{ "id": 617814, "title": "About 5 fold", "url": "/competitions/csiro-biomass/discussion/617814", "author_name": "", "post_date": "2025-11-12T07:26:54.269687200Z", "votes": 2, "comment_count": 4, "views": 0, "content": "

I achieved an LB score of 0.57 when making predictions using the weights from a single-fold training. However, the LB score dropped to 0.52 instead when I performed 5-fold ensemble.Why is the 5-fold ensemble performing worse than a single fold? It doesn’t make sense at all!

", "messages": [ { "id": "3320029", "postDate": "11/12/2025 07:26:54", "content": "

I achieved an LB score of 0.57 when making predictions using the weights from a single-fold training. However, the LB score dropped to 0.52 instead when I performed 5-fold ensemble.Why is the 5-fold ensemble performing worse than a single fold? It doesn’t make sense at all!

", "rawMarkdown": "I achieved an LB score of 0.57 when making predictions using the weights from a single-fold training. However, the LB score dropped to 0.52 instead when I performed 5-fold ensemble.Why is the 5-fold ensemble performing worse than a single fold? It doesn’t make sense at all!", "votes": null }, { "id": "3320157", "postDate": "11/12/2025 09:28:29", "content": "

It does make sense.

\n

The single fold model was trained on a data distribution that's similar to the public LB test dataset (while the other fold models were less similar).

\n

that's why you're getting such a huge jump

\n

But to remain safe I would advise you to use the 5-fold model as your final solution instead of that single-fold model because the distribution might be different in the private test dataset

", "rawMarkdown": "It does make sense.\n\nThe single fold model was trained on a data distribution that's similar to the public LB test dataset (while the other fold models were less similar).\n\nthat's why you're getting such a huge jump\n\nBut to remain safe I would advise you to use the 5-fold model as your final solution instead of that single-fold model because the distribution might be different in the private test dataset", "votes": null }, { "id": "3320188", "postDate": "11/12/2025 10:00:26", "content": "

thank you for your helpful advise.the key is how to fold the train set?

", "rawMarkdown": "thank you for your helpful advise.the key is how to fold the train set?", "votes": null }, { "id": "3320203", "postDate": "11/12/2025 10:17:56", "content": "

As seen in other discussion and from my own experience, K-Fold with group by sampling_date and stratify by state works the best. The CV-LB gap is much smaller with this split and we can see some correlation too!

", "rawMarkdown": "As seen in other discussion and from my own experience, K-Fold with group by `sampling_date` and stratify by `state` works the best. The CV-LB gap is much smaller with this split and we can see some correlation too!", "votes": null }, { "id": "3320352", "postDate": "11/12/2025 13:12:30", "content": "

I appreciate your genuine response. I prepare to try this method.how well it works

", "rawMarkdown": "I appreciate your genuine response. I prepare to try this method.how well it works", "votes": null } ], "comments": [ { "id": 3320157, "author_name": "mushfirat", "author_url": "", "post_date": "11/12/2025 09:28:29", "content": "

It does make sense.

\n

The single fold model was trained on a data distribution that's similar to the public LB test dataset (while the other fold models were less similar).

\n

that's why you're getting such a huge jump

\n

But to remain safe I would advise you to use the 5-fold model as your final solution instead of that single-fold model because the distribution might be different in the private test dataset

", "votes": null, "replies": [ { "id": 3320188, "author_name": "", "author_url": "", "post_date": "11/12/2025 10:00:26", "content": "

thank you for your helpful advise.the key is how to fold the train set?

", "votes": null, "replies": [ { "id": 3320203, "author_name": "mushfirat", "author_url": "", "post_date": "11/12/2025 10:17:56", "content": "

As seen in other discussion and from my own experience, K-Fold with group by sampling_date and stratify by state works the best. The CV-LB gap is much smaller with this split and we can see some correlation too!

", "votes": null, "replies": [ { "id": 3320352, "author_name": "", "author_url": "", "post_date": "11/12/2025 13:12:30", "content": "

I appreciate your genuine response. I prepare to try this method.how well it works

", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "3320029": "I achieved an LB score of 0.57 when making predictions using the weights from a single-fold training. However, the LB score dropped to 0.52 instead when I performed 5-fold ensemble.Why is the 5-fold ensemble performing worse than a single fold? It doesn’t make sense at all!", "3320157": "It does make sense.\n\nThe single fold model was trained on a data distribution that's similar to the public LB test dataset (while the other fold models were less similar).\n\nthat's why you're getting such a huge jump\n\nBut to remain safe I would advise you to use the 5-fold model as your final solution instead of that single-fold model because the distribution might be different in the private test dataset", "3320188": "thank you for your helpful advise.the key is how to fold the train set?", "3320203": "As seen in other discussion and from my own experience, K-Fold with group by `sampling_date` and stratify by `state` works the best. The CV-LB gap is much smaller with this split and we can see some correlation too!", "3320352": "I appreciate your genuine response. I prepare to try this method.how well it works" }, "source": "meta" }