![\"feature](\"https://storage.googleapis.com/dbi-share/kaggle/feature_selection.png\")
I started this competition because of it’s unique task. I liked the idea of „not overfitting“, very few data and it felt like some sort of „real world“ data science problem. Unfortunately I didn't have much time so I‘ve decided to wait and learn from others.
\n\nAs it turns out lb probing was key for winning which was a little disappointing (though I really liked @zachmayer & @cdeotte solutions and findings!)
\n\nNow I have more time again so I’ve decided to climb lb and build a model without probing. \nProps to @featureblind & @nightwolfbrooks - I used your models for some quick and dirty ensembling!\nYou can find my solution here link to kernel (feels like using a sledgehammer to crack a nut).
\n\nThe key part is feature selection. I used „shadow“/ NULL – feature importance to select the final feature set link to kernel.
\n\n
+Yes I know ... my code is really messy ;)\n+It is always easier to climb lb after competition ended :)\n+have fun!
", "comments": [ { "id": 531460, "author_name": "melondonkey", "author_url": "", "post_date": "2019-05-15T00:44:41.870000", "content": "I’ve said it before but it’s worth repeating: any solution modified after a submission to LB is “probing” in some sense, so it is not helpful to think in the binary sense of probing/no probing. In fact this solution even with its ensembling is built on other probed solutions. Not to take away from the other merits of your solution but labeling it a “no probe” solution is just not accurate.
", "votes": 5, "replies": [ { "id": 532301, "author_name": "Zach", "author_url": "", "post_date": "2019-05-16T15:54:30.577000", "content": "I completely agree.
\n\nThe original don't overfit had 3 targets. The 1st target was for practice, the second target was the public LB, and the 3rd target was the private LB.
\n\nThis meant that you only got a 5 submissions total on the private LB, which made probing almost impossible.
", "votes": 1, "replies": [] } ] } ], "comment_count": 2, "messages": [ { "id": 531460, "postDate": "2019-05-15T00:44:41.870Z", "content": "I’ve said it before but it’s worth repeating: any solution modified after a submission to LB is “probing” in some sense, so it is not helpful to think in the binary sense of probing/no probing. In fact this solution even with its ensembling is built on other probed solutions. Not to take away from the other merits of your solution but labeling it a “no probe” solution is just not accurate.
", "rawMarkdown": "I’ve said it before but it’s worth repeating: any solution modified after a submission to LB is “probing” in some sense, so it is not helpful to think in the binary sense of probing/no probing. In fact this solution even with its ensembling is built on other probed solutions. Not to take away from the other merits of your solution but labeling it a “no probe” solution is just not accurate.", "votes": 5, "replies": [ { "id": 532301, "postDate": "2019-05-16T15:54:30.577Z", "content": "I completely agree.
\n\nThe original don't overfit had 3 targets. The 1st target was for practice, the second target was the public LB, and the 3rd target was the private LB.
\n\nThis meant that you only got a 5 submissions total on the private LB, which made probing almost impossible.
", "rawMarkdown": "I completely agree.\n\nThe original don't overfit had 3 targets. The 1st target was for practice, the second target was the public LB, and the 3rd target was the private LB.\n\nThis meant that you only got a 5 submissions total on the private LB, which made probing almost impossible.", "votes": 1 } ] }, { "id": 531298, "postDate": "2019-05-14T16:04:06.503Z", "content": "I started this competition because of it’s unique task. I liked the idea of „not overfitting“, very few data and it felt like some sort of „real world“ data science problem. Unfortunately I didn't have much time so I‘ve decided to wait and learn from others.
\n\nAs it turns out lb probing was key for winning which was a little disappointing (though I really liked @zachmayer & @cdeotte solutions and findings!)
\n\nNow I have more time again so I’ve decided to climb lb and build a model without probing. \nProps to @featureblind & @nightwolfbrooks - I used your models for some quick and dirty ensembling!\nYou can find my solution here link to kernel (feels like using a sledgehammer to crack a nut).
\n\nThe key part is feature selection. I used „shadow“/ NULL – feature importance to select the final feature set link to kernel.
\n\n
+Yes I know ... my code is really messy ;)\n+It is always easier to climb lb after competition ended :)\n+have fun!
", "rawMarkdown": "I started this competition because of it’s unique task. I liked the idea of „not overfitting“, very few data and it felt like some sort of „real world“ data science problem. Unfortunately I didn't have much time so I‘ve decided to wait and learn from others. \n\nAs it turns out lb probing was key for winning which was a little disappointing (though I really liked @zachmayer & @cdeotte solutions and findings!)\n\nNow I have more time again so I’ve decided to climb lb and build a model without probing. \nProps to @featureblind & @nightwolfbrooks - I used your models for some quick and dirty ensembling!\nYou can find my solution here [link to kernel](https://www.kaggle.com/springmanndaniel/no-probing-autoenc-mlp-ensembled?scriptVersionId=14150758) **(feels like using a sledgehammer to crack a nut)**. \n\nThe key part is feature selection. I used „shadow“/ NULL – feature importance to select the final feature set [link to kernel]( https://www.kaggle.com/springmanndaniel/shadow-null-importance-feature-selection).\n\n\n\n____ \n+Yes I know ... my code is really messy ;)\n+It is always easier to climb lb after competition ended :)\n+have fun!", "votes": 3 } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:13:03Z", "competition_id": 12896, "competition_slug": "dont-overfit-ii", "team_id": 2821543, "solution_url": "/competitions/dont-overfit-ii/writeups/danzel-this-is-how-you-place-3rd-without-probing" }