Last modified on 21 Jun 2020.

Introduction

In this challenge, we are going to answer the question: “What sorts of people were more likely to survive?” using passenger data. Datasets to be used: train.csv (for training and predicting), test.csv (for submitting).

Open this html fileOpen In Colab

First 10 rows of dataset First 10 rows of the dataset.

Variable Definition Key
survival Survival 0 = No, 1 = Yes
pclass Ticket class 1 = 1st, 2 = 2nd, 3 = 3rd
sex Sex  
Age Age in years  
sibsp # of siblings / spouses aboard the Titanic  
parch # of parents / children aboard the Titanic  
ticket Ticket number  
fare Passenger fare  
cabin Cabin number  
embarked Port of Embarkation C = Cherbourg, Q = Queenstown, S = Southampton

TL;DR;

  • Take an overview about dataset.
    • .describe for numerical / categorical features.
    • Find percentage of missing data on each feature.
    • Survival based on some categorical features.
    • Visualize survival based on Age.
    • Check if the result depends on the titles indicated in the Name?
  • Preprocessing data:
    • Drop unnecessary features (columns) (Name, Ticket, Cabin) using df.drop().
    • Convert categorical variables to dummy ones using pd.get_dummies().
    • Impute missing continuous values using sklearn.impute.SimpleImputer.
    • Take an idea to change Age to a categorical feature and then also convert to dummy.
  • Using GridSearchCV to find the optimal hyper parameters and apply some algorithms, e.g. Random Forest.
  • Export the result to an output file.

Preliminaries

import numpy as np
import matplotlib.pyplot as plt # plot
import pandas as pd # working with dataset

from sklearn import preprocessing
from sklearn.impute import SimpleImputer # impute missing data

from sklearn.model_selection import GridSearchCV, cross_val_score

Overview datasets

Read data

train = pd.read_csv("train.csv")
test = pd.read_csv("test.csv")

Take a look

train.head(10)
train.info()
train.info()
train.describe() # for numerical features
train.describe(include=['O']) # for categorical features

Find the percentage of missing data on each feature,

total = train.isnull().sum().sort_values(ascending=False)
percent = (round(train.isnull().sum()/train.isnull().count()*100, 1)).sort_values(ascending=False)
pd.concat([total, percent], axis=1, keys=['Total', '% of missing'])

Survival based on some categorical features,

train.pivot_table(index="Sex", values="Survived")
train.pivot_table(index="Pclass", values="Survived")
train.pivot_table(index="SibSp", values="Survived")
train.pivot_table(index="Parch", values="Survived")

Visualize survival based on Age (numerical),

train[train["Survived"]==1]['Age'].plot.hist(alpha=0.5, color='blue', bins=50) # survived
train[train["Survived"]==0]['Age'].plot.hist(alpha=0.5, color='blue', bins=50) # died

List of titles (Mr., Mrs., Dr.,…) from Name,

train.Name.str.extract(' ([A-Za-z]+)\.', expand=False)

Preprocessing data

In this task, you have to do the same techniques for both train and test sets!

Drop unnecessary features

Drop some unnecessary features (columns),

train.drop(['Name', 'Ticket', 'Cabin'], axis=1, inplace=True)
test.drop(['Name', 'Ticket', 'Cabin'], axis=1, inplace=True)

Convert to dummy

Convert categorical features to dummy variables,

def create_dummies(df, column_name):
  # Convert the column_name training feature into dummies using one-hot
  #   and leave one first category to prevent perfect collinearity
  dummies = pd.get_dummies(df[column_name], prefix=column_name, drop_first=True)
  df = pd.concat([df, dummies], axis=1)
  return df
# Sex
train = create_dummies(train, 'Sex')
test = create_dummies(test, 'Sex')
# Embarked
train = create_dummies(train, 'Embarked')
test = create_dummies(test, 'Embarked')
# Social Class
train = create_dummies(train, 'Pclass')
test = create_dummies(test, 'Pclass')

Impute Missing Values

For continuous variables, we wanna fill missing data with the mean value.

def impute_data(df_train, df_test, column_name):
  imputer = SimpleImputer(missing_values=np.nan, strategy='mean', verbose=0)
  # Fit the imputer object on the training data
  imputer.fit(df_train[column_name].values.reshape(-1, 1)) # transform single column to 1
  # Apply the imputer object to the df_train and df_test
  df_train[column_name] = imputer.transform(df_train[column_name].values.reshape(-1, 1))
  df_test[column_name] = imputer.transform(df_test[column_name].values.reshape(-1, 1))
  return df_train, df_test
# Age
train, test = impute_data(train, test, 'Age')
# Fare
train, test = impute_data(train, test, 'Fare')

Continuous to categorical

In the case, for example, you wanna convert Age feature which is initially a numerical feature to a categorical feature (many ranges of ages, for example).

def process_age(df, cut_points, label_names):
    df["Age"] = df["Age"].fillna(-0.5)
    df["Age_categories"] = pd.cut(df["Age"], cut_points, labels=label_names)
    return df

cut_points = [-1, 0, 5, 12, 18, 35, 60, 100]
label_names = ["Missing", 'Infant', "Child", 'Teenager', "Young_Adult", 'Adult', 'Senior']

train = process_age(main, cut_points, label_names)
test = process_age(test, cut_points, label_names)

Convert to a dummy variable,

main = create_dummies(main, 'Age_categories')
test = create_dummies(test, 'Age_categories')

Training with Random Forest

We will use Grid Search to test with different parameters and then choose the best ones.

# Create a dictionary containing all the candidate values of the parameters
parameter_grid = dict(n_estimators=list(range(1, 5001, 1000)),
                      criterion=['gini','entropy'],
                      max_features=list(range(1, len(features), 2)),
                      max_depth= [None] + list(range(5, 25, 1)))

# Creata a random forest object
random_forest = RandomForestClassifier(random_state=0, n_jobs=-1)

# Create a gridsearch object with 5-fold cross validation, and uses all cores (n_jobs=-1)
clf = GridSearchCV(estimator=random_forest, param_grid=parameter_grid, cv=5, verbose=1, n_jobs=-1)

Split into X_train, y_train:

X_train = train[train.columns.difference(['Survived'])]
y_train = train['Survived']
# Nest the gridsearchCV in a 3-fold CV for model evaluation
cv_scores = cross_val_score(clf, X_train, y_train)

# Print results
print('Accuracy scores:', cv_scores)
print('Mean of score:', np.mean(cv_scores))
print('Variance of scores:', np.var(cv_scores))

Retrain The Random Forest With The Optimum Parameters

# Retrain the model on the whole dataset
clf.fit(X_train, y_train)
# Predict who survived in the test dataset
predictions = clf.predict(test)

Create an output file

final_ids = test["PassengerId"]
submission_df = {"PassengerId": final_ids,
                 "Survived": predictions}
submission = pd.DataFrame(submission_df)

submission.to_csv('titanic_submission.csv', index=False)

Another way, check the last section of this post.

Other approaches

  • Based on the number of family/sibling members: combination of SibSp and Parch.
  • Go alone?
  • Consider the title from Name.
  • Use Decision Tree with K-fold.

References