PizzaBoi

The least German 🇩🇪 German Kaggle competition master
324 kaggle competition rating

[PizzaBoi @vonRudenberg]

A Kaggle-data paper published earlier this year found a counterintuitive cost to merging teams early: less variety in the final solution. It helps to think of a competition as two phases: independent search, then joint search. https://t.co/waFlY2XuwT

[PizzaBoi @vonRudenberg]

One last thing: don’t share the details of this discussion. Kaggle’s rules are flexible, and staff could easily interpret this as private sharing. The same principles apply when looking for co-founders for your new lab.

[PizzaBoi @vonRudenberg]

Who should you merge with? A Kaggle merge is a game of poker in the dark: you don’t want to reveal your ideas in case they’re stolen, but you do want to learn as much as possible about a potential teammate’s solution.

[PizzaBoi @vonRudenberg]

For serious competitors: ask about vacation plans. Merging with people who’ve already ruled out a final push is genuinely unwise.

[PizzaBoi @vonRudenberg]

Don’t merge if you have 4 team golds, have no strong partners left, have endless compute and time, or the competition is exactly your niche. Or if interns do the work while you keep the ideas. Kaggle allows it. 🍌 for each intern who wins you a medal and gets a good review.

[PizzaBoi @vonRudenberg]

A Kaggle-data paper published earlier this year found a counterintuitive cost to merging teams early: less variety in the final solution. It helps to think of a competition as two phases: independent search, then joint search. https://t.co/waFlY2XuwT

[PizzaBoi @vonRudenberg]

Merge late if you don’t know teammates well: wait for proven leaderboard results (but beware leaderboard overfit). Or merge when ideas stall near the deadline. For RSNA, we teamed up with shin0000 and yuto083, then combined both teams’ best ideas into one model.