Clouress example: Titanic #0

This notebook is a variation of Pradeep Tripathi's Titanic Kaggle solution in R. Instead of writing it in R as the original, we write it in Clojure, and call R from Clojure.

The goal is to study the Clojure-R interop, and expecially experiment with various ways to define Clojure functions corresponding to R functions.

In this first, naive version, the corresponding Clojure functions are rather simple. They expect a varying number of arguments, and pass those arguments to R function calls by a rather generic way (as defined by Clojisrs)

We do not try to replace the rather imperative style of the original tutorial. Rather, we try to write something that is as close as possible to the original.

We have leanred a lot from this use case. It did expose lots of issues and open questions about the Clojisr API and implementation. Note, however, that the piece of R code that we are mimicing here is not so typical to the current tidyverse trends – there is no heavy use of dplyr, tidy evaluation, etc. It may be a good idea to study other examples that have more of those.

This notebook has been written in notespace – an experimental Clojure library that allows one to use regular Clojure namespaces as notebooks, thus enabling interactive literate programming from ones' favourite Clojure editor and REPL.

daslu, Jan. 2020

Bringing the neecessary R functions

Here are most of the functions that we need, brought by the standard require-r mechanism, inspired by libpython-clj's require-python (though not as sophisticated at the moment). In function names, dots are changed to hyphens.

Introduction – Prediction of Titanic Survival using Random Forest

Pradeep Tripathi's solution will use randomForest to create a model predicting survival on the Titanic.

Reading test and train data

This step assumes that the Titanic data lies under the resources/data/ path under your Clojure project.

# Original code:
list.files('../input')

# Original code:
train <-read.csv('../input/train.csv', stringsAsFactors = F)
test  <-read.csv('../input/test.csv', stringsAsFactors = F)

Combining test and train data

As explained by Thripathi, the Random Forest algorithm will use the Bagging method to create multiple random samples with replacement from the dataset, that will be treated as training data, while the out of bag samples will be treated as test data.

# Original code:
titanic<-bind_rows(train,test)

Data check

# Original code:
str(titanic)
summary(titanic)
head(titanic)

'data.frame':	1309 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" ...

Tripathi: We've got a sense of our variables, their class type, 3 and the first few observations of each. We know we're working with 1309 observations of 12 variables.

Feature engineering

Thripathi's explanation: We can break down Passenger name into additional meaningful variables which can feed predictions or be used in the creation of additional new variables. For instance, passenger title is contained within the passenger name variable and we can use surname to represent families.

# Original code:
colnames(titanic)

Retrieve title from passenger names

# Original code:
titanic$title<-gsub('(.*, )|(\..*)', '', titanic$Name)

Show title counts by sex

# Original code:
table(titanic$Sex, titanic$title)

Clojisr can covert an R frequency table to a Clojure data structure:

Sometimes, it is convenient to first convert it to an R data frame:

Sometimes, it is convenient to use the way R prints a frequency table.

Convert titles with low count into a new title, and rename/reassign Mlle, Ms and Mme.

# Original code:
unusual_title<-c('Dona', 'Lady', 'the Countess','Capt', 'Col', 'Don', 
                 'Dr', 'Major', 'Rev', 'Sir', 'Jonkheer')

# Original code:
titanic$title[titanic$title=='Mlle']<-'Miss'
titanic$title[titanic$title=='Ms']<-'Miss'
titanic$title[titanic$title=='Mme']<-'Mrs'
titanic$title[titanic$title %in% unusual_title]<-'Unusual Title'

Check the title count again:

# Original code:
table(titanic$Sex, titanic$title)

Create a variable which contain the surnames of passengers.

# Original code:
titanic$surname<-sapply(titanic$Name, function(x) strsplit(x,split='[,.]')[[1]][1])
nlevels(factor(titanic$surname)) ## 875 unique surnames

Tripathi: Family size variable: We are going to create a variable "famsize" to know the number of family members. It includes number of sibling/number of parents and children+ passenger themselves

# Original code:
titanic$famsize <- titanic$SibSp + titanic$Parch + 1

Create a family variable:

# Original code:
titanic$family <- paste(titanic$surname, titanic$famsize, sep='_')

Visualize the relationship between family size & survival:

 ggplot(titanic[1:891,], aes(x = famsize, fill = factor(Survived))) +
   geom_bar(stat='count', position='dodge') +
   scale_x_continuous(breaks=c(1:11)) +
   labs(x = 'Family Size') +
   theme_few()

Tripathi: Explanation: We can see that there's a survival penalty to single/alone, and those with family sizes above 4. We can collapse this variable into three levels which will be helpful since there are comparatively fewer large families.

Discretize family size:

# Original code:
titanic$fsizeD[titanic$famsize == 1] <- 'single'
titanic$fsizeD[titanic$famsize < 5 & titanic$famsize> 1] <- 'small'
titanic$fsizeD[titanic$famsize> 4] <- 'large'

Let us check if it makes sense:

And let us make sure there are no missing values:

Tripathi: There's could be some useful information in the passenger cabin variable including about their deck, so Retrieve deck from Cabin variable.

# Original code:
titanic$Cabin[1:28]

The first character is the deck:

# Original code:
strsplit(titanic$Cabin[2], NULL) [[1]]

Deck variable:

# Original R code:
titanic$deck<-factor(sapply(titanic$Cabin, function(x) strsplit(x, NULL)[[1]][1]))

Let us check:

Missing values

Thripathi's explanation, following the summary:

• Age : 263 missing values
• Fare : 1 missing values
• Embarked : 2 missing values
• survived:too many
• Cabin : too many

Missing value in Embarkment – Tripathi: Now we will explore missing values and rectify it through imputation. There are a number of different ways we could go about doing this. Given the small size of the dataset, we probably should not opt for deleting either entire observations (rows) or variables (columns) containing missing values. We're left with the option of replacing missing values with sensible values given the distribution of the data, e.g., the mean, median or mode.

To know which passengers have no listed embarkment port:

# Original code:
titanic$Embarked[titanic$Embarked == ""] <- NA
titanic[(which(is.na(titanic$Embarked))), 1]

Marking as missing:

Checking which has missing port:

Tripathi: Passengers 62 and 830 are missing Embarkment.

# Original code:
titanic[c(62, 830), 'Embarked']

Tripathi: So Passenger numbers 62 and 830 are each missing their embarkment ports. Let's look at their class of ticket and their fare.

# Original code:
titanic[c(62, 830), c(1,3,10)]

Alternatively:

Thripathi's explanation: Both passengers had first class tickets that they spent 80 (pounds?) on. Let's see the embarkment ports of others who bought similar kinds of tickets.

First way of handling missing value in Embarked:

# Original code:
titanic%>%
  group_by(Embarked, Pclass) %>%
  filter(Pclass == "1") %>%
  filter(Pclass == "1") %>%
  filter(Pclass == "1") %>%
  summarise(mfare = median(Fare),n = n())

Tripathi: Looks like the median price for a first class ticket departing from 'C' (Charbourg) was 77 (in comparison to our 80). While first class tickets departing from 'Q' were only slightly more expensiive (median price 90), only 3 first class passengers departed from that port. It seems far more likely that passengers 62 and 830 departed with the other 141 first-class passengers from Charbourg.

Second Way of handling missing value in Embarked:

# Original code:
embark_fare <- titanic %>%
  filter(PassengerId != 62 & PassengerId != 830)
embark_fare

Use ggplot2 to visualize embarkment, passenger class, & median fare:

# Original code:
ggplot(embark_fare, aes(x = Embarked, y = Fare, fill = factor(Pclass))) +
geom_boxplot() +
geom_hline(aes(yintercept=80), 
              colour='red', linetype='dashed', lwd=2) +
scale_y_continuous(labels=dollar_format()) +
theme_few()

Tripathi: From plot we can see that The median fare for a first class passenger departing from Charbourg ('C') coincides nicely with the $80 paid by our embarkment-deficient passengers. I think we can safely replace the NA values with 'C'. Since their fare was $80 for 1st class, they most likely embarked from 'C'.

# Original code:
titanic$Embarked[c(62, 830)] <- 'C'

A missing value in fare. Thripathi's explanation: To know Which passenger has no fare information:

# Original code:
titanic[(which(is.na(titanic$Fare))) , 1] 

Tripathi: Looks like Passenger number 1044 has no listed Fare

Where did this passenger leave from? What was their class?

# Original code:
 titanic[1044, c(3, 12)]

Tripathi: Another way to know about passenger id 1044 :Show row 1044

# Original code:
titanic[1044, ]

Thripathi's explanation: Looks like he left from 'S' (Southampton) as a 3rd class passenger. Let's see what other people of the same class and embarkment port paid for their tickets.

First way:

titanic%>% filter(Pclass == '3' & Embarked == 'S') %>% summarise(missing_fare = median(Fare, na.rm = TRUE))

Tripathi: Looks like the median cost for a 3rd class passenger leaving out of Southampton was 8.05. That seems like a logical value for this passenger to have paid.

Second way:

# Original code:
ggplot(titanic[titanic$Pclass == '3' & titanic$Embarked == 'S', ],
       aes(x = Fare)) +
  geom_density(fill = '#99d6ff', alpha=0.4) +
  geom_vline(aes(xintercept=median(Fare, na.rm=T)),
             colour='red', linetype='dashed', lwd=1) +
  scale_x_continuous(labels=dollar_format()) +
  theme_few()

Tripathi: From this visualization, it seems quite reasonable to replace the NA Fare value with median for their class and embarkment which is $8.05.

Replace that NA with 8.05

# Original code:
titanic$Fare[1044] <- 8.05
summary(titanic$Fare)

Tripathi: Another way of Replace missing fare value with median fare for class/embarkment:

# Original code:
titanic$Fare[1044] <- median(titanic[titanic$Pclass == '3' & titanic$Embarked == 'S', ]$Fare, na.rm = TRUE)

Missing Value in Age.

Tripathi: Show number of missing Age values.

# Original code:
sum(is.na(titanic$Age)) 

Tripathi: 263 passengers have no age listed. Taking a median age of all passengers doesn't seem like the best way to solve this problem, so it may be easiest to try to predict the passengers' age based on other known information.

To predict missing ages, I'm going to use the mice package. To start with I will factorize the factor variables and then perform mice(multiple imputation using chained equations).

Set a random seed:

# Original code:
set.seed(129)

Tripathi: Perform mice imputation, excluding certain less-than-useful variables:

# Original code:
mice_mod <- mice(titanic[, !names(titanic) %in% c('PassengerId','Name','Ticket','Cabin','Family','Surname','Survived')], method='rf') 

Save the complete output.

# Original code:
mice_output <- complete(mice_mod)

Tripathi: Let's compare the results we get with the original distribution of passenger ages to ensure that nothing has gone completely awry.

Plot age distributions:

# Original code:
par(mfrow=c(1,2))
hist(titanic$Age, freq=F, main='Age: Original Data',
     col='darkred', ylim=c(0,0.04))
hist(mice_output$Age, freq=F, main='Age: MICE Output',
     col='lightgreen', ylim=c(0,0.04))

Tripathi: Things look good, so let's replace our age vector in the original data with the output from the mice model.

Replace Age variable from the mice model:

# Original code:
titanic$Age <- mice_output$Age

Show new number of missing Age values

# Original code:
sum(is.na(titanic$Age))

Feature Enginnering: Part 2

Tripathi: I will create a couple of new age-dependent variables: Child and Mother. A child will simply be someone under 18 years of age and a mother is a passenger who is 1) female, 2) is over 18, 3) has more than 0 children and 4) does not have the title 'Miss'.

Relationship between age & survival: I include Sex since we know it's a significant predictor.

# Original code:
ggplot(titanic[1:891,], aes(Age, fill = factor(Survived))) +
  geom_histogram() + facet_grid(.~Sex) + theme_few()

Tripathi: Create the column Child, and indicate whether child or adult:

# Original code:
 titanic$Child[titanic$Age < 18] <- 'Child'
 titanic$Child[titanic$Age >= 18] <- 'Adult'

Show counts:

# Original code:
table(titanic$Child, titanic$Survived)

Adding Mother variable:

# Original code:
titanic$Mother <- 'Not Mother'
titanic$Mother[titanic$Sex == 'female' & titanic$Parch >0 & titanic$Age > 18 & titanic$title != 'Miss'] <- 'Mother'

Show counts:

# Original code:
table(titanic$Mother, titanic$Survived)

Factorizing variables:

# Original code:
titanic$Child <- factor(titanic$Child)
titanic$Mother <- factor(titanic$Mother)
titanic$Pclass <- factor(titanic$Pclass)
titanic$Sex <- factor(titanic$Sex)
titanic$Embarked <- factor(titanic$Embarked)
titanic$Survived <- factor(titanic$Survived)
titanic$title <- factor(titanic$title)
titanic$fsizeD <- factor(titanic$fsizeD)

Check classes of all columns:

Prediction

Split into training & test sets:

# Original code:
train <- titanic[1:891,]
test <- titanic[892:1309,]

Building the model:

Tripathi: We then build our model using randomForest on the training set.

Set a random seed:

# Original code:
set.seed(754)

Tripathi: Build the model (note: not all possible variables are used):

# Original code:
titanic_model <- randomForest(Survived ~ Pclass + Sex + Age + SibSp + Parch +
                                Fare + Embarked + title +
                                fsizeD + Child + Mother,
                              data = train)

Show model error:

# Original code:
plot(titanic_model, ylim=c(0,0.36))
legend('topright', colnames(titanic_model$err.rate), col=1:3, fill=1:3)

Tripathi: The black line shows the overall error rate which falls below 20%. The red and green lines show the error rate for 'died' and 'survived', respectively. We can see that right now we're much more successful predicting death than we are survival.

Variable Importance

Get importance:

# Original code:
importance    <- importance(titanic_model)
varImportance <- data.frame(Variables = row.names(importance),
                            Importance = round(importance[ ,'MeanDecreaseGini'],2))

Variable importance

Create a rank variable based on importance:

# Original code:
rankImportance <- varImportance %>%
   mutate(Rank = paste0('#',dense_rank(desc(Importance))))

Tripathi: Use ggplot2 to visualize the relative importance of variables

# Original code:
ggplot(rankImportance, aes(x = reorder(Variables, Importance),
                           y = Importance, fill = Importance)) +
  geom_bar(stat='identity') +
  geom_text(aes(x = Variables, y = 0.5, label = Rank),
            hjust=0, vjust=0.55, size = 4, colour = 'red') +
  labs(x = 'Variables') +
  coord_flip() +
  theme_few()

Tripathi: From the plot we can see that the 'title' variable has the highest relative importance out of all of our predictor variables.

Final Prediction

Predict using the test set:

# Original code:
prediction <- predict(titanic_model, test)
prediction

Tripathi: Save the solution to a dataframe with two columns: PassengerId and Survived (prediction).

# Original code:
Output<- data.frame(PassengerID = test$PassengerId, Survived = prediction)
Output

Write the Output to file:

# Original code:
write.csv(Output, file = 'pradeep_titanic_output.csv', row.names = F)

Conclusion

Tripathi: Thank you for taking the time to read through my first exploration of a Titanic Kaggle dataset. Again, this newbie welcomes comments and suggestions!