Zeu Park

A curious developer documenting my journey through code, from Python basics to machine learning and beyond

2 minute read

Introduction

When I first encountered machine learning (ML), it was exciting — and also totally overwhelming.
What does it mean to train a model? How do I handle raw data? Do I really need preprocessing if the algorithm is strong enough?

This post is a reflection of the questions I asked, the false starts I had, and how I’m finally beginning to see why feature engineering is the real power behind effective ML.

The Early Questions I Had

Here are some of the honest, beginner questions I asked while starting ML:

  • “Isn’t it ok to not do preprocessing”
  • “feature engineering을 why does it matter? the graphs won’t be used for the ML”
  • “What is group X and y? How is it used in the algorith?”
  • “What is the differences among many ML models? What are the purpose of each model and why are they important?”

And one of the biggest:

  • “How many features are needed for the ML to get high credits?”

These weren’t just questions about code — they were about understanding how ML thinks.

Working With the Russian Number Plate Dataset

To make this more real, I worked on a dataset containing Russian car auction records, which included:

  • plate (numberplate)
  • date, price
  • region_code, region_name
  • government_code, agency
  • forbidden, priority, significance

What I Tried

  • Extracted meaningful segments from the plate (e.g. prefix letters)
  • Determined if a plate was a government car using partial string matching (startswith)
  • Merged regional dictionaries using region_code
  • Removed or transformed outliers in price
  • Created new binary columns from categoricals (like forbidden == 1)
  • Converted date into month and year features

What I Learned

  • Raw columns ≠ features. Just having data isn’t enough — you have to transform it.
  • Feature engineering means knowing the domain. In this case, understanding Russian plates helped me create useful features.
  • Visualizing outliers helped me decide what to drop, not just guess.
  • Government vehicle classification wasn’t in the original data — I had to create that signal from patterns.

What I Thought About

All this made me realize:

  • Machine learning without feature engineering is like cooking without prepping ingredients
  • You can’t rely on the algorithm to figure out messy signals
  • Even unsupervised tasks benefit from careful feature shaping

I used to think ML was about fancy algorithms. Now I think it’s about thoughtful representation.

What I Want to Do Next

  • Finish building a preprocessing pipeline for the plate dataset
  • Apply LabelEncoder or OneHotEncoder where appropriate
  • Use correlation matrices to filter out redundant features
  • Build a basic classification model (e.g. RandomForest) with these features
  • Practice on another public dataset and compare modeling results with and without preprocessing

Machine learning didn’t “click” for me because of math — it clicked because of feature engineering.
And now, I finally understand why preprocessing is half the work.

You May Also Enjoy