![\"shakeup\"](\"https://pbs.twimg.com/media/D2Q_xagXQAAjSiF.png\")
Hmm, interesting. I would take both ;)) Similar to blending, you know ;))
", "rawMarkdown": "Hmm, interesting. I would take both ;)) Similar to blending, you know ;))", "votes": null }, { "id": "496698", "postDate": "03/22/2019 13:33:19", "content": "Haha ... Actually learn a trick from you!
", "rawMarkdown": "Haha ... Actually learn a trick from you!", "votes": null }, { "id": "496705", "postDate": "03/22/2019 13:42:35", "content": "There is a reason why we are allowed two submissions. ;)
", "rawMarkdown": "There is a reason why we are allowed two submissions. ;)", "votes": null }, { "id": "496753", "postDate": "03/22/2019 14:48:31", "content": "Haha, interesting! What if both of them perform bad on private LB?
\n\nThe second one in table, is neither 'Best Public LB' nor 'Best Local CV'. But it is good on private LB. \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85205
", "rawMarkdown": "Haha, interesting! What if both of them perform bad on private LB? \n\nThe second one in table, is neither 'Best Public LB' nor 'Best Local CV'. But it is good on private LB. \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85205", "votes": null }, { "id": "496757", "postDate": "03/22/2019 14:52:16", "content": "Huh not this one.\nThe same public kernel run twice gives me top-60 or top-1100 score. What do we suppose from competitions like this?
", "rawMarkdown": "Huh not this one.\nThe same public kernel run twice gives me top-60 or top-1100 score. What do we suppose from competitions like this?", "votes": null }, { "id": "496758", "postDate": "03/22/2019 14:55:26", "content": "Apparently you were supposed to average over different seeds and different K-fold splits. :)
", "rawMarkdown": "Apparently you were supposed to average over different seeds and different K-fold splits. :)", "votes": null }, { "id": "496776", "postDate": "03/22/2019 15:12:51", "content": "Generalization and uncertainty in modeling are crucial to make anything useful in real world. Shakeups are needed to bring someone in bottom to top and those at top go down just based on generalization. I have noticed that some top notch Kagglers may hold position in prize range but they try to be in gold and once they are in gold they spend most of their time in generalization of model when shakeup occurs these top Kagglers are hardly affected. I have recently started to look into this domain I guess Kaggle changes the distribution of test and train and more the difference in distributions is, lower the rank goes. one way to measure this difference among probability density function or probability distribution function is KL divergence https://en.wikipedia.org/wiki/Kullback–Leibler_divergence
", "rawMarkdown": "Generalization and uncertainty in modeling are crucial to make anything useful in real world. Shakeups are needed to bring someone in bottom to top and those at top go down just based on generalization. I have noticed that some top notch Kagglers may hold position in prize range but they try to be in gold and once they are in gold they spend most of their time in generalization of model when shakeup occurs these top Kagglers are hardly affected. I have recently started to look into this domain I guess Kaggle changes the distribution of test and train and more the difference in distributions is, lower the rank goes. one way to measure this difference among probability density function or probability distribution function is KL divergence https://en.wikipedia.org/wiki/Kullback–Leibler_divergence", "votes": null }, { "id": "497047", "postDate": "03/22/2019 22:20:57", "content": "Good one @tunguz. Did you also have a tree based model with a measly public score that did better than your final selections on the private LB? I think most people had.
", "rawMarkdown": "Good one @tunguz. Did you also have a [tree based model with a measly public score](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85146#496224) that did better than your final selections on the private LB? I think most people had.", "votes": null }, { "id": "497059", "postDate": "03/22/2019 22:41:14", "content": "Hmm! @tunguz, I openned a discussion thread to brainstorm why so many of us got caught up in the shakeup but have not got much attention. Will you care to add your 2 cents there?
", "rawMarkdown": "Hmm! @tunguz, I [openned a discussion thread to brainstorm](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85143) why so many of us got caught up in the shakeup but have not got much attention. Will you care to add your 2 cents [there](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85143)?", "votes": null }, { "id": "497183", "postDate": "03/23/2019 04:26:51", "content": "Yeah, that happened to me. I had a pure xgboost submission with a Private score of 0.67240, but a Public score of only 0.48401. Had I selected it I would have finished 23rd instead of 47th. But I had no reason to believe it would do so well, and I was lucky to finish where I did.
", "rawMarkdown": "Yeah, that happened to me. I had a pure xgboost submission with a Private score of 0.67240, but a Public score of only 0.48401. Had I selected it I would have finished 23rd instead of 47th. But I had no reason to believe it would do so well, and I was lucky to finish where I did.", "votes": null }, { "id": "497185", "postDate": "03/23/2019 04:34:03", "content": "Thanks @dslate for sharing and congrats you faired well considering. Reading though all the discussions, my conclusion is that tree based models performed really well on this data. The problem is the lack of very good local CV setup.
", "rawMarkdown": "Thanks @dslate for sharing and congrats you faired well considering. Reading though all the discussions, my conclusion is that tree based models performed really well on this data. The problem is the lack of very good local CV setup.", "votes": null }, { "id": "497330", "postDate": "03/23/2019 11:40:11", "content": "I actually “decoupled” from this competition a while back, and have not paid much attention to it, so my 2c would be worth much less than anyone else’s.
", "rawMarkdown": "I actually “decoupled” from this competition a while back, and have not paid much attention to it, so my 2c would be worth much less than anyone else’s.", "votes": null }, { "id": "539399", "postDate": "05/30/2019 02:58:38", "content": "Lol True Story!!!
", "rawMarkdown": "Lol True Story!!!", "votes": null } ], "comments": [ { "id": 496695, "author_name": "sggpls", "author_url": "", "post_date": "03/22/2019 13:30:35", "content": "Hmm, interesting. I would take both ;)) Similar to blending, you know ;))
", "votes": null, "replies": [ { "id": 496698, "author_name": "ratthachat", "author_url": "", "post_date": "03/22/2019 13:33:19", "content": "Haha ... Actually learn a trick from you!
", "votes": null, "replies": [] }, { "id": 496705, "author_name": "tunguz", "author_url": "", "post_date": "03/22/2019 13:42:35", "content": "There is a reason why we are allowed two submissions. ;)
", "votes": null, "replies": [] } ] }, { "id": 496753, "author_name": "gitshe11", "author_url": "", "post_date": "03/22/2019 14:48:31", "content": "Haha, interesting! What if both of them perform bad on private LB?
\n\nThe second one in table, is neither 'Best Public LB' nor 'Best Local CV'. But it is good on private LB. \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85205
", "votes": null, "replies": [] }, { "id": 496757, "author_name": "yaroshevskiy", "author_url": "", "post_date": "03/22/2019 14:52:16", "content": "Huh not this one.\nThe same public kernel run twice gives me top-60 or top-1100 score. What do we suppose from competitions like this?
", "votes": null, "replies": [ { "id": 496758, "author_name": "tunguz", "author_url": "", "post_date": "03/22/2019 14:55:26", "content": "Apparently you were supposed to average over different seeds and different K-fold splits. :)
", "votes": null, "replies": [] }, { "id": 497059, "author_name": "sheriytm", "author_url": "", "post_date": "03/22/2019 22:41:14", "content": "Hmm! @tunguz, I openned a discussion thread to brainstorm why so many of us got caught up in the shakeup but have not got much attention. Will you care to add your 2 cents there?
", "votes": null, "replies": [] }, { "id": 497330, "author_name": "tunguz", "author_url": "", "post_date": "03/23/2019 11:40:11", "content": "I actually “decoupled” from this competition a while back, and have not paid much attention to it, so my 2c would be worth much less than anyone else’s.
", "votes": null, "replies": [] } ] }, { "id": 496776, "author_name": "cyberia", "author_url": "", "post_date": "03/22/2019 15:12:51", "content": "Generalization and uncertainty in modeling are crucial to make anything useful in real world. Shakeups are needed to bring someone in bottom to top and those at top go down just based on generalization. I have noticed that some top notch Kagglers may hold position in prize range but they try to be in gold and once they are in gold they spend most of their time in generalization of model when shakeup occurs these top Kagglers are hardly affected. I have recently started to look into this domain I guess Kaggle changes the distribution of test and train and more the difference in distributions is, lower the rank goes. one way to measure this difference among probability density function or probability distribution function is KL divergence https://en.wikipedia.org/wiki/Kullback–Leibler_divergence
", "votes": null, "replies": [] }, { "id": 497047, "author_name": "sheriytm", "author_url": "", "post_date": "03/22/2019 22:20:57", "content": "Good one @tunguz. Did you also have a tree based model with a measly public score that did better than your final selections on the private LB? I think most people had.
", "votes": null, "replies": [ { "id": 497183, "author_name": "dslate", "author_url": "", "post_date": "03/23/2019 04:26:51", "content": "Yeah, that happened to me. I had a pure xgboost submission with a Private score of 0.67240, but a Public score of only 0.48401. Had I selected it I would have finished 23rd instead of 47th. But I had no reason to believe it would do so well, and I was lucky to finish where I did.
", "votes": null, "replies": [] }, { "id": 497185, "author_name": "sheriytm", "author_url": "", "post_date": "03/23/2019 04:34:03", "content": "Thanks @dslate for sharing and congrats you faired well considering. Reading though all the discussions, my conclusion is that tree based models performed really well on this data. The problem is the lack of very good local CV setup.
", "votes": null, "replies": [] } ] }, { "id": 539399, "author_name": "akhileshrai", "author_url": "", "post_date": "05/30/2019 02:58:38", "content": "Lol True Story!!!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "496686": "", "496695": "Hmm, interesting. I would take both ;)) Similar to blending, you know ;))", "496698": "Haha ... Actually learn a trick from you!", "496705": "There is a reason why we are allowed two submissions. ;)", "496753": "Haha, interesting! What if both of them perform bad on private LB? \n\nThe second one in table, is neither 'Best Public LB' nor 'Best Local CV'. But it is good on private LB. \nhttps://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85205", "496757": "Huh not this one.\nThe same public kernel run twice gives me top-60 or top-1100 score. What do we suppose from competitions like this?", "496758": "Apparently you were supposed to average over different seeds and different K-fold splits. :)", "496776": "Generalization and uncertainty in modeling are crucial to make anything useful in real world. Shakeups are needed to bring someone in bottom to top and those at top go down just based on generalization. I have noticed that some top notch Kagglers may hold position in prize range but they try to be in gold and once they are in gold they spend most of their time in generalization of model when shakeup occurs these top Kagglers are hardly affected. I have recently started to look into this domain I guess Kaggle changes the distribution of test and train and more the difference in distributions is, lower the rank goes. one way to measure this difference among probability density function or probability distribution function is KL divergence https://en.wikipedia.org/wiki/Kullback–Leibler_divergence", "497047": "Good one @tunguz. Did you also have a [tree based model with a measly public score](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85146#496224) that did better than your final selections on the private LB? I think most people had.", "497059": "Hmm! @tunguz, I [openned a discussion thread to brainstorm](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85143) why so many of us got caught up in the shakeup but have not got much attention. Will you care to add your 2 cents [there](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/85143)?", "497183": "Yeah, that happened to me. I had a pure xgboost submission with a Private score of 0.67240, but a Public score of only 0.48401. Had I selected it I would have finished 23rd instead of 47th. But I had no reason to believe it would do so well, and I was lucky to finish where I did.", "497185": "Thanks @dslate for sharing and congrats you faired well considering. Reading though all the discussions, my conclusion is that tree based models performed really well on this data. The problem is the lack of very good local CV setup.", "497330": "I actually “decoupled” from this competition a while back, and have not paid much attention to it, so my 2c would be worth much less than anyone else’s.", "539399": "Lol True Story!!!" }, "source": "meta" }