Experimenting with machine learning in R with tidymodels and the Kaggle titanic dataset

1 First version August 13, 2021, updated August 23, 2021

Since my first post, I’ve been reading notebooks shared by folks who ranked high in the challenge, and added two features that they used. Eventually, these new predictors did not help (I must be doing something wrong). I also explored some other ML algorithms. Last, I tuned the parameters more efficiently with a clever grid-search algorithm. All in all, I slightly improved my score, but most importantly, I now have a clean template for further use.

2 Motivation

I would like to familiarize myself with machine learning (ML) techniques in R. So I have been reading and learning by doing. I thought I’d share my experience for others who’d like to give it a try. All material available from GitHub at https://github.com/oliviergimenez/learning-machine-learning.

The two great books I’m using are:

• An Introduction to Statistical Learning with Applications in R by Gareth James, Daniela Witten, Trevor Hastie and Robert Tibshirani

• Tidy models in R by Max Kuhn and Julia Silge

I also recommend checking out the material (codes, screencasts) shared by David Robinson and Julia Silge from whom I picked some useful tricks that I put to use below.

To try things, I’ve joined the Kaggle online community which gathers folks with lots of experience in ML from whom you can learn. Kaggle also hosts public datasets that can be used for playing around.

I use the tidymodels metapackage that contains a suite of packages for modeling and machine learning using tidyverse principles. Check out all possibilities here, and parsnip models in particular there.

Let’s start with the famous Titanic dataset. We need to predict if a passenger survived the sinking of the Titanic (1) or not (0). A dataset is provided for training our models (train.csv). Another dataset is provided (test.csv) for which we do not know the answer. We will predict survival for each passenger, submit our answer to Kaggle and see how well we did compared to other folks. The metric for comparison is the percentage of passengers we correctly predict – aka as accuracy.

First things first, let’s load some packages to get us started.

library(tidymodels) # metapackage for ML 
library(tidyverse) # metapackage for data manipulation and visulaisation
library(stacks) # stack ML models for better perfomance
theme_set(theme_light())
doParallel::registerDoParallel(cores = 4) # parallel computations

3 Data

Read in training data.

rawdata <- read_csv("dat/titanic/train.csv")
glimpse(rawdata)

## Rows: 891
## Columns: 12
## $ PassengerId <dbl> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,…
## $ Survived    <dbl> 0, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1…
## $ Pclass      <dbl> 3, 1, 3, 1, 3, 3, 1, 3, 3, 2, 3, 1, 3, 3, 3, 2, 3, 2, 3, 3…
## $ Name        <chr> "Braund, Mr. Owen Harris", "Cumings, Mrs. John Bradley (Fl…
## $ Sex         <chr> "male", "female", "female", "female", "male", "male", "mal…
## $ Age         <dbl> 22, 38, 26, 35, 35, NA, 54, 2, 27, 14, 4, 58, 20, 39, 14, …
## $ SibSp       <dbl> 1, 1, 0, 1, 0, 0, 0, 3, 0, 1, 1, 0, 0, 1, 0, 0, 4, 0, 1, 0…
## $ Parch       <dbl> 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 1, 0, 0, 5, 0, 0, 1, 0, 0, 0…
## $ Ticket      <chr> "A/5 21171", "PC 17599", "STON/O2. 3101282", "113803", "37…
## $ Fare        <dbl> 7.2500, 71.2833, 7.9250, 53.1000, 8.0500, 8.4583, 51.8625,…
## $ Cabin       <chr> NA, "C85", NA, "C123", NA, NA, "E46", NA, NA, NA, "G6", "C…
## $ Embarked    <chr> "S", "C", "S", "S", "S", "Q", "S", "S", "S", "C", "S", "S"…

naniar::miss_var_summary(rawdata)

After some data exploration (not shown), I decided to take care of missing values, gather the two family variables in a single variable, and create a variable title.

# Get most frequent port of embarkation
uniqx <- unique(na.omit(rawdata$Embarked))
mode_embarked <- as.character(fct_drop(uniqx[which.max(tabulate(match(rawdata$Embarked, uniqx)))]))


# Build function for data cleaning and handling NAs
process_data <- function(tbl){
  
  tbl %>%
    mutate(class = case_when(Pclass == 1 ~ "first",
                             Pclass == 2 ~ "second",
                             Pclass == 3 ~ "third"),
           class = as_factor(class),
           gender = factor(Sex),
           fare = Fare,
           age = Age,
           ticket = Ticket,
           alone = if_else(SibSp + Parch == 0, "yes", "no"), # alone variable
           alone = as_factor(alone),
           port = factor(Embarked), # rename embarked as port
           title = str_extract(Name, "[A-Za-z]+\\."), # title variable
           title = fct_lump(title, 4)) %>% # keep only most frequent levels of title
    mutate(port = ifelse(is.na(port), mode_embarked, port), # deal w/ NAs in port (replace by mode)
           port = as_factor(port)) %>%
    group_by(title) %>%
    mutate(median_age_title = median(age, na.rm = T)) %>%
    ungroup() %>%
    mutate(age = if_else(is.na(age), median_age_title, age)) %>% # deal w/ NAs in age (replace by median in title)
    mutate(ticketfreq = ave(1:nrow(.), FUN = length),
           fareadjusted = fare / ticketfreq) %>%
    mutate(familyage = SibSp + Parch + 1 + age/70)
    
}

# Process the data
dataset <- rawdata %>%
  process_data() %>%
  mutate(survived = as_factor(if_else(Survived == 1, "yes", "no"))) %>%
  mutate(survived = relevel(survived, ref = "yes")) %>% # first event is survived = yes
  select(survived, class, gender, age, alone, port, title, fareadjusted, familyage) 

# Have a look again
glimpse(dataset)

## Rows: 891
## Columns: 9
## $ survived     <fct> no, yes, yes, yes, no, no, no, no, yes, yes, yes, yes, no…
## $ class        <fct> third, first, third, first, third, third, first, third, t…
## $ gender       <fct> male, female, female, female, male, male, male, male, fem…
## $ age          <dbl> 22, 38, 26, 35, 35, 30, 54, 2, 27, 14, 4, 58, 20, 39, 14,…
## $ alone        <fct> no, no, yes, no, yes, yes, yes, no, no, no, no, yes, yes,…
## $ port         <fct> 3, 1, 3, 3, 3, 2, 3, 3, 3, 1, 3, 3, 3, 3, 3, 3, 2, 3, 3, …
## $ title        <fct> Mr., Mrs., Miss., Mrs., Mr., Mr., Mr., Master., Mrs., Mrs…
## $ fareadjusted <dbl> 0.008136925, 0.080003704, 0.008894501, 0.059595960, 0.009…
## $ familyage    <dbl> 2.314286, 2.542857, 1.371429, 2.500000, 1.500000, 1.42857…

naniar::miss_var_summary(dataset)

Let’s apply the same treatment to the test dataset.

rawdata <- read_csv("dat/titanic/test.csv") 
holdout <- rawdata %>%
  process_data() %>%
  select(PassengerId, class, gender, age, alone, port, title, fareadjusted, familyage) 

glimpse(holdout)

## Rows: 418
## Columns: 9
## $ PassengerId  <dbl> 892, 893, 894, 895, 896, 897, 898, 899, 900, 901, 902, 90…
## $ class        <fct> third, third, second, third, third, third, third, second,…
## $ gender       <fct> male, female, male, male, female, male, female, male, fem…
## $ age          <dbl> 34.5, 47.0, 62.0, 27.0, 22.0, 14.0, 30.0, 26.0, 18.0, 21.…
## $ alone        <fct> yes, no, yes, yes, no, yes, yes, no, yes, no, yes, yes, n…
## $ port         <fct> 2, 3, 2, 3, 3, 3, 2, 3, 1, 3, 3, 3, 3, 3, 3, 1, 2, 1, 3, …
## $ title        <fct> Mr., Mrs., Mr., Mr., Mrs., Mr., Miss., Mr., Mrs., Mr., Mr…
## $ fareadjusted <dbl> 0.018730144, 0.016746411, 0.023175837, 0.020723684, 0.029…
## $ familyage    <dbl> 1.492857, 2.671429, 1.885714, 1.385714, 3.314286, 1.20000…

naniar::miss_var_summary(holdout)

4 Exploratory data analysis

skimr::skim(dataset)

Data summary

Name	dataset
Number of rows	891
Number of columns	9
_______________________
Column type frequency:
factor	6
numeric	3
________________________
Group variables	None

Variable type: factor

skim_variable	n_missing	complete_rate	ordered	n_unique	top_counts
survived	0	1	FALSE	2	no: 549, yes: 342
class	0	1	FALSE	3	thi: 491, fir: 216, sec: 184
gender	0	1	FALSE	2	mal: 577, fem: 314
alone	0	1	FALSE	2	yes: 537, no: 354
port	0	1	FALSE	4	3: 644, 1: 168, 2: 77, S: 2
title	0	1	FALSE	5	Mr.: 517, Mis: 182, Mrs: 125, Mas: 40

Variable type: numeric

skim_variable	n_missing	complete_rate	mean	sd	p0	p25	p50	p75	p100	hist
age	0	1	29.39	13.26	0.42	21.00	30.00	35.00	80.00	▂▇▃▁▁
fareadjusted	0	1	0.04	0.06	0.00	0.01	0.02	0.03	0.58	▇▁▁▁▁
familyage	0	1	2.32	1.57	1.07	1.41	1.57	2.62	11.43	▇▁▁▁▁

Let’s explore the data.

dataset %>%
  count(survived)

dataset %>%
  group_by(gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n)

dataset %>%
  group_by(title) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
  arrange(desc(pct_surv))

dataset %>%
  group_by(class, gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
  arrange(desc(pct_surv))

Some informative graphs.

dataset %>%
  group_by(class, gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
    mutate(class = fct_reorder(class, pct_surv)) %>%
    ggplot(aes(pct_surv, class, fill = class, color = class)) +
    geom_col(position = position_dodge()) +
    scale_x_continuous(labels = percent) +
    labs(x = "% in category that survived", fill = NULL, color = NULL, y = NULL) +
  facet_wrap(~gender)

dataset %>%
  mutate(age = cut(age, breaks = c(0, 20, 40, 60, 80))) %>%
  group_by(age, gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
    mutate(age = fct_reorder(age, pct_surv)) %>%
    ggplot(aes(pct_surv, age, fill = age, color = age)) +
    geom_col(position = position_dodge()) +
    scale_x_continuous(labels = percent) +
    labs(x = "% in category that survived", fill = NULL, color = NULL, y = NULL) +
  facet_wrap(~gender)

dataset %>%
    ggplot(aes(fareadjusted, group = survived, color = survived, fill = survived)) +
    geom_histogram(alpha = .4, position = position_dodge()) +
    labs(x = "fare", y = NULL, color = "survived?", fill = "survived?")

dataset %>%
    ggplot(aes(familyage, group = survived, color = survived, fill = survived)) +
    geom_histogram(alpha = .4, position = position_dodge()) +
    labs(x = "family aged", y = NULL, color = "survived?", fill = "survived?")

5 Training/testing datasets

Split our dataset in two, one dataset for training and the other one for testing. We will use an additionnal splitting step for cross-validation.

set.seed(2021)
spl <- initial_split(dataset, strata = "survived")
train <- training(spl)
test <- testing(spl)

train_5fold <- train %>%
  vfold_cv(5)

6 Gradient boosting algorithms - xgboost

Let’s start with gradient boosting methods which are very popular in the ML community.

6.1 Tuning

Set up defaults.

mset <- metric_set(accuracy) # metric is accuracy
control <- control_grid(save_workflow = TRUE,
                        save_pred = TRUE,
                        extract = extract_model) # grid for tuning

First a recipe.

xg_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a gradient boosting model.

xg_model <- boost_tree(mode = "classification", # binary response
                       trees = tune(),
                       mtry = tune(),
                       tree_depth = tune(),
                       learn_rate = tune(),
                       loss_reduction = tune(),
                       min_n = tune()) # parameters to be tuned

Now set our workflow.

xg_wf <- 
  workflow() %>% 
  add_model(xg_model) %>% 
  add_recipe(xg_rec)

Use cross-validation to evaluate our model with different param config.

xg_tune <- xg_wf %>%
  tune_grid(train_5fold,
            metrics = mset,
            control = control,
            grid = crossing(trees = 1000,
                            mtry = c(3, 5, 8), # finalize(mtry(), train)
                            tree_depth = c(5, 10, 15),
                            learn_rate = c(0.01, 0.005),
                            loss_reduction = c(0.01, 0.1, 1),
                            min_n = c(2, 10, 25)))

Visualize the results.

autoplot(xg_tune) + theme_light()

Collect metrics.

xg_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

The tuning takes some time. There are other ways to explore the parameter space more efficiently. For example, we will use the function dials::grid_max_entropy() in the last section about ensemble modelling. Here, I will use finetune::tune_race_anova.

library(finetune)
xg_tune <-
  xg_wf %>%
  tune_race_anova(
    train_5fold,
    grid = 50,
    param_info = xg_model %>% parameters(),
    metrics = metric_set(accuracy),
    control = control_race(verbose_elim = TRUE))

Visualize the results.

autoplot(xg_tune)

Collect metrics.

xg_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

6.2 Fit model

Use best config to fit model to training data.

xg_fit <- xg_wf %>%
  finalize_workflow(select_best(xg_tune)) %>%
  fit(train)

## [20:15:43] WARNING: amalgamation/../src/learner.cc:1095: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

Check out accuracy on testing dataset to see if we overfitted.

xg_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Check out important features (aka predictors).

importances <- xgboost::xgb.importance(model = extract_fit_engine(xg_fit))
importances %>%
  mutate(Feature = fct_reorder(Feature, Gain)) %>%
  ggplot(aes(Gain, Feature)) +
  geom_col()

6.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle. Note that I use the whole dataset, not just the training dataset.

xg_wf %>%
  finalize_workflow(select_best(xg_tune)) %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/xgboost.csv")

## [20:15:45] WARNING: amalgamation/../src/learner.cc:1095: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

I got and accuracy of 0.74162. Cool. Let’s train a random forest model now.

7 Random forests

Let’s continue with random forest methods.

7.1 Tuning

First a recipe.

rf_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a random forest model.

rf_model <- rand_forest(mode = "classification", # binary response
                        engine = "ranger", # by default
                        mtry = tune(),
                        trees = tune(),
                        min_n = tune()) # parameters to be tuned

Now set our workflow.

rf_wf <- 
  workflow() %>% 
  add_model(rf_model) %>% 
  add_recipe(rf_rec)

Use cross-validation to evaluate our model with different param config.

rf_tune <-
  rf_wf %>%
  tune_race_anova(
    train_5fold,
    grid = 50,
    param_info = rf_model %>% parameters(),
    metrics = metric_set(accuracy),
    control = control_race(verbose_elim = TRUE))

Visualize the results.

autoplot(rf_tune)

Collect metrics.

rf_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

7.2 Fit model

Use best config to fit model to training data.

rf_fit <- rf_wf %>%
  finalize_workflow(select_best(rf_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

rf_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Check out important features (aka predictors).

library(vip)
finalize_model(
  x = rf_model,
  parameters = select_best(rf_tune)) %>%
  set_engine("ranger", importance = "permutation") %>%
  fit(survived ~ ., data = juice(prep(rf_rec))) %>%
  vip(geom = "point")

7.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

rf_wf %>%
  finalize_workflow(select_best(rf_tune)) %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/randomforest.csv")

I got and accuracy of 0.77990, a bit better than gradient boosting.

Let’s continue with cat boosting methods.

8 Gradient boosting algorithms - catboost

8.1 Tuning

Set up defaults.

mset <- metric_set(accuracy) # metric is accuracy
control <- control_grid(save_workflow = TRUE,
                        save_pred = TRUE,
                        extract = extract_model) # grid for tuning

First a recipe.

cb_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a cat boosting model.

library(treesnip)
cb_model <- boost_tree(mode = "classification",
                       engine = "catboost",
                       mtry = tune(),
                       trees = tune(),
                       min_n = tune(),
                       tree_depth = tune(),
                       learn_rate = tune()) # parameters to be tuned

Now set our workflow.

cb_wf <- 
  workflow() %>% 
  add_model(cb_model) %>% 
  add_recipe(cb_rec)

Use cross-validation to evaluate our model with different param config.

cb_tune <- cb_wf %>%
  tune_race_anova(
    train_5fold,
    grid = 30,
    param_info = cb_model %>% parameters(),
    metrics = metric_set(accuracy),
    control = control_race(verbose_elim = TRUE))

Visualize the results.

autoplot(cb_tune)

Collect metrics.

cb_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

8.2 Fit model

Use best config to fit model to training data.

cb_fit <- cb_wf %>%
  finalize_workflow(select_best(cb_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

cb_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

8.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

cb_wf %>%
  finalize_workflow(select_best(cb_tune)) %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/catboost.csv")

I got and accuracy of 0.76076. Cool.

9 Regularization methods

Let’s continue with elastic net regularization.

9.1 Tuning

First a recipe.

en_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_normalize(all_numeric_predictors()) %>% # normalize
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a regularization model. We tune parameter mixture, with ridge regression for mixture = 0, and lasso for mixture = 1.

en_model <- logistic_reg(penalty = tune(), 
                         mixture = tune()) %>% # param to be tuned
  set_engine("glmnet") %>% # elastic net
  set_mode("classification") # binary response

Now set our workflow.

en_wf <- 
  workflow() %>% 
  add_model(en_model) %>% 
  add_recipe(en_rec)

Use cross-validation to evaluate our model with different param config.

en_tune <- en_wf %>%
  tune_grid(train_5fold,
            metrics = mset,
            control = control,
            grid = crossing(penalty = 10 ^ seq(-8, -.5, .5),
                            mixture = seq(0, 1, length.out = 10)))

Visualize the results.

autoplot(en_tune)

Collect metrics.

en_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

9.2 Fit model

Use best config to fit model to training data.

en_fit <- en_wf %>%
  finalize_workflow(select_best(en_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

en_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Check out important features (aka predictors).

library(broom)
en_fit$fit$fit$fit %>%
  tidy() %>%
  filter(lambda >= select_best(en_tune)$penalty) %>%
  filter(lambda == min(lambda),
         term != "(Intercept)") %>%
  mutate(term = fct_reorder(term, estimate)) %>%
  ggplot(aes(estimate, term, fill = estimate > 0)) +
  geom_col() +
  theme(legend.position = "none")

9.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

en_wf %>%
  finalize_workflow(select_best(en_tune)) %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/regularization.csv")

I got and accuracy of 0.76315.

10 Logistic regression

And what about a good old-fashioned logistic regression (not a ML algo)?

First a recipe.

logistic_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_normalize(all_numeric_predictors()) %>% # normalize
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a logistic regression.

logistic_model <- logistic_reg() %>% # no param to be tuned
  set_engine("glm") %>% # elastic net
  set_mode("classification") # binary response

Now set our workflow.

logistic_wf <- 
  workflow() %>% 
  add_model(logistic_model) %>% 
  add_recipe(logistic_rec)

Fit model.

logistic_fit <- logistic_wf %>%
  fit(train)

Inspect significant features (aka predictors).

tidy(logistic_fit, exponentiate = TRUE) %>%
  filter(p.value < 0.05)

Same thing, but graphically.

library(broom)
logistic_fit %>%
  tidy() %>%
  mutate(term = fct_reorder(term, estimate)) %>%
  ggplot(aes(estimate, term, fill = estimate > 0)) +
  geom_col() +
  theme(legend.position = "none")

Check out accuracy on testing dataset to see if we overfitted.

logistic_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Confusion matrix.

logistic_fit %>%
  augment(test, type.predict = "response") %>%
  conf_mat(survived, .pred_class)

##           Truth
## Prediction yes  no
##        yes  59  13
##        no   27 125

ROC curve.

logistic_fit %>%
  augment(test, type.predict = "response") %>%
  roc_curve(truth = survived, estimate = .pred_yes) %>%
  autoplot()

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

logistic_wf %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/logistic.csv")

I got and accuracy of 0.76076. Oldies but goodies!

11 Neural networks

We go on with neural networks.

11.1 Tuning

Set up defaults.

mset <- metric_set(accuracy) # metric is accuracy
control <- control_grid(save_workflow = TRUE,
                        save_pred = TRUE,
                        extract = extract_model) # grid for tuning

First a recipe.

nn_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_normalize(all_numeric_predictors()) %>%
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a neural network.

nn_model <- mlp(epochs = tune(), 
                hidden_units = tune(), 
                dropout = tune()) %>% # param to be tuned
  set_mode("classification") %>% # binary response var
  set_engine("keras", verbose = 0)

Now set our workflow.

nn_wf <- 
  workflow() %>% 
  add_model(nn_model) %>% 
  add_recipe(nn_rec)

Use cross-validation to evaluate our model with different param config.

nn_tune <- nn_wf %>%
  tune_race_anova(
    train_5fold,
    grid = 30,
    param_info = nn_model %>% parameters(),
    metrics = metric_set(accuracy),
    control = control_race(verbose_elim = TRUE))

Visualize the results.

autoplot(nn_tune)

Collect metrics.

nn_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

11.2 Fit model

Use best config to fit model to training data.

nn_fit <- nn_wf %>%
  finalize_workflow(select_best(nn_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

nn_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

11.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

nn_wf %>%
  finalize_workflow(select_best(nn_tune)) %>%
  fit(train) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/nn.csv")

I got and accuracy of 0.78708. My best score so far.

12 Support vector machines

We go on with support vector machines.

12.1 Tuning

Set up defaults.

mset <- metric_set(accuracy) # metric is accuracy
control <- control_grid(save_workflow = TRUE,
                        save_pred = TRUE,
                        extract = extract_model) # grid for tuning

First a recipe.

svm_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  # remove any zero variance predictors
  step_zv(all_predictors()) %>% 
  # remove any linear combinations
  step_lincomb(all_numeric()) %>%
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a svm.

svm_model <- svm_rbf(cost = tune(), 
                     rbf_sigma = tune()) %>% # param to be tuned
  set_mode("classification") %>% # binary response var
  set_engine("kernlab")

Now set our workflow.

svm_wf <- 
  workflow() %>% 
  add_model(svm_model) %>% 
  add_recipe(svm_rec)

Use cross-validation to evaluate our model with different param config.

svm_tune <- svm_wf %>%
  tune_race_anova(
    train_5fold,
    grid = 30,
    param_info = svm_model %>% parameters(),
    metrics = metric_set(accuracy),
    control = control_race(verbose_elim = TRUE))

Visualize the results.

autoplot(svm_tune)

Collect metrics.

svm_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

12.2 Fit model

Use best config to fit model to training data.

svm_fit <- svm_wf %>%
  finalize_workflow(select_best(svm_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

svm_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

12.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

svm_wf %>%
  finalize_workflow(select_best(svm_tune)) %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/svm.csv")

I got and accuracy of 0.77511.

13 Decision trees

We go on with decision trees.

13.1 Tuning

Set up defaults.

mset <- metric_set(accuracy) # metric is accuracy
control <- control_grid(save_workflow = TRUE,
                        save_pred = TRUE,
                        extract = extract_model) # grid for tuning

First a recipe.

dt_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_zv(all_predictors()) %>% 
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a decision tree model.

library(baguette)
dt_model <- bag_tree(cost_complexity = tune(),
                     tree_depth = tune(),
                     min_n = tune()) %>% # param to be tuned
  set_engine("rpart", times = 25) %>% # nb bootstraps
  set_mode("classification") # binary response var

Now set our workflow.

dt_wf <- 
  workflow() %>% 
  add_model(dt_model) %>% 
  add_recipe(dt_rec)

Use cross-validation to evaluate our model with different param config.

dt_tune <- dt_wf %>%
  tune_race_anova(
    train_5fold,
    grid = 30,
    param_info = dt_model %>% parameters(),
    metrics = metric_set(accuracy),
    control = control_race(verbose_elim = TRUE))

Visualize the results.

autoplot(dt_tune)

Collect metrics.

dt_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

13.2 Fit model

Use best config to fit model to training data.

dt_fit <- dt_wf %>%
  finalize_workflow(select_best(dt_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

dt_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

13.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

dt_wf %>%
  finalize_workflow(select_best(dt_tune)) %>%
  fit(dataset) %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/dt.csv")

I got and accuracy of 0.76794.

14 Stacked ensemble modelling

Let’s do some ensemble modelling with all algo but logistic and catboost. Tune again with a probability-based metric. Start with xgboost.

library(finetune)
library(stacks)
# xgboost
xg_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)
xg_model <- boost_tree(mode = "classification", # binary response
                       trees = tune(),
                       mtry = tune(),
                       tree_depth = tune(),
                       learn_rate = tune(),
                       loss_reduction = tune(),
                       min_n = tune()) # parameters to be tuned
xg_wf <- 
  workflow() %>% 
  add_model(xg_model) %>% 
  add_recipe(xg_rec)
xg_grid <- grid_latin_hypercube(
  trees(),
  finalize(mtry(), train),
  tree_depth(),
  learn_rate(),
  loss_reduction(),
  min_n(),
  size = 30)
xg_tune <- xg_wf %>%
  tune_grid(
    resamples = train_5fold,
    grid = xg_grid,
    metrics = metric_set(roc_auc),
    control = control_stack_grid())

Then random forests.

# random forest
rf_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% 
  step_dummy(all_nominal_predictors()) 
rf_model <- rand_forest(mode = "classification", # binary response
                        engine = "ranger", # by default
                        mtry = tune(),
                        trees = tune(),
                        min_n = tune()) # parameters to be tuned
rf_wf <- 
  workflow() %>% 
  add_model(rf_model) %>% 
  add_recipe(rf_rec)
rf_grid <- grid_latin_hypercube(
  finalize(mtry(), train),
  trees(),
  min_n(),
  size = 30)
rf_tune <- rf_wf %>%
  tune_grid(
    resamples = train_5fold,
    grid = rf_grid,
    metrics = metric_set(roc_auc),
    control = control_stack_grid())

Regularisation methods (between ridge and lasso).

# regularization methods
en_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_normalize(all_numeric_predictors()) %>% # normalize
  step_dummy(all_nominal_predictors()) 
en_model <- logistic_reg(penalty = tune(), 
                         mixture = tune()) %>% # param to be tuned
  set_engine("glmnet") %>% # elastic net
  set_mode("classification") # binary response
en_wf <- 
  workflow() %>% 
  add_model(en_model) %>% 
  add_recipe(en_rec)
en_grid <- grid_latin_hypercube(
  penalty(),
  mixture(),
  size = 30)
en_tune <- en_wf %>%
  tune_grid(
    resamples = train_5fold,
    grid = en_grid,
    metrics = metric_set(roc_auc),
    control = control_stack_grid())

Neural networks (takes time, so pick only a few values for illustration purpose).

# neural networks
nn_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_normalize(all_numeric_predictors()) %>%
  step_dummy(all_nominal_predictors()) 
nn_model <- mlp(epochs = tune(), 
                hidden_units = 2, 
                dropout = tune()) %>% # param to be tuned
  set_mode("classification") %>% # binary response var
  set_engine("keras", verbose = 0)
nn_wf <- 
  workflow() %>% 
  add_model(nn_model) %>% 
  add_recipe(nn_rec)
# nn_grid <- grid_latin_hypercube(
#   epochs(),
#   hidden_units(),
#   dropout(),
#   size = 10)
nn_tune <- nn_wf %>%
  tune_grid(
    resamples = train_5fold,
    grid = crossing(dropout = c(0.1, 0.2), epochs = c(250, 500, 1000)), # nn_grid
    metrics = metric_set(roc_auc),
    control = control_stack_grid())
#autoplot(nn_tune)

Support vector machines.

# support vector machines
svm_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% 
  step_normalize(all_numeric_predictors()) %>%
  step_dummy(all_nominal_predictors()) 
svm_model <- svm_rbf(cost = tune(), 
                     rbf_sigma = tune()) %>% # param to be tuned
  set_mode("classification") %>% # binary response var
  set_engine("kernlab")
svm_wf <- 
  workflow() %>% 
  add_model(svm_model) %>% 
  add_recipe(svm_rec)
svm_grid <- grid_latin_hypercube(
  cost(),
  rbf_sigma(),
  size = 30)
svm_tune <- svm_wf %>%
  tune_grid(
    resamples = train_5fold,
    grid = svm_grid,
    metrics = metric_set(roc_auc),
    control = control_stack_grid())

Last, decision trees.

# decision trees
dt_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% 
  step_zv(all_predictors()) %>% 
  step_dummy(all_nominal_predictors()) 
library(baguette)
dt_model <- bag_tree(cost_complexity = tune(),
                     tree_depth = tune(),
                     min_n = tune()) %>% # param to be tuned
  set_engine("rpart", times = 25) %>% # nb bootstraps
  set_mode("classification") # binary response var
dt_wf <- 
  workflow() %>% 
  add_model(dt_model) %>% 
  add_recipe(dt_rec)
dt_grid <- grid_latin_hypercube(
  cost_complexity(),
  tree_depth(),
  min_n(),
  size = 30)
dt_tune <- dt_wf %>%
  tune_grid(
    resamples = train_5fold,
    grid = dt_grid,
    metrics = metric_set(roc_auc),
    control = control_stack_grid())

Get best config.

xg_best <- xg_tune %>% filter_parameters(parameters = select_best(xg_tune))
rf_best <- rf_tune %>% filter_parameters(parameters = select_best(rf_tune))
en_best <- en_tune %>% filter_parameters(parameters = select_best(en_tune))
nn_best <- nn_tune %>% filter_parameters(parameters = select_best(nn_tune))
svm_best <- svm_tune %>% filter_parameters(parameters = select_best(svm_tune))
dt_best <- dt_tune %>% filter_parameters(parameters = select_best(dt_tune))

Do the stacked ensemble modelling.

Pile all models together.

blended <- stacks() %>% # initialize
  add_candidates(en_best) %>% # add regularization model
  add_candidates(xg_best) %>% # add gradient boosting
  add_candidates(rf_best) %>% # add random forest
  add_candidates(nn_best) %>% # add neural network
  add_candidates(svm_best) %>% # add svm
  add_candidates(dt_best) # add decision trees
blended

Fit regularized model.

blended_fit <- blended %>%
  blend_predictions() # fit regularized model

Visualise penalized model. Note that neural networks are dropped, despite achieving best score when used in isolation. I’ll have to dig into that.

autoplot(blended_fit)

autoplot(blended_fit, type = "members")

autoplot(blended_fit, type = "weights")

Fit candidate members with non-zero stacking coef with full training dataset.

blended_regularized <- blended_fit %>%
  fit_members() 
blended_regularized

## # A tibble: 3 × 3
##   member                type         weight
##   <chr>                 <chr>         <dbl>
## 1 .pred_no_en_best_1_12 logistic_reg  2.67 
## 2 .pred_no_dt_best_1_11 bag_tree      1.95 
## 3 .pred_no_rf_best_1_05 rand_forest   0.922

Perf on testing dataset?

test %>%
  bind_cols(predict(blended_regularized, .)) %>%
  accuracy(survived, .pred_class)

Now predict.

holdout %>%
  bind_cols(predict(blended_regularized, .)) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/stacked.csv")

I got an 0.76076 accuracy.

15 Conclusions

I covered several ML algorithms and logistic regression with the awesome tidymodels metapackage in R. My scores at predicting Titanic survivors were ok I guess. Some folks on Kaggle got a perfect accuracy, so there is always room for improvement. Maybe better tuning, better features (or predictors) or other algorithms would increase accuracy. Of course, I forgot to use set.seed() so results are not exactly reproducible.