![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F263583%2F607f09d46bb4359b03899b6b1bf0699b%2Fmeme_kaggle.png?generation=1716054234839285&alt=media\")
As an outsider I'm quite confused, why couldn't the metric just be a weighted sum of weekly score with more weight toward the later weeks or something like that 🤔
", "rawMarkdown": "As an outsider I'm quite confused, why couldn't the metric just be a weighted sum of weekly score with more weight toward the later weeks or something like that 🤔", "votes": null }, { "id": "2822696", "postDate": "05/18/2024 17:55:27", "content": "pro tip: change 0.07 to 0.073 to get more damage output
", "rawMarkdown": "pro tip: change 0.07 to 0.073 to get more damage output", "votes": null }, { "id": "2822728", "postDate": "05/18/2024 18:20:51", "content": "Because the organizer is confident their \"metric\" is so powerful that is much better than your proposal. Your idea, and many others were proposed to them, they rejected them all 😀
\nNow pay the price.
", "rawMarkdown": "Because the organizer is **confident** their \"metric\" is so powerful that is much better than your proposal. Your idea, and many others were proposed to them, they rejected them all 😀\n\nNow pay the price.", "votes": null }, { "id": "2822830", "postDate": "05/18/2024 19:39:56", "content": "This is funny, and sad, and true at the same time. However, the (sub.loc[con, 'score']-0.07).clip(0) (or the -0.05 variation) doesn't help all ensembles… in my case my best scoring ensemble saw a 0.02 score drop after implementing this. This makes me wonder how much of this tremendous score bump observed in the \"This is the way\" various notebooks is simply overfitting to the public dataset. I would not want to rely just on this notebook for sure.
I guess it matters a lot your model calibrated probability, this \"trick\" relies on that, and would require tuning, also, sometimes, the model could pick up things by mistake that artificially try to reproduce this, hence, not bringing new value.
\nEither way, its a waste of time, come for the learnings, ignore the leaderboard.
", "rawMarkdown": "I guess it matters a lot your model calibrated probability, this \"trick\" relies on that, and would require tuning, also, sometimes, the model could pick up things by mistake that artificially try to reproduce this, hence, not bringing new value.\n\nEither way, its a waste of time, come for the learnings, ignore the leaderboard.", "votes": null }, { "id": "2822922", "postDate": "05/18/2024 21:02:06", "content": "I think it's because your model is doing \"better\" on \"stability\", so would not be penalized that much on the falling rate. Applying the same parameter will not do much on improving your score but more damage to your average ginis. In short, you need to test your optimal parameter if you want to increase public score with the hack. But I highly doubt the optimal parameter would also be the one on private dataset.
", "rawMarkdown": "I think it's because your model is doing \"better\" on \"stability\", so would not be penalized that much on the falling rate. Applying the same parameter will not do much on improving your score but more damage to your average ginis. In short, you need to test your optimal parameter if you want to increase public score with the hack. But I highly doubt the optimal parameter would also be the one on private dataset.", "votes": null }, { "id": "2823643", "postDate": "05/19/2024 10:37:37", "content": "I think there will be a big shake with the full dataset,these hackings may be useless for the rest data.
", "rawMarkdown": "I think there will be a big shake with the full dataset,these hackings may be useless for the rest data.", "votes": null }, { "id": "2824126", "postDate": "05/19/2024 15:46:39", "content": "![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1098475%2F3529784a960c860aed58ae2c65a33cc7%2FScreenshot%202024-05-19%20at%2018.45.55.png?generation=1716133584245115&alt=media\")
As an outsider I'm quite confused, why couldn't the metric just be a weighted sum of weekly score with more weight toward the later weeks or something like that 🤔
", "votes": null, "replies": [ { "id": 2822728, "author_name": "carloshuertas", "author_url": "", "post_date": "05/18/2024 18:20:51", "content": "Because the organizer is confident their \"metric\" is so powerful that is much better than your proposal. Your idea, and many others were proposed to them, they rejected them all 😀
\nNow pay the price.
", "votes": null, "replies": [] } ] }, { "id": 2822696, "author_name": "shyhandsome", "author_url": "", "post_date": "05/18/2024 17:55:27", "content": "pro tip: change 0.07 to 0.073 to get more damage output
", "votes": null, "replies": [] }, { "id": 2822830, "author_name": "eduardastefanescu", "author_url": "", "post_date": "05/18/2024 19:39:56", "content": "This is funny, and sad, and true at the same time. However, the (sub.loc[con, 'score']-0.07).clip(0) (or the -0.05 variation) doesn't help all ensembles… in my case my best scoring ensemble saw a 0.02 score drop after implementing this. This makes me wonder how much of this tremendous score bump observed in the \"This is the way\" various notebooks is simply overfitting to the public dataset. I would not want to rely just on this notebook for sure.
I guess it matters a lot your model calibrated probability, this \"trick\" relies on that, and would require tuning, also, sometimes, the model could pick up things by mistake that artificially try to reproduce this, hence, not bringing new value.
\nEither way, its a waste of time, come for the learnings, ignore the leaderboard.
", "votes": null, "replies": [] }, { "id": 2822922, "author_name": "shyhandsome", "author_url": "", "post_date": "05/18/2024 21:02:06", "content": "I think it's because your model is doing \"better\" on \"stability\", so would not be penalized that much on the falling rate. Applying the same parameter will not do much on improving your score but more damage to your average ginis. In short, you need to test your optimal parameter if you want to increase public score with the hack. But I highly doubt the optimal parameter would also be the one on private dataset.
", "votes": null, "replies": [] } ] }, { "id": 2823643, "author_name": "zzhisthebest", "author_url": "", "post_date": "05/19/2024 10:37:37", "content": "I think there will be a big shake with the full dataset,these hackings may be useless for the rest data.
", "votes": null, "replies": [] }, { "id": 2824126, "author_name": "snnclsr", "author_url": "", "post_date": "05/19/2024 15:46:39", "content": "