Lesson 15: Challenge 60 minutes — Advanced House Prices

Machine Learning: From Basics to Advanced

Part 2: Industry standard workflow

xdev.asia

Introduction

This is the first combined challenge. You will return to the real estate problem but do it at a more mature level: pipeline, missing handling, feature engineering, cross-validation and tuning. The goal is to combine discrete skills into a complete workflow.

Lesson objectives

• Complete a relatively complete regression challenge.
• Combine preprocessing, pipeline and tuning in the same flow.
• Write a clear summary like a mini technical report.

Request challenge

You need to submit 3 outputs: a notebook or training script, a file describing the technical decision, and a table comparing at least 3 model experiments.

Suggested directions

1. Start with a simple baseline.
2. Check for missing values ​​and data types.
3. Create a standard pipeline.
4. Run cross-validation.
5. Tuning few important parameters.
6. Analyze the best model and outstanding errors.

Questions you must answer for yourself

• Which feature is the most important?
• Which group of houses does the biggest error fall into?
• Does adding feature engineering really improve it?
• Is the current model reliable enough to use as a preliminary estimate?

Common mistakes

• Write a long notebook but without a clear conclusion.
• Tuning a lot but not saving old results.
• Evaluate based on feeling without using consistent metrics.

Practice exercises

• Do the challenge as an independent submission.
• Create a test table with columns: model, preprocessing, CV score, notes.
• Write a lesson learned section about 10 lines long.

Completion criteria

• Have a complete pipeline that runs end-to-end.
• There are structured experimental comparisons.
• There is a clear conclusion as to why the final model was chosen.

Practice step by step (advanced)

1. Time-box 60 minutes in three phases: preparation, modeling, summary.
2. Design at least 3 experiments with clear differences.
3. Fully write down the assumptions for each experiment.
4. Select the final model using metric + stability criteria.
5. Post-mortem writing: if you had 2 more hours what would you do?

Artifact should be submitted

• Experiment tracking table has at least 3 lines.
• Clean notebook, runs from start to finish.
• 1-page summary in technical + business format.

Self-test questions

• Have you really controlled leakage?
• Which tuning brings the greatest benefit?
• Are the current results enough to ship the beta version?