Feature Engineering: Enhancing the Titanic Dataset for Survival Predictions

Welcome back to our course - Data Cleaning and Preprocessing in Machine Learning. Today's mission revolves around Feature Engineering on the Titanic dataset. By the end of today's lesson, your toolkit will be loaded with skills that revolve around feature creation, modification, and encoding. Your expertise in Python and Pandas will also be put into practice, reinforcing your knowledge in the process.

Feature engineering is the process of creating optimized features that improve the effectiveness of machine learning algorithms. This process utilizes the data to create new features and modify existing ones. This might involve creating new features, transforming existing features, or identifying and removing irrelevant ones. For instance, in our Titanic dataset, we have properties or indicators like age, sex, pclass, etc., which might need some optimizing.

Let's take sibsp and parch as an example: sibsp shows the number of siblings/spouses aboard while parch shows the number of parents/children onboard. Because these features both indicate the number of family members onboard for each individual, one might see them as similar features or even overlapping. Hence, we can combine these two features to create a new feature: family_size.

To start creating new features, we need to have the Titanic dataset in our hands. This can be quickly done in Python using Seaborn, as shown below:

import seaborn as sns

titanic_df = sns.load_dataset("titanic")

With the dataset ready, let's compute the family_size feature using the number of siblings/spouses (sibsp) and parents/children (parch).

titanic_df['family_size'] = titanic_df['sibsp'] + titanic_df['parch'] + 1 # +1 for self

This snippet of code will create a new column 'family_size' and add it to the dataset titanic_df.

Another valuable task in feature engineering is to modify existing features. For instance, the fare feature in our Titanic dataset may contain values that are significantly larger than the average, making the distribution of fare values skewed. This skewness could be unhelpful for our machine-learning algorithms and might induce overfitting.

In such a case, we can consider applying a log transformation to the fare column, with the aim of reducing skewness. The code snippet for this is shown below:

import numpy as np

# Adding a small constant to handle zero fares
titanic_df['log_fare'] = np.log(titanic_df['fare'] + 0.1)

Here, we add 0.1 to the fare column to adjust for zero values, then apply a logarithmic transformation.