I'm analyzing the recent jump in Public LB scores and it seems largely driven by post-processing rules like setting Clover to 0. While this works for the Public Test set, I doubt that the Private Test set has absolutely zero clover samples. This feels like a huge risk for a shake-up.
\nFor those aiming for a gold/silver medal, are you incorporating these public notebook strategies, or are you focusing purely on improving your model's generalization capability? I'd love to hear opinions from experienced Kagglers.
", "messages": [ { "id": "3393494", "postDate": "01/19/2026 06:28:50", "content": "I'm analyzing the recent jump in Public LB scores and it seems largely driven by post-processing rules like setting Clover to 0. While this works for the Public Test set, I doubt that the Private Test set has absolutely zero clover samples. This feels like a huge risk for a shake-up.
\nFor those aiming for a gold/silver medal, are you incorporating these public notebook strategies, or are you focusing purely on improving your model's generalization capability? I'd love to hear opinions from experienced Kagglers.
", "rawMarkdown": "I'm analyzing the recent jump in Public LB scores and it seems largely driven by post-processing rules like setting Clover to 0. While this works for the Public Test set, I doubt that the Private Test set has absolutely zero clover samples. This feels like a huge risk for a shake-up.\n\nFor those aiming for a gold/silver medal, are you incorporating these public notebook strategies, or are you focusing purely on improving your model's generalization capability? I'd love to hear opinions from experienced Kagglers.", "votes": null }, { "id": "3393518", "postDate": "01/19/2026 08:05:17", "content": "I prefer no post processing but that too isn't shake-proof. I would love if grandmasters/masters were a bit more active in sharing their experience. Like I understand no one wants to share their exact approach but atleast guide us through your experience with past competitions on tackling shakeups.
", "rawMarkdown": "I prefer no post processing but that too isn't shake-proof. I would love if grandmasters/masters were a bit more active in sharing their experience. Like I understand no one wants to share their exact approach but atleast guide us through your experience with past competitions on tackling shakeups.", "votes": null }, { "id": "3393523", "postDate": "01/19/2026 08:22:04", "content": "Me too. I also haven't incorporated any post processing so far. \nI would say, if one is deciding between using a post processing strategy vs not to, they should check how much improvement are they getting. If it changes the second decimal, then go for it, else not.
", "rawMarkdown": "Me too. I also haven't incorporated any post processing so far. \nI would say, if one is deciding between using a post processing strategy vs not to, they should check how much improvement are they getting. If it changes the second decimal, then go for it, else not.", "votes": null }, { "id": "3393606", "postDate": "01/19/2026 11:54:01", "content": "In conclusion, since you can select two Subs, you can choose one that is intentionally biased towards the target, or one that is not.
\nAs an extremely bad example, at 『NeurIPS - Open Polymer Prediction 2025』, which ended four months ago,\nA tragic incident occurred where anyone could win a Gold Medal simply by multiplying the inference result by 2: \"{PredictionResult}*2\".\nThis was because a domain shift occurred in this competition, where public was celsius and private was fahrenheit.
\nhttps://www.kaggle.com/competitions/neurips-open-polymer-prediction-2025/discussion/607769#3289402
\nIdeally, the machine learning problem setting and data division would be appropriate for Train/PublicLB/PrivateLB, but there may be competitions where this is not the case.
", "rawMarkdown": "In conclusion, since you can select two Subs, you can choose one that is intentionally biased towards the target, or one that is not.\n\nAs an extremely bad example, at 『NeurIPS - Open Polymer Prediction 2025』, which ended four months ago,\nA tragic incident occurred where anyone could win a **Gold Medal simply by multiplying the inference result by 2: \"{PredictionResult}*2\".**\nThis was because a domain shift occurred in this competition, where public was celsius and private was fahrenheit.\n\nhttps://www.kaggle.com/competitions/neurips-open-polymer-prediction-2025/discussion/607769#3289402\n\nIdeally, the machine learning problem setting and data division would be appropriate for Train/PublicLB/PrivateLB, but there may be competitions where this is not the case.", "votes": null }, { "id": "3393627", "postDate": "01/19/2026 12:52:32", "content": "Oh I remember the angry downvotes, lol! But the host later clarified it was just a distribution shift where everyone's models were not able to extrapolate that well (basically not very good at predicting higher values)
", "rawMarkdown": "Oh I remember the angry downvotes, lol! But the host later clarified it was just a distribution shift where everyone's models were not able to extrapolate that well (basically not very good at predicting higher values)", "votes": null }, { "id": "3395597", "postDate": "01/23/2026 09:38:57", "content": "I don't think there will be a big shakeup, from my personal perspective, the score on the public leaderboard for this competition is too difficult to improve, and the sample size is also considerable (800 * 53% = 424).
", "rawMarkdown": "I don't think there will be a big shakeup, from my personal perspective, the score on the public leaderboard for this competition is too difficult to improve, and the sample size is also considerable (800 * 53% = 424).", "votes": null }, { "id": "3395692", "postDate": "01/23/2026 13:50:01", "content": "I am working on cv mostly, but some strategies seem to be also consistant with the public leaderboard
", "rawMarkdown": "I am working on cv mostly, but some strategies seem to be also consistant with the public leaderboard", "votes": null } ], "comments": [ { "id": 3393518, "author_name": "paritoshtripathi5", "author_url": "", "post_date": "01/19/2026 08:05:17", "content": "I prefer no post processing but that too isn't shake-proof. I would love if grandmasters/masters were a bit more active in sharing their experience. Like I understand no one wants to share their exact approach but atleast guide us through your experience with past competitions on tackling shakeups.
", "votes": null, "replies": [ { "id": 3393523, "author_name": "komilparmar", "author_url": "", "post_date": "01/19/2026 08:22:04", "content": "Me too. I also haven't incorporated any post processing so far. \nI would say, if one is deciding between using a post processing strategy vs not to, they should check how much improvement are they getting. If it changes the second decimal, then go for it, else not.
", "votes": null, "replies": [] } ] }, { "id": 3393606, "author_name": "hideyukizushi", "author_url": "", "post_date": "01/19/2026 11:54:01", "content": "In conclusion, since you can select two Subs, you can choose one that is intentionally biased towards the target, or one that is not.
\nAs an extremely bad example, at 『NeurIPS - Open Polymer Prediction 2025』, which ended four months ago,\nA tragic incident occurred where anyone could win a Gold Medal simply by multiplying the inference result by 2: \"{PredictionResult}*2\".\nThis was because a domain shift occurred in this competition, where public was celsius and private was fahrenheit.
\nhttps://www.kaggle.com/competitions/neurips-open-polymer-prediction-2025/discussion/607769#3289402
\nIdeally, the machine learning problem setting and data division would be appropriate for Train/PublicLB/PrivateLB, but there may be competitions where this is not the case.
", "votes": null, "replies": [ { "id": 3393627, "author_name": "paritoshtripathi5", "author_url": "", "post_date": "01/19/2026 12:52:32", "content": "Oh I remember the angry downvotes, lol! But the host later clarified it was just a distribution shift where everyone's models were not able to extrapolate that well (basically not very good at predicting higher values)
", "votes": null, "replies": [] } ] }, { "id": 3395597, "author_name": "jiazhuang", "author_url": "", "post_date": "01/23/2026 09:38:57", "content": "I don't think there will be a big shakeup, from my personal perspective, the score on the public leaderboard for this competition is too difficult to improve, and the sample size is also considerable (800 * 53% = 424).
", "votes": null, "replies": [] }, { "id": 3395692, "author_name": "litaihao", "author_url": "", "post_date": "01/23/2026 13:50:01", "content": "I am working on cv mostly, but some strategies seem to be also consistant with the public leaderboard
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3393494": "I'm analyzing the recent jump in Public LB scores and it seems largely driven by post-processing rules like setting Clover to 0. While this works for the Public Test set, I doubt that the Private Test set has absolutely zero clover samples. This feels like a huge risk for a shake-up.\n\nFor those aiming for a gold/silver medal, are you incorporating these public notebook strategies, or are you focusing purely on improving your model's generalization capability? I'd love to hear opinions from experienced Kagglers.", "3393518": "I prefer no post processing but that too isn't shake-proof. I would love if grandmasters/masters were a bit more active in sharing their experience. Like I understand no one wants to share their exact approach but atleast guide us through your experience with past competitions on tackling shakeups.", "3393523": "Me too. I also haven't incorporated any post processing so far. \nI would say, if one is deciding between using a post processing strategy vs not to, they should check how much improvement are they getting. If it changes the second decimal, then go for it, else not.", "3393606": "In conclusion, since you can select two Subs, you can choose one that is intentionally biased towards the target, or one that is not.\n\nAs an extremely bad example, at 『NeurIPS - Open Polymer Prediction 2025』, which ended four months ago,\nA tragic incident occurred where anyone could win a **Gold Medal simply by multiplying the inference result by 2: \"{PredictionResult}*2\".**\nThis was because a domain shift occurred in this competition, where public was celsius and private was fahrenheit.\n\nhttps://www.kaggle.com/competitions/neurips-open-polymer-prediction-2025/discussion/607769#3289402\n\nIdeally, the machine learning problem setting and data division would be appropriate for Train/PublicLB/PrivateLB, but there may be competitions where this is not the case.", "3393627": "Oh I remember the angry downvotes, lol! But the host later clarified it was just a distribution shift where everyone's models were not able to extrapolate that well (basically not very good at predicting higher values)", "3395597": "I don't think there will be a big shakeup, from my personal perspective, the score on the public leaderboard for this competition is too difficult to improve, and the sample size is also considerable (800 * 53% = 424).", "3395692": "I am working on cv mostly, but some strategies seem to be also consistant with the public leaderboard" }, "source": "meta" }