MLP project: heavy equipment price prediction
My Machine Learning Practice project from the IIT Madras BS in Data Science. A Kaggle regression problem: predict the auction price of used construction machines. Metric RMSLE, best leaderboard score 0.18223.
Notebook
- NotebookFinal project notebookEDA, feature engineering, 6 models compared, tuning, and the final 4-model ensemble
Guidance
- PageKaggle journeyHow the score went from 0.205 to 0.182, and what didn't work
- PageProject guideThe order I'd do things in: EDA, features, encoding, leakage, models, blending
Five things I learned
- Read the metric first. RMSLE means training on
log1p(target), and that shaped every later decision. - Trust the leaderboard over local CV when they disagree. Two of my best local results did worse on Kaggle.
- Seed bagging is the cheapest reliable gain. Heavy hyperparameter tuning gave almost nothing.
- Target encoding has to be out-of-fold and smoothed. Otherwise each row sees its own answer.
- Good features beat bigger models. The biggest jump came from features built on related sales.