3 2 ...
## $ Name : chr "Braund, Mr. Owen Harris" "Cumings, Mrs. John Bradley
(Florence Briggs Thayer)" "Heikkinen, Miss. Laina" "Futrelle, Mrs. Jacques
Heath (Lily May Peel)" ...
## $ Sex : chr "male" "female" "female" "female" ...
## $ Age : num 22 38 26 35 35 NA 54 2 27 14 ...
## $ SibSp : int 1 1 0 1 0 0 0 3 0 1 ...
## $ Parch : int 0 0 0 0 0 0 0 1 2 0 ...
## $ Ticket : chr "A/5 21171" "PC 17599" "STON/O2. 3101282" "113803"
...
## $ Fare : num 7.25 71.28 7.92 53.1 8.05 ...
## $ Cabin : chr "" "C85" "" "C123" ...
## $ Embarked : chr "S" "C" "S" "S" ...

str(titanic.test)

## 'data.frame': 418 obs. of 11 variables:

## $ PassengerId: int 892 893 894 895 896 897 898 899 900 901 ...
## $ Pclass : int 3 3 2 3 3 3 3 2 3 3 ...
## $ Name : chr "Kelly, Mr. James" "Wilkes, Mrs. James (Ellen Needs)"
"Myles, Mr. Thomas Francis" "Wirz, Mr. Albert" ...
## $ Sex : chr "male" "female" "male" "male" ...
## $ Age : num 34.5 47 62 27 22 14 30 26 18 21 ...
## $ SibSp : int 0 1 0 0 1 0 0 1 0 2 ...
## $ Parch : int 0 0 0 0 1 0 0 1 0 0 ...
## $ Ticket : chr "330911" "363272" "240276" "315154" ...
## $ Fare : num 7.83 7 9.69 8.66 12.29 ...
## $ Cabin : chr "" "" "" "" ...
## $ Embarked : chr "Q" "S" "Q" "S" ...
The structure of the training and test sets is almost exactly the same (as expected). In fact,
the only difference is the Survived column that is present in the training, but absent in the
test set - it is the response (outcome) variable, that is, the variable with the class values.

Detecting missing values

Let’s start by checking if the data is complete, that is, if there are some missing values. One
way to do that is through the summary f. which will let us know if a variable has NA values
summary(titanic.train)

## PassengerId Survived Pclass Name

## Min. : 1.0 Min. :0.0000 Min. :1.000 Length:891
## 1st Qu.:223.5 1st Qu.:0.0000 1st Qu.:2.000 Class :character
## Median :446.0 Median :0.0000 Median :3.000 Mode :character
## Mean :446.0 Mean :0.3838 Mean :2.309
## 3rd Qu.:668.5 3rd Qu.:1.0000 3rd Qu.:3.000
## Max. :891.0 Max. :1.0000 Max. :3.000
##
## Sex Age SibSp Parch
## Length:891 Min. : 0.42 Min. :0.000 Min. :0.0000
## Class :character 1st Qu.:20.12 1st Qu.:0.000 1st Qu.:0.0000
## Mode :character Median :28.00 Median :0.000 Median :0.0000
## Mean :29.70 Mean :0.523 Mean :0.3816
## 3rd Qu.:38.00 3rd Qu.:1.000 3rd Qu.:0.0000
## Max. :80.00 Max. :8.000 Max. :6.0000
## NA's :177
## Ticket Fare Cabin Embarked
## Length:891 Min. : 0.00 Length:891 Length:891
## Class :character 1st Qu.: 7.91 Class :character Class :character
## Mode :character Median : 14.45 Mode :character Mode :character
## Mean : 32.20
## 3rd Qu.: 31.00
## Max. :512.33
##

It seems that in the training set only Age has missing values, and quite a number of them
(177).
summary(titanic.test)

## PassengerId Pclass Name Sex

## Min. : 892.0 Min. :1.000 Length:418 Length:418
## 1st Qu.: 996.2 1st Qu.:1.000 Class :character Class :character
## Median :1100.5 Median :3.000 Mode :character Mode :character
## Mean :1100.5 Mean :2.266
## 3rd Qu.:1204.8 3rd Qu.:3.000
## Max. :1309.0 Max. :3.000
##
## Age SibSp Parch Ticket
## Min. : 0.17 Min. :0.0000 Min. :0.0000 Length:418
## 1st Qu.:21.00 1st Qu.:0.0000 1st Qu.:0.0000 Class :character
## Median :27.00 Median :0.0000 Median :0.0000 Mode :character
## Mean :30.27 Mean :0.4474 Mean :0.3923
## 3rd Qu.:39.00 3rd Qu.:1.0000 3rd Qu.:0.0000
## Max. :76.00 Max. :8.0000 Max. :9.0000
## NA's :86
## Fare Cabin Embarked
## Min. : 0.000 Length:418 Length:418
## 1st Qu.: 7.896 Class :character Class :character
## Median : 14.454 Mode :character Mode :character
## Mean : 35.627
## 3rd Qu.: 31.500
## Max. :512.329
## NA's :1

In the test set, in addition to the 86 NAs for Age, there is also one missing value for the Fare
variable.
So, based on the NA values, it seems that only Age variable has a serious issue with missing
values.
However, if you take a closer look at the output of the str() f., you’ll notice that for some
observations (passengers) the value for Cabin seems to be missing, that is, Cabin value is
equal to empty string (“”). Let’s inspect this more closely by checking how many “” values
we have for the Cabin variable in both datasets:
length(which(titanic.train$Cabin==""))

## [1] 687

length(which(titanic.test$Cabin==""))

## [1] 327

So, for 687 passengers in the training set and 327 passanges in the test, we have “” as the
Cabin value. Should we consider these as missing values?
Recall that on Titanic, there were three classes of passengers, and only those from the 1st
class were offered a cabin. So, some of the empty string values we have observed are due to
the fact that passengers were from the 2nd or the 3rd class, meaning that they really didn’t
have a cabin. In those cases empty string is not a missing value, but “not applicable” value.
However, passengers from the 1st class should have had a cabin; so, an empty string for the
Cabin value of a 1st class passenger is a ‘real’ missing value. Let’s check how many such
values we have in the training set:
train.class1.no.cabin <- which(titanic.train$Pclass==1 &
titanic.train$Cabin=="")
length(train.class1.no.cabin)

## [1] 40
Also, on the test set:
test.class1.no.cabin <- which(titanic.test$Pclass==1 &
titanic.test$Cabin=="")
length(test.class1.no.cabin)

## [1] 27

So, for 40 1st class passengers in the training set and 27 1st class passengers in the test set,
the Cabin value is missing. To make this explicit, let’s replace the missing Cabin values for
1st class passengers with NAs:
titanic.train$Cabin[train.class1.no.cabin] <- NA
titanic.test$Cabin[test.class1.no.cabin] <- NA

We can check the results of this transformation:

length(which(is.na(titanic.train$Cabin)))

## [1] 40

length(which(is.na(titanic.test$Cabin)))

## [1] 27

Note that we have discovered missing values of the Cabin variable by spotting a few empty
strings in the output of the str() f. However, if those values were not amongst the first
couple of values listed by str(), they would have passed unnoticed. So, let’s check other
string variables for missing values ‘hidden’ as empty strings:
apply(X = titanic.train[,c("Name","Sex","Ticket","Embarked")],
MARGIN = 2,
FUN = function(x) length(which(x=="")))

## Name Sex Ticket Embarked

## 0 0 0 2

In the training set, only for the Embarked variable, we have 2 missing values.
apply(X = titanic.test[,c("Name","Sex","Ticket","Embarked")],
MARGIN = 2,
FUN = function(x) length(which(x=="")))

## Name Sex Ticket Embarked

## 0 0 0 0

In the test set, none of the examined variables has missing values.
We’ll set the two missing values of Embarked to NA, as we did with the Cabin.
titanic.train$Embarked[titanic.train$Embarked==""] <- NA
We have now examined all the variables for the missing values. Before proceeding with
‘fixing’ the missing values, let’s see how we can make use of visualizations to more easily
spot missing values.
An easy way to get a high level view on the data completeness is to visualize the data using
some functions from the Amelia R package
#install.packages('Amelia')
library(Amelia)

## Warning: package 'Rcpp' was built under R version 3.4.3

We will use the missmap() f. to plot the missing data from the traning and test sets
par(mfrow=c(1,2)) # structure the display area to show two plots in the same
row
missmap(obj = titanic.train, main = "Training set", legend = FALSE)
missmap(obj = titanic.test, main = "Test set", legend = FALSE)

par(mfrow=c(1,1)) # reverting plotting area to the default (one plot per row)

Note: the detection of missing values in the missmap() f. is based on the NA values; so, if we
hadn’t transformed those empty strings (for Cabin and Embarked) into NAs, they wouldn’t
be visualized as missing.
Handling missing values
Let’s now see how to deal with missing values. We’ll start with those cases that are easier to
deal with, that is, variables where we have just a few missing values.

Categorical variables with a small number of missing values

In our datasets, Embarked variables falls into this category:
unique(titanic.train$Embarked)

## [1] "S" "C" "Q" NA

unique(titanic.test$Embarked)

## [1] "Q" "S" "C"

So, as we see, Embarked is essentially a nominal (categorical) variable with 3 possible

values (‘S’, ‘C’, and ‘Q’). And, we have seen that it has 2 missing values (in the train set).
In a situation like this, the missing values are replaced by the ‘majority class’, that is, the
most dominant value
xtabs(~Embarked, data = titanic.train)

## Embarked
## C Q S
## 168 77 644

So, “S” is the dominant value, and it will be used as a replacement for NAs
titanic.train$Embarked[is.na(titanic.train$Embarked)] <- 'S'
xtabs(~Embarked, data = titanic.train)

## Embarked
## C Q S
## 168 77 646

Let’s also make Embarked a ‘true’ categorical variable by transforming it into a factor
variable:
titanic.train$Embarked <- factor(titanic.train$Embarked)
titanic.test$Embarked <- factor(titanic.test$Embarked)

Numerical variables with a small number of missing values

In our data set, Fare variable belongs to this category - it is a numerical variable with 1
missing value (in the test set)
A typical way to deal with missing values in situations like this is to replace them with the
average value of the variable on a subset of observations that are the closest (most similar)
to the observation(s) with the missing value. One way to do this is to apply the kNN
method. However, we can opt for a simpler approach: we will replace the missing Fare
value with the average Fare value for the passengers of the same class (Pclass).
First, we need to check the distribution of the Fare variable, to decide if we should use mean
or median as the average value
shapiro.test(titanic.test$Fare)

##
## Shapiro-Wilk normality test
##
## data: titanic.test$Fare
## W = 0.5393, p-value < 2.2e-16

The variable is not normaly distributed -> use median

Now, identify the passenger class (Pclass) of the passenger whose Fare is missing
missing.fare.pclass <- titanic.test$Pclass[is.na(titanic.test$Fare)]

Compute median Fare for the other passengers of the same class
median.fare <- median(x = titanic.test$Fare[titanic.test$Pclass ==
missing.fare.pclass],
na.rm = T) # we have to set this to true as Fare has
one NA value

Set the missing Fare value to the computed median value

titanic.test$Fare[is.na(titanic.test$Fare)] <- median.fare

Check if the NA value was really replaced

summary(titanic.test$Fare)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 0.000 7.896 14.454 35.561 31.472 512.329

Variables with many missing values and/or missing values that are difficult to replace
The Age variable is an example of the first type: variable with many missing value; Cabin is
an example of the second type, as it is a categorical variable with many different values
(~150)
For such variables we apply the process known as imputation - the process of replacing
missing values with substituted (predicted) values. It is, in fact, the task of predicting (good
substitutes for) the missing values. R has several packages for imputation: MICE, Amelia,
HMisc,…
We are not going to do imputation (out of scope of this course), but will instead create new
variables (features) that will, in a way, serve as substitutes or proxies for Age and Cabin.
Feature selection
To select features to be used for creating a prediction model, we have to examine if and to
what extent they are associated with the response (outcome) variable.
If we are familiar with the domain of the problem (prediction task), we can start from the
knowledge and/or intuition about the predictors. Otherwise, that is, if the domain is
unknown to us (such as would be prediction of the outcome of some chemical reactions) or
the real names (labels) of the variables are withdrawn (e.g. for privacy reasons), we have to
rely on some well establish general methods for feature selection (such as forward or
backward selection).
Since the Titanic data set is associated with a familiar domain, we can start from some
intuition about potential predictors.

Examining the predictive power of variables from the data set

It’s well-known that in disasters woman and children are often the first to be rescued. Let’s
check if that was the case in the Titanic disaster.
We’ll start by looking at the survival based on the gender. First, let’s see the proportion of
males and females in the dataset
titanic.train$Sex <- factor(titanic.train$Sex)
summary( titanic.train$Sex )

## female male
## 314 577

prop.table(summary( titanic.train$Sex ))

## female male
## 0.352413 0.647587

Now, examine the survival counts based on the gender

xtabs(~Sex + Survived, data = titanic.train)

## Survived
## Sex 0 1
## female 81 233
## male 468 109

and the proportions

sex.surv.tbl <- prop.table(xtabs(~Sex + Survived,
data = titanic.train),
margin = 1) # proportions are computed at the row
level (each row sums to 1)
sex.surv.tbl
## Survived
## Sex 0 1
## female 0.2579618 0.7420382
## male 0.8110919 0.1889081

Obviously, gender is highly associated with the survival.

Before inspecting if/how age group has affected the chances for survival, let’s quickly take a
look at the potential impact of the passenger class (1st, 2nd or 3rd), as it is reasonable to
expect that those from a higher class would have had higher chances of survival. We can do
that again using tables, but it might be more effective to examine it visually, using the
ggplot2 package:
library(ggplot2)

For plotting the survival against the passenger class, we need to transform both variables
into factor variables (they are given as variables of type int)
titanic.train$Survived <- factor(titanic.train$Survived,
levels = c(0,1), labels = c('No','Yes'))

titanic.train$Pclass <- factor(titanic.train$Pclass,
levels = c(1,2,3),
labels = c("1st", "2nd", "3rd"))

gp1 <- ggplot(data = titanic.train,

mapping = aes(x = Pclass, fill=Survived)) +
geom_bar(position = "dodge", width = 0.4) +
ylab("Number of passengers") + xlab("Passenger class") +
theme_bw()
gp1

The chart suggests that passenger class is another relevant predictor.
Let’s examine passenger class and gender together
gp2 <- gp1 + facet_wrap(~Sex)
gp2

Let’s also inspect if the place of embarkment (the Embarked variable) affected the survival
gp3 <- ggplot(data = titanic.train,
mapping = aes(x = Embarked, fill = Survived)) +
geom_bar(position = "dodge", width = 0.45) +
ylab("Number of passengers") + xlab("Place of embarkment") +
theme_bw()
gp3

It seems that those who embarked in Cherbourg had higher chance of surviving than the
passengers who embarked in the other two ports. Though not as strong as Sex and Pclass,
this variable seems to be a viable candidate for a predictor.

Feature engineering
When creating new features (attributes) to be used for prediction purposes, we need to
base those features on the data from both the training and the test sets, so that the features
are available both for training the prediction model, and making predictions on the unseen
test data.
Hence, we should merge the training and the test sets and develop new features on the
merged data. But before we do that, we need to assure that the training and the test sets
have exactly the same structure. To that end, we will first add the Survived column to the
test data, as a factor variable with the same levels as in the training set:
titanic.test$Survived <- factor(NA, levels = levels(titanic.train$Survived))

Next, we need to transform the Pclass, Sex, and Embarked variables in the test set into
factors, since we’ve done that in the training set (the structure should be exactly the same)
titanic.test$Pclass <- factor(x = titanic.test$Pclass,
levels = c(1,2,3),
labels = levels(titanic.train$Pclass))
titanic.test$Sex <- factor(x = titanic.test$Sex,
levels = c("female", "male"),
 labels = levels(titanic.train$Sex))
titanic.test$Embarked <- factor(x = titanic.test$Embarked,
levels = c("S", "C", "Q"),
labels = levels(titanic.test$Embarked))

Make the order of the columns in the test set the same as in the train set:
titanic.test <- titanic.test[,names(titanic.train)]

Now, we can merge the two datasets

titanic.all <- rbind(titanic.train, titanic.test)

Creating an age proxy variable

Recall that the Age variable has a lot of missing values, and simple imputation methods we
considered cannot be used in such cases. So, we will create a new variable that
approximates the passengers’ age group. We’ll do that by making use of the Name variable.
To start, let’s first inspect the values of this variable
titanic.all$Name[1:10]

## [1] "Braund, Mr. Owen Harris"

## [2] "Cumings, Mrs. John Bradley (Florence Briggs Thayer)"
## [3] "Heikkinen, Miss. Laina"
## [4] "Futrelle, Mrs. Jacques Heath (Lily May Peel)"
## [5] "Allen, Mr. William Henry"
## [6] "Moran, Mr. James"
## [7] "McCarthy, Mr. Timothy J"
## [8] "Palsson, Master. Gosta Leonard"
## [9] "Johnson, Mrs. Oscar W (Elisabeth Vilhelmina Berg)"
## [10] "Nasser, Mrs. Nicholas (Adele Achem)"

We can observe that the Name variable consists of surname, title, first name, and in some
cases additional name (maiden name of married woman).
The idea is to use the title of a person as a rough proxy for his/her age.
First, we need to extract title from the Name variable; to that end, we’ll split the Name
string using “,” or “.” as delimiters; lets’ try it first:
strsplit(x = titanic.all$Name[1], split = "[,|.]")

## [[1]]
## [1] "Braund" " Mr" " Owen Harris"

We get a list of vectors, where each vector consists of pieces of a person’s name. To extract
the title, we need to simplify the output, so that instead of a list, we get a vector (with the
elements of a person’s name)
unlist(strsplit(x = titanic.all$Name[1], split = "[,|.]"))
## [1] "Braund" " Mr" " Owen Harris"

and then, take the second element of that vector:

unlist(strsplit(x = titanic.all$Name[1], split = "[,|.]"))[2]

## [1] " Mr"

You might have noticed a space before the title, we’ll remove that quickly, but before that,
we’ll apply this procedure to all the rows in the titanic.all dataset to create a new feature:
titanic.all$Title <- sapply(titanic.all$Name,
FUN = function(x) unlist(strsplit(x, split =
"[,|.]"))[2] )

Now, let’s remove that leading blank space

titanic.all$Title <- trimws(titanic.all$Title, which = "left")

Note: if trimws() f. is not working on your computer, use str_trim() f. from the stringr R
package.
We can now inspect different kinds of titles we have in the dataset
table(titanic.all$Title)

##
## Capt Col Don Dona Dr
## 1 4 1 1 8
## Jonkheer Lady Major Master Miss
## 1 1 2 61 260
## Mlle Mme Mr Mrs Ms
## 2 1 757 197 2
## Rev Sir the Countess
## 8 1 1

There are some rarely occuring titles that won’t be much usefull for creating a model; so,
we’ll aggregate those titles into broader categories that represent some basic age-gender
groups:
adult.women <- c("Dona", "Lady", "Mme", "Mrs", "the Countess")
girls <- c("Ms", "Mlle", "Miss")
adult.men <- c("Capt", "Col", "Don", "Dr", "Major", "Mr", "Rev", "Sir")
boys <- c("Master", "Jonkheer")

First, we’ll introduce a new variable (feature) to represent the age-gender group
titanic.all$AgeGender <- vector(mode = "character", length =
nrow(titanic.all))

and, now define each age-gender group using the Title groupings we defined above
titanic.all$AgeGender[ titanic.all$Title %in% adult.women ] <- "AdultWomen"
titanic.all$AgeGender[ titanic.all$Title %in% adult.men ] <- "AdultMen"
titanic.all$AgeGender[ titanic.all$Title %in% girls ] <- "Girls"
titanic.all$AgeGender[ titanic.all$Title %in% boys ] <- "Boys"

Note: the %in% operator checks to see if a value is an element of the given vector
Let’s see how passengers are distributed across our age-gender groups:
table(titanic.all$AgeGender)

##
## AdultMen AdultWomen Boys Girls
## 782 201 62 264

We observe a high disproportion in the number of boys and girls, and man and woman.
Let’s take a closer look at the groups with unexpectedly high number of passengers, namely
Girls and AdultMen groups.
We’ll make use of the available values of the Age variable to see how our Girls group is
distributed with respect age.
ggplot(data = titanic.all[titanic.all$AgeGender=="Girls",],
mapping = aes(x = Age)) +
geom_density() +
theme_bw()

## Warning: Removed 51 rows containing non-finite values (stat_density).

It is obvious from the graph that the Girls group includes a considerable number of adult
women. We’ll need to fix this. But before that, let’s also inspect the AdultMen group.
ggplot(data = titanic.all[titanic.all$AgeGender=="AdultMen", ],
mapping = aes(x = Age)) +
geom_density() +
scale_x_continuous(breaks = seq(5,80,5)) +
theme_bw()

## Warning: Removed 177 rows containing non-finite values (stat_density).

From this plot we can see that the AdultMen group also includes some males who cannot be
qualified as adults.
We will try to fix both problems using the available values of the Age variable.
First, let’s check for how many passengers in the ‘Girls’ group the Age value is available:
nrow(titanic.all[titanic.all$AgeGender=="Girls" & !is.na(titanic.all$Age),])

## [1] 213

So, we have Age value for 213 out of 264 Girls, which is not bad at all (80%). We’ll make use
of these available Age values to move some Girls to AdultWomen group, using 18 years of
age as the threshold:
titanic.all$AgeGender[titanic.all$AgeGender=="Girls" &
!is.na(titanic.all$Age) &
titanic.all$Age >= 18] <- "AdultWomen"

We’ll do a similar thing for the AdultMen group. First, check the number of AdultMen
passengers for whom age is available:
nrow(titanic.all[titanic.all$AgeGender=="AdultMen" &
!is.na(titanic.all$Age),])

## [1] 605

We have Age value for 605 out of 782 AdultMen passengers (77%). Let’s make use of those
values to move some passengers from AdultMen to Boys group using, again, the 18 year
threshold
titanic.all$AgeGender[titanic.all$AgeGender=="AdultMen" &
!is.na(titanic.all$Age) &
titanic.all$Age < 18] <- "Boys"

Let’s check the AgeGender proportions after these modifications

table(titanic.all$AgeGender)

##
## AdultMen AdultWomen Boys Girls
## 753 347 91 118

round(prop.table(table(titanic.all$AgeGender)), digits = 2)

##
## AdultMen AdultWomen Boys Girls
## 0.58 0.27 0.07 0.09

This looks far more realistic.

Finally, we’ll transform AgeGender into a factor variable, so that it can be better used for
data exploration and prediction purposes
titanic.all$AgeGender <- factor(titanic.all$AgeGender)
summary(titanic.all$AgeGender)

## AdultMen AdultWomen Boys Girls

## 753 347 91 118

Let’s see if our efforts in creating the AgeGender variable were worthwhile, that is, if
AgeGender is likely to be a significant predictor. To that end, we will plot the AgeGender
groups against the Survival variable.
ggplot(data = titanic.all[1:891,],
mapping = aes(x = AgeGender, fill=Survived)) +
geom_bar(position = "dodge") +
theme_bw()

Note: we are using only the first 891 observations in the merged dataset as these are
observations from the training set for which we know the outcome (i.e., survival).
Let’s examine this also as percentages. First, we need to compute the percentages
age.gen.surv.tbl <- prop.table(table(AgeGender =
titanic.all$AgeGender[1:891],
Survived = titanic.all$Survived[1:891]),
margin = 1)
age.gen.surv.tbl

## Survived
## AgeGender No Yes
## AdultMen 0.8349515 0.1650485
## AdultWomen 0.2212389 0.7787611
## Boys 0.6031746 0.3968254
## Girls 0.3563218 0.6436782

Note that we are setting the margin parameter to 1 as we want to have percentages of
survived and not-survived (column values) computed for each AgeGender group (row)
individually. Try setting margin to 2 and not setting it at all to observe the effect.
For plotting, we’ll transform the table into a data frame
age.gen.surv.df <- as.data.frame(age.gen.surv.tbl)
age.gen.surv.df
## AgeGender Survived Freq
## 1 AdultMen No 0.8349515
## 2 AdultWomen No 0.2212389
## 3 Boys No 0.6031746
## 4 Girls No 0.3563218
## 5 AdultMen Yes 0.1650485
## 6 AdultWomen Yes 0.7787611
## 7 Boys Yes 0.3968254
## 8 Girls Yes 0.6436782

Note the difference in the structure of the table and the data frame
ggplot(data = age.gen.surv.df,
mapping = aes(x = AgeGender, y = Freq, fill=Survived)) +
geom_col(position = "dodge", width = 0.5) +
ylab("Proportion") +
theme_bw()

Obviously, the age/gender group affects survival.

Creating FamilySize variable

Recall that we have two variable related to the number of family members one is travelling
with:
• SibSp - the number of siblings and spouses a passenger is travelling with
• Parch - the number of parents and children one is travelling with
summary(titanic.all$SibSp)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 0.0000 0.0000 0.0000 0.4989 1.0000 8.0000

summary(titanic.all$Parch)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 0.000 0.000 0.000 0.385 0.000 9.000

To get a better insight into the number of family members passengers were travelling with,
we’ll create a new variable FamilySize by simply adding the value of the SibSp and Parch
variables:
titanic.all$FamilySize <- titanic.all$SibSp + titanic.all$Parch
summary(titanic.all$FamilySize)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 0.0000 0.0000 0.0000 0.8839 1.0000 10.0000

We can observe that large majority of passengers didn’t travel with family members.
table(titanic.all$FamilySize)

##
## 0 1 2 3 4 5 6 7 10
## 790 235 159 43 22 25 16 8 11

It can be also observed that those who travelled with 3+ family members were not that
numerous
length(which(titanic.all$FamilySize>=3))/length(titanic.all$FamilySize)

## [1] 0.09549274

Only 10% of passengers travelled with 3+ family members. In situations like this - several
values of a variable spread across a small proportion of the observations - it is
recommended to aggregate those values. We’ll apply that practice to the FamilySize
variable and aggregate observations with 3+ family members:
titanic.all$FamilySize[titanic.all$FamilySize > 3] <- 3

and turn FamilySize into a factor:

titanic.all$FamilySize <- factor(titanic.all$FamilySize,
levels = c(0,1,2,3), labels = c("0", "1",
"2", "3+"))
table(titanic.all$FamilySize)

##
## 0 1 2 3+
## 790 235 159 125
Let’s see how this new feature affects the survival prospects
ggplot(data = titanic.all[1:891,],
mapping = aes(x = FamilySize, fill = Survived)) +
geom_bar(position = "dodge", width = 0.5) +
theme_light()

We can see that those who travelled with 1 or 2 family members had better prospects than
those who travelled without family members or with 3+ family members.

Making use of the Ticket variable

Let’s examine the Ticket variable and see if we can make some use of it
titanic.all$Ticket[1:20]

## [1] "A/5 21171" "PC 17599" "STON/O2. 3101282"

## [4] "113803" "373450" "330877"
## [7] "17463" "349909" "347742"
## [10] "237736" "PP 9549" "113783"
## [13] "A/5. 2151" "347082" "350406"
## [16] "248706" "382652" "244373"
## [19] "345763" "2649"
We can observe that some tickets start with letters, while others consist of digits only.
length(unique(titanic.all$Ticket))

## [1] 929

929 unique ticket values for 1309 passengers suggests that some passengers were
travelling on the same ticket. Let’s examine this further as a shared ticket is an indicator
that a passenger was not travelling alone, and we saw that the number of people one was
travelling with might have had affected their survival prospects.
# tapply, as applied here, computes the number of occurrences of each unique
Ticket value
ticket.count <- tapply(titanic.all$Ticket,
INDEX = titanic.all$Ticket,
FUN = function(x) sum( !is.na(x) ))
ticket.count.df <- data.frame(ticket=names(ticket.count),
count=as.integer(ticket.count))
head(ticket.count.df)

## ticket count
## 1 110152 3
## 2 110413 3
## 3 110465 2
## 4 110469 1
## 5 110489 1
## 6 110564 1

Let’s examine the number of passengers per single and shared tickets:
table(ticket.count.df$count)

##
## 1 2 3 4 5 6 7 8 11
## 713 132 49 16 7 4 5 2 1

We can see that majority of passengers travelled on a single person ticket, a considerable
number of them shared a ticket with one person, and a small number shared their ticket
with 3+ people.
We’ll add ticket count to each passenger by merging titanic.all dataset with the
ticket.count.df based on the ticket value:
titanic.all <- merge(x = titanic.all, y = ticket.count.df,
by.x = "Ticket", by.y = "ticket",
all.x = TRUE, all.y = TRUE)

# change the name of the newly added column:
colnames(titanic.all)[16] <- "PersonPerTicket"

As we did with FamilySize, we’ll aggregate infrequent values of PersonPerTicket and

transform the variable into a factor
titanic.all$PersonPerTicket[titanic.all$PersonPerTicket > 3] <- 3
titanic.all$PersonPerTicket <- factor(titanic.all$PersonPerTicket,
levels = c(1,2,3), labels = c("1", "2",
"3+"))
table(titanic.all$PersonPerTicket)

##
## 1 2 3+
## 713 264 332

Out of curiosity, we can crosstab this variable with FamilySize to see if there were some
passengers who were not travelling with family members but still had company, as well as
those who really travelled alone
xtabs(~ PersonPerTicket + FamilySize, data = titanic.all)

## FamilySize
## PersonPerTicket 0 1 2 3+
## 1 663 31 16 3
## 2 62 170 25 7
## 3+ 65 34 118 115

Let’s examine the PersonPerTicket feature from the perspective of its relevance for a
passenger’s survival
ggplot(data = titanic.all[!is.na(titanic.all$Survived),],
mapping = aes(x = PersonPerTicket, fill=Survived)) +
geom_bar(position = "dodge", width = 0.5) +
theme_light()

It seems that this feature could be a useful predictor. Note that when we merged the
titanic.all and ticket.count.df data frames, the order of rows in the titanic.all changed, so it is
not the case any more that the first 891 observations are those taken from the training set
and the rest are from the test set. Therefore, in the data argument (of ggplot()) we had to
select observations based on having value for the Survived attribute.
Let’s also check what a plot based on percentages would look like.
Compute first the percentages of survived and not survived for each PersonPerTicket value:
tcount.surv.tbl <- prop.table(table(PersonPerTicket =
titanic.all$PersonPerTicket,
Survived = titanic.all$Survived,
useNA = "no"),
margin = 1)
tcount.surv.tbl

## Survived
## PersonPerTicket No Yes
## 1 0.7297297 0.2702703
## 2 0.4861878 0.5138122
## 3+ 0.4803493 0.5196507

In the table() f. we used the useNA argument to restrict the computations to only those
observations where the Survived variable is not NA (that is, observations are from the
training set).
Transform the table into a data frame (required for plotting):
tcount.surv.df <- as.data.frame(tcount.surv.tbl)
tcount.surv.df

## PersonPerTicket Survived Freq

## 1 1 No 0.7297297
## 2 2 No 0.4861878
## 3 3+ No 0.4803493
## 4 1 Yes 0.2702703
## 5 2 Yes 0.5138122
## 6 3+ Yes 0.5196507

ggplot(data = tcount.surv.df,
mapping = aes(x = PersonPerTicket, y = Freq*100, fill=Survived)) +
geom_col(width = 0.5, position = "dodge") +
theme_light() +
ylab("Percentage")

It seems that this variable can, indeed, be worth including in a prediction model.

Save the augmented data set

Finally, let’s split the augmented data set again into training and test parts and save them.
Training observations are those that have the Survived value set; test observations have NA
value for the Survived attribute
ttrain.new <- titanic.all[!is.na(titanic.all$Survived),]
ttest.new <- titanic.all[is.na(titanic.all$Survived),]
saveRDS(ttrain.new, file = "data/titanic/train_new.RData")
saveRDS(ttest.new, file = "data/titanic/test_new.RData")