![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3597887%2Fc2dd9cfe51d806442e7d6e75cbb7f8c5%2FCapture.PNG?generation=1597115447601918&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3597887%2F56e1fb8496df4d78e2ed4e0344708125%2F2.PNG?generation=1597115467267630&alt=media\")
\nwas from this notebook .\nIt is really annoying for the ones on the bottom of 0.9619 to loose 170 positions despite the hard work. Kaggle Team should find a way to solve this issue.99% of submissions from position 314 to 493 (179 positions ) \n\nwas from this notebook .\nIt is really annoying for the ones on the bottom of 0.9619 to loose 170 positions despite the hard work. Kaggle Team should find a way to solve this issue.
Watch them fall in private LB
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F75976%2Fbdf92ddd7b94fbf31736e1e6507ba8f8%2Fpopcorn.jpg?generation=1597164604080305&alt=media\")
99% of submissions from position 314 to 493 (179 positions ) \n\nwas from this notebook .\nIt is really annoying for the ones on the bottom of 0.9619 to loose 170 positions despite the hard work. Kaggle Team should find a way to solve this issue.
Like tom said, Don't believe it π
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1984321%2F8ae4abae81fb76fb00ab8495fe5c71e5%2Fuy.jpg?generation=1597267542976136&alt=media\")
Porto Seguro was among my first competitions on kaggle.
\nI left the competion about 1 month before the end, after submitting my best CV.
\nBut during the last month, many people were submitting a highly overfitting public blend kernel. But they fell on private LB and I jumped up + 800 on that private LB ^^
", "rawMarkdown": "[Porto Seguro](https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/leaderboard) was among my first competitions on kaggle. \n\nI left the competion about 1 month before the end, after submitting my best CV. \n\nBut during the last month, many people were submitting a highly overfitting public blend kernel. But they fell on private LB and I jumped up + 800 on that private LB ^^\n", "votes": 8, "replies": [ { "id": 966801, "postDate": "2020-08-11T17:06:23.853Z", "content": "Seems like Masters and Grand masters have a lot of such sweet stories! ^_^
", "rawMarkdown": "Seems like Masters and Grand masters have a lot of such sweet stories! ^_^" }, { "id": 966816, "postDate": "2020-08-11T17:17:25.110Z", "content": "Porto Seguro is also my prototypical examples as well for how dangerous blends of blends of blends of public notebooks can be. I moved from 1178 public LB to 29 private LB there. I would not mind if something similar happens in this competition. LOL.
", "rawMarkdown": "Porto Seguro is also my prototypical examples as well for how dangerous blends of blends of blends of public notebooks can be. I moved from 1178 public LB to 29 private LB there. I would not mind if something similar happens in this competition. LOL.", "votes": 15 }, { "id": 966825, "postDate": "2020-08-11T17:20:02.923Z", "content": "\nI was novive back then ;) And to be fair I was a bit discouraged and decided to leave the competition when I saw I dropped +100 places just few hours after the kernel was published.
\nIn retrospective , I think it was a good decision , because as novice, it's easy to be fooled by public LB and public best kernels score and I would likely take part of the blending game if I stayed.
", "rawMarkdown": "@sarques \n\nI was novive back then ;) And to be fair I was a bit discouraged and decided to leave the competition when I saw I dropped +100 places just few hours after the kernel was published. \n\nIn retrospective , I think it was a good decision , because as novice, it's easy to be fooled by public LB and public best kernels score and I would likely take part of the blending game if I stayed. ", "votes": 1 }, { "id": 966832, "postDate": "2020-08-11T17:25:20.120Z", "content": "\n\nI would not mind if something similar happens in this competition. LOL.
\n
May be, but 3 submissions choice may mitigate a bit the shake up ;)
", "rawMarkdown": "> I would not mind if something similar happens in this competition. LOL.\n\nMay be, but 3 submissions choice may mitigate a bit the shake up ;)", "votes": 1 }, { "id": 966841, "postDate": "2020-08-11T17:28:39.500Z", "content": "3 chances of overfitting instead of 2?
", "rawMarkdown": "3 chances of overfitting instead of 2?", "votes": 5, "replies": [ { "id": 967342, "postDate": "2020-08-12T07:10:56.573Z", "content": "The thing is with 3 subs, everyone (incl. top in public LB) can easily take a risk with their best \"over-fitted\" model. 1 sub - best over-fit and other 2 subs - good cv + indicative LB + intuition. Plus the variance in models and huge class imbalance - anything can happen on private LB.
", "rawMarkdown": "The thing is with 3 subs, everyone (incl. top in public LB) can easily take a risk with their best \"over-fitted\" model. 1 sub - best over-fit and other 2 subs - good cv + indicative LB + intuition. Plus the variance in models and huge class imbalance - anything can happen on private LB." }, { "id": 967677, "postDate": "2020-08-12T12:21:52.770Z", "content": "You're confirming that there will be more overfitting with 3 subs than with 2, thanks ;)
", "rawMarkdown": "You're confirming that there will be more overfitting with 3 subs than with 2, thanks ;)", "votes": 1 } ] }, { "id": 966861, "postDate": "2020-08-11T17:41:30.337Z", "content": "True @serigne sama, although I have participated only in 1 competition before this one(Recursion), I have this habit of reading the notebook, at least the comments before believing some blend. Similar happened here.
\nAlso, at a time, I was around 340 in this competition, after that Trap came in, I'm pushed to 600+. That's why I think it'll be fun to watch the shake up! :)
", "rawMarkdown": "True @serigne sama, although I have participated only in 1 competition before this one(Recursion), I have this habit of reading the notebook, at least the comments before believing some blend. Similar happened here.\n\nAlso, at a time, I was around 340 in this competition, after that Trap came in, I'm pushed to 600+. That's why I think it'll be fun to watch the shake up! :)" }, { "id": 967709, "postDate": "2020-08-12T12:59:23Z", "content": "@cpmpml, Porto Seguro was when I realized there are so many focusing on the gamification aspect of kaggling. I just closed my eyes and ears to remain disciplined, conservative etc. But when you jumped from 1178 public LB to 29 private LB, you earned my respect. I only moved up 9 places to 62nd place on the private LB with simple averaging of diversed submissions making porto my best performance up until that time.
\nHowever, my protypical example of the danger of blends and overfitting was the home credit competiton. I was also so discouraged during the competition like you @serigne but jumped 400+ places to silver in the private. Discouraged because I had limited time to catchup since I was also working on the TrackML competition that was so so exciting. But my biggest suprise there was how so many experienced kagglers in teams that fell so badly in the end.
", "rawMarkdown": "@cpmpml, [Porto Seguro](https://www.kaggle.com/c/porto-seguro-safe-driver-prediction) was when I realized there are so many focusing on the gamification aspect of kaggling. I just closed my eyes and ears to remain disciplined, conservative etc. But when you jumped from 1178 public LB to 29 private LB, you earned my respect. I only moved up 9 places to 62nd place on the private LB with simple averaging of diversed submissions making porto my best performance up until that time. \n\nHowever, my protypical example of the danger of blends and overfitting was the [home credit competiton](https://www.kaggle.com/c/home-credit-default-risk/). I was also so discouraged during the competition like you @serigne but jumped 400+ places to silver in the private. Discouraged because I had limited time to catchup since I was also working on the TrackML competition that was so so exciting. But my biggest suprise there was how so many experienced kagglers in teams that fell so badly in the end.", "votes": 3 }, { "id": 967727, "postDate": "2020-08-12T13:11:20.093Z", "content": "@sheriytm thanks.
", "rawMarkdown": "@sheriytm thanks.", "votes": 1 }, { "id": 967753, "postDate": "2020-08-12T13:33:47.770Z", "content": "In what seems like a long time ago, in a galaxy far, far away, I was a beneficiary of the big shakeup in Porto Seguro, moving up 1226 places from rank 1249 on Public LB to 23 on Private. I have no hope for myself of repeating that feat in this contest, but many other participants still have a chance, so don't let the Public LB discourage you, or seduce you into making poor selections.
", "rawMarkdown": "In what seems like a long time ago, in a galaxy far, far away, I was a beneficiary of the big shakeup in Porto Seguro, moving up 1226 places from rank 1249 on Public LB to 23 on Private. I have no hope for myself of repeating that feat in this contest, but many other participants still have a chance, so don't let the Public LB discourage you, or seduce you into making poor selections.", "votes": 3 }, { "id": 967757, "postDate": "2020-08-12T13:38:54.443Z", "content": "I remember your move that was even bigger than mine. It wa s no luck as far as I remember.
", "rawMarkdown": "I remember your move that was even bigger than mine. It wa s no luck as far as I remember." }, { "id": 967853, "postDate": "2020-08-12T14:40:28.877Z", "content": "@dslate as rightly said by @cpmpml, that was not luck. No disrespect your jump in porto was amazing as well, I noticed @cpmpml's because we were in TrackML together where I asked so many questions of him during the competition.
", "rawMarkdown": "@dslate as rightly said by @cpmpml, that was not luck. No disrespect your jump in porto was amazing as well, I noticed @cpmpml's because we were in [TrackML](https://www.kaggle.com/c/trackml-particle-identification) together where I asked so many questions of him during the competition." }, { "id": 972677, "postDate": "2020-08-16T18:25:01.590Z", "content": "@cpmpml, @sheriytm: I didn't mean to imply that my jump in porto was substantially due to luck. I attribute it to effort (my own and my computers), persistence, a combination of techniques and methods I had practiced in previous contests and projects, and a stubborn refusal to pay much attention to the Public LB. I also felt more like I knew what I was doing in that contest than in some of the others I've been in (including this one), perhaps because it featured a decent number of records and variables, but no images, which was like my last big predictive analytics project in the \"real world\" before I retired: building and maintaining a credit card fraud detection system for a major card issuer.
\nI haven't yet gotten comfortable with computer vision problems, although I enjoy working on them, and for some reason I'm reluctant to make use of external data, so I would be very surprised if I do well in this contest even if I rise up a few places at the end.
Have you tried ensembling your model with the public notebook? 0.95 * public_notebook_predictions + 0.05 * your_best_sub ? It will probably be overfitted but you can use that as one of your 3 submissions and jump over everyone on the public LB.
I think everybody will use one sub this way ;)
", "rawMarkdown": "I think everybody will use one sub this way ;)", "votes": 1 }, { "id": 971365, "postDate": "2020-08-15T13:03:05.550Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] }, { "id": 968293, "postDate": "2020-08-12T20:51:24.710Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1984321%2F256798817bc1483797a11db080ad72b1%2Finbox_352359_a2f8fba5a96fd8e6af73c4ecfbe0bf5f_photo5794133576814539089.jpg?generation=1597265419195160&alt=media\")
Don't worry, the heavy green jump in Private LB after shakeup will be a delight to see. ππ
", "rawMarkdown": "Don't worry, the heavy green jump in Private LB after shakeup will be a delight to see. ππ", "votes": 6, "replies": [ { "id": 966018, "postDate": "2020-08-11T03:58:58.637Z", "content": "But what if their submission isn't overfitted one? Will kaggle remove the duplicated submissions?
", "rawMarkdown": "But what if their submission isn't overfitted one? Will kaggle remove the duplicated submissions?", "votes": -2 }, { "id": 966022, "postDate": "2020-08-11T04:04:14.107Z", "content": "Nope, I don't think Kaggle will remove these solutions, although, I can tell you that this score is overfitting, I saw the notebook which shared this solution, you too can see the comments there if not the notebook itself! A thing is clearly written there, \"It's a Trap\". :)
\nSee this.
", "rawMarkdown": "Nope, I don't think Kaggle will remove these solutions, although, I can tell you that this score is overfitting, I saw the notebook which shared this solution, you too can see the comments there if not the notebook itself! A thing is clearly written there, \"It's a Trap\". :)\n\nSee [this](https://www.kaggle.com/paklau9/minmax-highest-public-lb-9619#961474).", "votes": 2 } ] }, { "id": 967910, "postDate": "2020-08-12T15:25:21.710Z", "content": "Please, don't get frustrated. I recently participated in the Global wheat Detection challenge, stories are always similar. A big shakeup in the private leaderboard. So trust your CV and it will bring you ultimate glory.
", "rawMarkdown": "Please, don't get frustrated. I recently participated in the Global wheat Detection challenge, stories are always similar. A big shakeup in the private leaderboard. So trust your CV and it will bring you ultimate glory.", "votes": 2 }, { "id": 966006, "postDate": "2020-08-11T03:44:58.920Z", "content": "Blend blend blend β¦β¦usually will overfit and cant do well in private LB
", "rawMarkdown": "Blend blend blend ......usually will overfit and cant do well in private LB ", "votes": 2 }, { "id": 965991, "postDate": "2020-08-11T03:21:07.070Z", "content": "I agree with you. Let's keep our faith and continue to focus on improving our CV results. It has been shown that it is possible to reverse engineer the public LB so some of the submission files publicly made available could have been altered to inflate public LB score
", "rawMarkdown": "I agree with you. Let's keep our faith and continue to focus on improving our CV results. It has been shown that it is possible to reverse engineer the public LB so some of the submission files publicly made available could have been altered to inflate public LB score", "votes": 2 }, { "id": 977246, "postDate": "2020-08-19T11:23:09.647Z", "content": "hmmmmmmmmmmmm
", "rawMarkdown": "hmmmmmmmmmmmm" }, { "id": 970599, "postDate": "2020-08-14T15:44:45.410Z", "content": "They probably will drop down due to overfitting.
", "rawMarkdown": "They probably will drop down due to overfitting." }, { "id": 967742, "postDate": "2020-08-12T13:24:32.720Z", "content": "Agree. Well, probably slight overfit, but we'll see.
", "rawMarkdown": "Agree. Well, probably slight overfit, but we'll see." }, { "id": 966960, "postDate": "2020-08-11T19:52:34.770Z", "content": "The forum discussion after the so waited shake up is going to be great.
", "rawMarkdown": "The forum discussion after the so waited shake up is going to be great." }, { "id": 966766, "postDate": "2020-08-11T16:47:07.900Z", "content": "Its okay the notebook was public by author's choice !
", "rawMarkdown": "Its okay the notebook was public by author's choice !" }, { "id": 969043, "postDate": "2020-08-13T12:28:38.403Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 966775, "author_name": "CPMP", "author_url": "", "post_date": "2020-08-11T16:50:05.807000", "content": "Watch them fall in private LB
\n
Like tom said, Don't believe it π
\n
Porto Seguro was among my first competitions on kaggle.
\nI left the competion about 1 month before the end, after submitting my best CV.
\nBut during the last month, many people were submitting a highly overfitting public blend kernel. But they fell on private LB and I jumped up + 800 on that private LB ^^
", "votes": 8, "replies": [ { "id": 966801, "author_name": "Gajendra Saraswat", "author_url": "", "post_date": "2020-08-11T17:06:23.853000", "content": "Seems like Masters and Grand masters have a lot of such sweet stories! ^_^
", "votes": 0, "replies": [] }, { "id": 966816, "author_name": "CPMP", "author_url": "", "post_date": "2020-08-11T17:17:25.110000", "content": "Porto Seguro is also my prototypical examples as well for how dangerous blends of blends of blends of public notebooks can be. I moved from 1178 public LB to 29 private LB there. I would not mind if something similar happens in this competition. LOL.
", "votes": 15, "replies": [] }, { "id": 966825, "author_name": "Serigne ", "author_url": "", "post_date": "2020-08-11T17:20:02.923000", "content": "\nI was novive back then ;) And to be fair I was a bit discouraged and decided to leave the competition when I saw I dropped +100 places just few hours after the kernel was published.
\nIn retrospective , I think it was a good decision , because as novice, it's easy to be fooled by public LB and public best kernels score and I would likely take part of the blending game if I stayed.
", "votes": 1, "replies": [] }, { "id": 966832, "author_name": "Serigne ", "author_url": "", "post_date": "2020-08-11T17:25:20.120000", "content": "\n\nI would not mind if something similar happens in this competition. LOL.
\n
May be, but 3 submissions choice may mitigate a bit the shake up ;)
", "votes": 1, "replies": [] }, { "id": 966841, "author_name": "CPMP", "author_url": "", "post_date": "2020-08-11T17:28:39.500000", "content": "3 chances of overfitting instead of 2?
", "votes": 5, "replies": [ { "id": 967342, "author_name": "Manjesh Gupta", "author_url": "", "post_date": "2020-08-12T07:10:56.573000", "content": "The thing is with 3 subs, everyone (incl. top in public LB) can easily take a risk with their best \"over-fitted\" model. 1 sub - best over-fit and other 2 subs - good cv + indicative LB + intuition. Plus the variance in models and huge class imbalance - anything can happen on private LB.
", "votes": 0, "replies": [] }, { "id": 967677, "author_name": "CPMP", "author_url": "", "post_date": "2020-08-12T12:21:52.770000", "content": "You're confirming that there will be more overfitting with 3 subs than with 2, thanks ;)
", "votes": 1, "replies": [] } ] }, { "id": 966861, "author_name": "Gajendra Saraswat", "author_url": "", "post_date": "2020-08-11T17:41:30.337000", "content": "True @serigne sama, although I have participated only in 1 competition before this one(Recursion), I have this habit of reading the notebook, at least the comments before believing some blend. Similar happened here.
\nAlso, at a time, I was around 340 in this competition, after that Trap came in, I'm pushed to 600+. That's why I think it'll be fun to watch the shake up! :)
", "votes": 0, "replies": [] }, { "id": 967709, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-08-12T12:59:23", "content": "@cpmpml, Porto Seguro was when I realized there are so many focusing on the gamification aspect of kaggling. I just closed my eyes and ears to remain disciplined, conservative etc. But when you jumped from 1178 public LB to 29 private LB, you earned my respect. I only moved up 9 places to 62nd place on the private LB with simple averaging of diversed submissions making porto my best performance up until that time.
\nHowever, my protypical example of the danger of blends and overfitting was the home credit competiton. I was also so discouraged during the competition like you @serigne but jumped 400+ places to silver in the private. Discouraged because I had limited time to catchup since I was also working on the TrackML competition that was so so exciting. But my biggest suprise there was how so many experienced kagglers in teams that fell so badly in the end.
", "votes": 3, "replies": [] }, { "id": 967727, "author_name": "CPMP", "author_url": "", "post_date": "2020-08-12T13:11:20.093000", "content": "@sheriytm thanks.
", "votes": 1, "replies": [] }, { "id": 967753, "author_name": "David J. Slate", "author_url": "", "post_date": "2020-08-12T13:33:47.770000", "content": "In what seems like a long time ago, in a galaxy far, far away, I was a beneficiary of the big shakeup in Porto Seguro, moving up 1226 places from rank 1249 on Public LB to 23 on Private. I have no hope for myself of repeating that feat in this contest, but many other participants still have a chance, so don't let the Public LB discourage you, or seduce you into making poor selections.
", "votes": 3, "replies": [] }, { "id": 967757, "author_name": "CPMP", "author_url": "", "post_date": "2020-08-12T13:38:54.443000", "content": "I remember your move that was even bigger than mine. It wa s no luck as far as I remember.
", "votes": 0, "replies": [] }, { "id": 967853, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-08-12T14:40:28.877000", "content": "@dslate as rightly said by @cpmpml, that was not luck. No disrespect your jump in porto was amazing as well, I noticed @cpmpml's because we were in TrackML together where I asked so many questions of him during the competition.
", "votes": 0, "replies": [] }, { "id": 972677, "author_name": "David J. Slate", "author_url": "", "post_date": "2020-08-16T18:25:01.590000", "content": "@cpmpml, @sheriytm: I didn't mean to imply that my jump in porto was substantially due to luck. I attribute it to effort (my own and my computers), persistence, a combination of techniques and methods I had practiced in previous contests and projects, and a stubborn refusal to pay much attention to the Public LB. I also felt more like I knew what I was doing in that contest than in some of the others I've been in (including this one), perhaps because it featured a decent number of records and variables, but no images, which was like my last big predictive analytics project in the \"real world\" before I retired: building and maintaining a credit card fraud detection system for a major card issuer.
\nI haven't yet gotten comfortable with computer vision problems, although I enjoy working on them, and for some reason I'm reluctant to make use of external data, so I would be very surprised if I do well in this contest even if I rise up a few places at the end.
Have you tried ensembling your model with the public notebook? 0.95 * public_notebook_predictions + 0.05 * your_best_sub ? It will probably be overfitted but you can use that as one of your 3 submissions and jump over everyone on the public LB.
I think everybody will use one sub this way ;)
", "votes": 1, "replies": [] }, { "id": 971365, "author_name": "", "author_url": "", "post_date": "2020-08-15T13:03:05.550000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 968293, "author_name": "Innat", "author_url": "", "post_date": "2020-08-12T20:51:24.710000", "content": "
Don't worry, the heavy green jump in Private LB after shakeup will be a delight to see. ππ
", "votes": 6, "replies": [ { "id": 966018, "author_name": "Aakash V", "author_url": "", "post_date": "2020-08-11T03:58:58.637000", "content": "But what if their submission isn't overfitted one? Will kaggle remove the duplicated submissions?
", "votes": -2, "replies": [] }, { "id": 966022, "author_name": "Gajendra Saraswat", "author_url": "", "post_date": "2020-08-11T04:04:14.107000", "content": "Nope, I don't think Kaggle will remove these solutions, although, I can tell you that this score is overfitting, I saw the notebook which shared this solution, you too can see the comments there if not the notebook itself! A thing is clearly written there, \"It's a Trap\". :)
\nSee this.
", "votes": 2, "replies": [] } ] }, { "id": 967910, "author_name": "Ahmmad Rashid", "author_url": "", "post_date": "2020-08-12T15:25:21.710000", "content": "Please, don't get frustrated. I recently participated in the Global wheat Detection challenge, stories are always similar. A big shakeup in the private leaderboard. So trust your CV and it will bring you ultimate glory.
", "votes": 2, "replies": [] }, { "id": 966006, "author_name": "Salaryman", "author_url": "", "post_date": "2020-08-11T03:44:58.920000", "content": "Blend blend blend β¦β¦usually will overfit and cant do well in private LB
", "votes": 2, "replies": [] }, { "id": 965991, "author_name": "Alan Choon", "author_url": "", "post_date": "2020-08-11T03:21:07.070000", "content": "I agree with you. Let's keep our faith and continue to focus on improving our CV results. It has been shown that it is possible to reverse engineer the public LB so some of the submission files publicly made available could have been altered to inflate public LB score
", "votes": 2, "replies": [] }, { "id": 977246, "author_name": "hustlerkaggler", "author_url": "", "post_date": "2020-08-19T11:23:09.647000", "content": "hmmmmmmmmmmmm
", "votes": 0, "replies": [] }, { "id": 970599, "author_name": "Nickolay Safronov", "author_url": "", "post_date": "2020-08-14T15:44:45.410000", "content": "They probably will drop down due to overfitting.
", "votes": 0, "replies": [] }, { "id": 967742, "author_name": "lr", "author_url": "", "post_date": "2020-08-12T13:24:32.720000", "content": "Agree. Well, probably slight overfit, but we'll see.
", "votes": 0, "replies": [] }, { "id": 966960, "author_name": "Santiago Viquez", "author_url": "", "post_date": "2020-08-11T19:52:34.770000", "content": "The forum discussion after the so waited shake up is going to be great.
", "votes": 0, "replies": [] }, { "id": 966766, "author_name": "Utkarsh Chandra Srivastava", "author_url": "", "post_date": "2020-08-11T16:47:07.900000", "content": "Its okay the notebook was public by author's choice !
", "votes": 0, "replies": [] }, { "id": 969043, "author_name": "", "author_url": "", "post_date": "2020-08-13T12:28:38.403000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "966775": "Watch them fall in private LB\n\n", "965989": "99% of submissions from position 314 to 493 (179 positions ) \n\nwas from this [notebook ](https://www.kaggle.com/paklau9/minmax-highest-public-lb-9619).\nIt is really annoying for the ones on the bottom of 0.9619 to loose 170 positions despite the hard work. Kaggle Team should find a way to solve this issue.", "968309": "Like tom said, **Don't believe it** π\n\n", "966797": "[Porto Seguro](https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/leaderboard) was among my first competitions on kaggle. \n\nI left the competion about 1 month before the end, after submitting my best CV. \n\nBut during the last month, many people were submitting a highly overfitting public blend kernel. But they fell on private LB and I jumped up + 800 on that private LB ^^\n", "968493": "Have you tried ensembling your model with the public notebook? `0.95 * public_notebook_predictions + 0.05 * your_best_sub` ? It will probably be overfitted but you can use that as one of your 3 submissions and jump over everyone on the public LB.", "968293": "", "966016": "Don't worry, the heavy green jump in Private LB after shakeup will be a delight to see. ππ", "967910": "Please, don't get frustrated. I recently participated in the Global wheat Detection challenge, stories are always similar. A big shakeup in the private leaderboard. So trust your CV and it will bring you ultimate glory.", "966006": "Blend blend blend ......usually will overfit and cant do well in private LB ", "965991": "I agree with you. Let's keep our faith and continue to focus on improving our CV results. It has been shown that it is possible to reverse engineer the public LB so some of the submission files publicly made available could have been altered to inflate public LB score", "977246": "hmmmmmmmmmmmm", "970599": "They probably will drop down due to overfitting.", "967742": "Agree. Well, probably slight overfit, but we'll see.", "966960": "The forum discussion after the so waited shake up is going to be great.", "966766": "Its okay the notebook was public by author's choice !", "969043": "" } }