Thread by @svpino

Yesterday, @PrasoonPratham posted a step-by-step guide to solve the Titanic challenge on Kaggle.

I thought it'd be fun to engineer some features that can help build an even better model.

Here are some ideas worth considering.

↓ 1/10

Attached you can find the original set of input variables that come with the data to solve the problem.

We are going to transform some of these into features that should help our model produce better results.

This is what Feature Engineering is all about.

↓ 2/13

Keep in mind that these are just hypotheses that you'll have to try and validate.

Some of these suggestions might not improve the results or could even make the model perform worse.

This is an exercise to try and think creatively about the data we are getting.

↓ 3/13

1. Let's start with "Age."

It makes sense for the age of a passenger to influence whether the person survives the crash.

People that were too old or young children probably had different chances.

But there's really no difference between a 25 and a 27 years old.

↓ 4/13

We can consolidate the information we want our model to learn by turning "Age" into a one-hot encoded feature with three possible values:

• [1, 0, 0]: Anyone younger than 10.
• [0, 1, 0]: Anyone between 11 and 64.
• [0, 0, 1]: Anyone older than 64.

↓ 5/13

2. We are given a "ticket" variable that represents the ticket number of the passenger.

I don't see how this feature would influence the model's outcome. We can drop it.

What you don't feed to your model is as important as what you give to it.

↓ 6/13

3. Let's work now with the "cabin" number.

The actual number here is not important. We mostly care about whether or not the passenger had a cabin.

This is probably important to determine whether you had time to save yourself based on where you were.

↓ 7/13

We also know that affluent people had a better chance of surviving, and having a cabin is definitely correlated to that.

We can transform "cabin" into a 0/1 feature.

↓ 8/13

Another way to process the "cabin" variable is by extracting the floor where the cabin was located in the ship. (Assuming it's provided.)

I'd guess that people on the top floors had a different chance of survival than those on the bottom floors.

↓ 9/13

4. The "sibsp" and "parch" variables tell us the number of siblings plus spouses and parents plus children, respectively.

Does this matter?

Here is a theory: people with family on the ship might have had a different chance of survival than those traveling alone.

↓ 10/13

Maybe some people couldn't save themselves because they had to protect others.

Maybe some people survived because others protected them.

Either way, we could combine these two variables into a 0/1 feature indicating whether the passenger traveled alone.

↓ 11/13

You can probably take it from here.

The Titanic challenge is a great exercise to explore different hypotheses and practice feature engineering.

↓ 12/13

Two more things:

First, be careful while engineering features not to introduce biases. Use the data you have as much as you can to drive these decisions.

Second, here is Pratham's thread walking through the steps to solve the challenge:

https://twitter.com/PrasoonPratham/status/1384145017145462785?s=20

.

13/13

If you found this thread helpful, follow me @svpino for weekly posts touching on machine learning and how to use it to build real-life systems.

It’s all about building value, and that’s way more fun if we do it together!

That's a good insight. Probably a good idea to explore the following and see if any of them (or both) work:

• has_a_cabin: 0 | 1
• side_of_ship: [0, 0] | [1, 0] | [0, 1]

https://twitter.com/trump1_fan/status/1384493706502447105?s=20