Workspace rental · price tier prediction
The competition, step by step
How to approach this Kaggle project from scratch: what to look at, how to validate, which models to try, and why. Every lesson shows the explanation next to the exact code and its real output.
Roadmap
The whole project in 8 steps, mapped to the viva checklist.
WHYThe story: from 0.70
Why CV said 0.758 but Kaggle said 0.70, and how we fixed it.
00The first model
The original notebook, annotated: what's good, and the one hidden flaw.
01Setup & data
Load the data, rebuild the operator id, make operator-grouped folds.
02Exploratory data analysis
Target, metros, size, terms, and the operator split that explains everything.
03Preprocessing & features
Encodings, missing values, ratios, geo, operator profile, out-of-fold target features.
04Feature experiments
Two plausible feature ideas, measured honestly. Neither helped.
05Models
Logistic Regression baseline vs XGBoost, XGBoost regression, CatBoost. Pros, cons, importance.
06Operator smoothing
A level-2 stacker that uses the predictions of each operator's other listings.
07Scenarios
Blends × decoders (class weights, cut-points, per-metro), cross-fitted on grouped CV.
08Submission
The best scenario is re-tuned on all data and written to submission.csv.
09Error analysis
Confusion matrix, per-tier and per-metro scores: where the model still struggles.
10What else could be tried
Next ideas, and what's not allowed.
v3v3 · the Colab run
The full run with real outputs: grouped CV 0.7214, the submitted model.
VIVAViva prep
The 5 components of each model, key concepts, live-coding practice.