Kaggle Grandmaster Panel - H2O AI World London 2018
- Generative AI
- H2O Driverless AI
Overview
A panel of Kaggle Grandmasters, several of them H2O.ai data scientists, fields audience questions at H2O World London 2018 on how they actually work competitions: setting up validation before touching a model, how many hours a week they put in, favorite tools, and how to spot leakage. The conversation also covers whether Kaggle skills transfer to real production work and how panelists use Driverless AI alongside manual feature engineering.
Key takeaways
- Why nearly every panelist starts a competition by building a solid cross-validation setup before any serious modeling, and why a validation score that doesn't track the leaderboard is the first red flag
- How much time top Kagglers actually spend, ranging from a couple hours a week to eight-to-ten hours a day during an active competition
- Favorite tools across the panel: XGBoost, LightGBM, PyTorch, Keras/TensorFlow, and H2O's own Datatable for fast feature engineering
- Why data leakage found on Kaggle mirrors leakage that shows up in real production models, and why learning to spot it is a transferable skill
- How panelists use Driverless AI to get a fast baseline and feature-importance read before manually engineering features, while noting it can't be used for final Kaggle submissions
- Practical competition habits: not joining a competition too early, using both final submission slots for different ideas, and treating validation setup as the most important skill
Start owning your AI