AIOZ AI Housing Prices Challenge: Rules, Scoring, Submission, Best Practice
Success in the Housing Prices Challenge goes beyond clever models — it's about playing fair, submitting correctly, and avoiding common traps like data leakage.
This guide breaks down the rules, evaluation mechanics, submission specs, and a leakage playbook to ensure your offline experiments align with leaderboard results.
With multiple daily entries (up to 5 per day), it's essential to focus on thoughtful iterations: Train on labeled data, make blind predictions, and validate rigorously.
Let's show you how to rise to the top ranks smoothly and ethically.
Evaluation: How Your Model Gets Scored
Your predictions are evaluated using Root Mean Squared Error (RMSE) on log-transformed prices (both predicted and actual). This metric rewards precision across scales — nailing a $50K apartment counts as much as a $5M estate.
Lower RMSE wins; it's straightforward and rewards robust variance handling.
- Why Log-Transform? It normalizes skewed prices, preventing high-end outliers from dominating.
- Offline tip: Mirror this in your code — log prices before RMSE calculation for accurate local evaluations.
Scores update on the Public Leaderboard after each submission, fueling community benchmarking.
Submission Guidelines: Get It Right the First Time
Keep it simple: A CSV with a header and exactly two columns.
The platform auto-scores, and mismatches (e.g., wrong IDs) result in a zero score.
- house_index: use exact matches from test.csv (69,649 entries).
- price: Submit prices in the original scale; the platform applies the log transform for RMSE.
- Upload via the AIOZ AI platform; expect near‑instant leaderboard feedback.
Rules Recap:
- One account per participant — fair play only.
- Share code and datasets only in public forums; avoid private exchanges.
- Maximum of 5 submissions per day — quality over quantity.
We're building a collaborative space — lean on discussions for tips, not secrets.
Leakage Playbook: Safeguard Your Scores
Leakage can inflate offline results and tank live performance—typically by pulling “future” information or mixing pipelines. Stay clean with these:
- Identifier Leakage: Use house_index only for alignment; exclude it from your features to avoid ID overfitting.
- Incomputable Features: Derive inputs from fields available in both train and test; avoid statistics computed only on train (e.g., mean price by city from train alone).
- Train-Test Contamination: Fit on train.csv exclusively; keep test.csv blind; do not access ground_truth.csv
- Target Mismatch: Model "final sale price" — use purpose to filter sales, avoiding rent dilution.
- Transform Mismatch: Submit raw prices; log offline to sync with scoring.
Pro tip: Split train.csv 80/20 for internal train/validation — treat validation like test for realism.
Quality Checklist: Pre-Submission Audit
Tick these for leaderboard-ready work:
- Train/validate on train.csv (with price); test.csv unlabeled only.
- Exclude house_index from the feature set.
- Apply identical preprocessing to train and test.
- Ensure the target corresponds to the final sale price (e.g., select sale via purpose).
- CSV format: header house_index, price; IDs matching test.csv; original‑scale prices.
- Offline RMSE: On log‑transformed prices for true comparisons.
Quick 5 Key Facts
- Rows: 69,649 train, 29,850 test — diverse and doable.
- Target: Final sale price (present only in train.csv).
- Scoring: RMSE on log‑transformed prices — scale-agnostic.
- Purpose: Sale/rent flag — vital for purity.
- Submit: Original prices in house_index, price CSV.
Final Thoughts
Mastering rules and submissions turns potential pitfalls into power moves. Respect boundaries, validate smartly, and collaborate openly — this challenge is about shared growth in AI for real impact.
Submit your first model today, climb the ranks, and watch your predictions shape housing wisdom.
Join in at: aiozai.network/challenges — your clean run starts now!