Kaggle is a good platform and I appreciate the experts in the fields. But blending actually leads to mega submissions which discourage the bronze position candidates who are intermediate in machine learning.
\n\nFurther, it defeats the purpose of learning and committing the kernel without prior knowledge can actually land you higher than someone who is attempting the competition from scratch. This isnt ethical
\n\nOr, if some are very keen showing the blending expertise please do post an EDA on weights. Avoid submit a kernel submission leading to false leaderboard positions without a minute's effort. This might turn down enthusiast candidates and encourage just a numbers competition.
\n\nI request Kaggle to avoid kernel public if it surpasses some x% of bronze position. Please do share your opinions.
", "messages": [ { "id": "301800", "postDate": "03/23/2018 09:19:39", "content": "Kaggle is a good platform and I appreciate the experts in the fields. But blending actually leads to mega submissions which discourage the bronze position candidates who are intermediate in machine learning.
\n\nFurther, it defeats the purpose of learning and committing the kernel without prior knowledge can actually land you higher than someone who is attempting the competition from scratch. This isnt ethical
\n\nOr, if some are very keen showing the blending expertise please do post an EDA on weights. Avoid submit a kernel submission leading to false leaderboard positions without a minute's effort. This might turn down enthusiast candidates and encourage just a numbers competition.
\n\nI request Kaggle to avoid kernel public if it surpasses some x% of bronze position. Please do share your opinions.
", "rawMarkdown": "Kaggle is a good platform and I appreciate the experts in the fields. But blending actually leads to mega submissions which discourage the bronze position candidates who are intermediate in machine learning. \n\nFurther, it defeats the purpose of learning and committing the kernel without prior knowledge can actually land you higher than someone who is attempting the competition from scratch. This isnt ethical \n\n## Make Blends educational instead\n- Why do these weights work?\n- How you got that weights?\n\nOr, if some are very keen showing the blending expertise please do post an EDA on weights. Avoid submit a kernel submission leading to false leaderboard positions without a minute's effort. This might turn down **enthusiast** candidates and encourage just a numbers competition.\n\nI request Kaggle to avoid kernel public if it surpasses some x% of bronze position. Please do share your opinions.", "votes": null }, { "id": "301931", "postDate": "03/23/2018 13:33:51", "content": "Totally agree, posting blindly blending to public is not useful, but we could learn a lot if it is evidence based
", "rawMarkdown": "Totally agree, posting blindly blending to public is not useful, but we could learn a lot if it is evidence based", "votes": null }, { "id": "302195", "postDate": "03/23/2018 19:29:38", "content": "At the end of the day it really just creates lots of leaderboard noise, since typically the blends are overfit to the leaderboard so in the private leaderboard update there'll be a huge upset.
", "rawMarkdown": "At the end of the day it really just creates lots of leaderboard noise, since typically the blends are overfit to the leaderboard so in the private leaderboard update there'll be a huge upset.", "votes": null }, { "id": "302411", "postDate": "03/24/2018 02:14:33", "content": "Unfortunatelly, just because in theory it is true and we hope it to be true for all cases, it doesn't mean that it is always true. As an example -> the toxic comp
", "rawMarkdown": "Unfortunatelly, just because in theory it is true and we hope it to be true for all cases, it doesn't mean that it is always true. As an example -> the toxic comp", "votes": null }, { "id": "302470", "postDate": "03/24/2018 05:11:49", "content": "Actually I totally agree with that, I do a lot of work on structure to improve my nn scores,and I made top 3% which I was proud of.Howerver after a blending kernel submission which is higher than me by 0.001 point and 1 hour later,I was going down from top 30 to top150,I think there must be people just submit the blending submission.
", "rawMarkdown": "Actually I totally agree with that, I do a lot of work on structure to improve my nn scores,and I made top 3% which I was proud of.Howerver after a blending kernel submission which is higher than me by 0.001 point and 1 hour later,I was going down from top 30 to top150,I think there must be people just submit the blending submission.", "votes": null }, { "id": "302502", "postDate": "03/24/2018 07:20:21", "content": "I have posted my view in spongebobs thread as below. I respect everyone's views.
I agree Blending is also an art and a piece of data science solutions framework.
\nBut we should also encourage the Diverse solutions and the effort put in them.
My suggestion to @Kaggle is
\nto make an intermediate dead line like ( one/two month after launch of competition) ,
Agreed! Wonderful idea
", "rawMarkdown": "Agreed! Wonderful idea", "votes": null }, { "id": "303519", "postDate": "03/26/2018 10:57:49", "content": "I vote for totally blocking out kernal submissions that achieves bronze or higher on public LB, either do this completely throughout the competition or do this near to the ending or middle of the competition, this will prevent actual capable bronze achievers to be degraded by fake achievers that just submits kernal submissions. If not I feel it will ultimately degrade the actual value and skillset needed to achieve a bronze medal or higher. Or maybe kaggle can just clean them up after the competitions lol
", "rawMarkdown": "I vote for totally blocking out kernal submissions that achieves bronze or higher on public LB, either do this completely throughout the competition or do this near to the ending or middle of the competition, this will prevent actual capable bronze achievers to be degraded by fake achievers that just submits kernal submissions. If not I feel it will ultimately degrade the actual value and skillset needed to achieve a bronze medal or higher. Or maybe kaggle can just clean them up after the competitions lol", "votes": null }, { "id": "303783", "postDate": "03/26/2018 16:54:51", "content": "Ok so people will clone kernels into private kernels, change a few insignificant lines and submit that instead. If you do exact submission comparison then somebody will add tiny amount of noise so you checker doesn't trigger.\nOn the other hand, if you have a good model , nothing is stopping you from blending it with all the other models that appear - you should end up above them.
\n\nI think end of competition release of information seems to be incentivised by kaggle. I don't like it, and clearly many other people don't but it is now a definite part of the game.
", "rawMarkdown": "Ok so people will clone kernels into private kernels, change a few insignificant lines and submit that instead. If you do exact submission comparison then somebody will add tiny amount of noise so you checker doesn't trigger.\nOn the other hand, if you have a good model , nothing is stopping you from blending it with all the other models that appear - you should end up above them.\n\nI think end of competition release of information seems to be incentivised by kaggle. I don't like it, and clearly many other people don't but it is now a definite part of the game.", "votes": null }, { "id": "304053", "postDate": "03/27/2018 00:09:06", "content": "I propose a solution that does not prejudice the dissemination of knowledge through kernel sharing and that could be useful to avoid several distortions of the competitions in Kaggle, besides benefiting the development of original kernels.
\n\nMy proposal would simply block or disqualify submissions that have high correlation with other submissions already made in a competition. In order for the processing of this new criterion not to consume many computational resources, it would be sufficient to compare a small part of the rows of the new submission with the submissions of the public kernels.
", "rawMarkdown": "I propose a solution that does not prejudice the dissemination of knowledge through kernel sharing and that could be useful to avoid several distortions of the competitions in Kaggle, besides benefiting the development of original kernels.\n\nMy proposal would simply block or disqualify submissions that have high correlation with other submissions already made in a competition. In order for the processing of this new criterion not to consume many computational resources, it would be sufficient to compare a small part of the rows of the new submission with the submissions of the public kernels.", "votes": null }, { "id": "304167", "postDate": "03/27/2018 05:55:14", "content": "Some of my submissions are highly correlated with each other. I'm not sure I would like them t be removed.
", "rawMarkdown": "Some of my submissions are highly correlated with each other. I'm not sure I would like them t be removed.", "votes": null }, { "id": "304185", "postDate": "03/27/2018 06:26:00", "content": "\n\n\nMy proposal would simply block or disqualify submissions that have high correlation with other submissions already made in a competition.
\n
@PaoloPinto, IMHO, in this kind of competition, it is not a good idea. In the matter of improving the importance of the original over the blending kernels, the effect would be the opposite - the results of a little bit improved original method would be rejected, and blending kernel results, as less correlated to the original submissions, would be accepted.
", "rawMarkdown": "> My proposal would simply block or disqualify submissions that have high correlation with other submissions already made in a competition.\n\n@PaoloPinto, IMHO, in this kind of competition, it is not a good idea. In the matter of improving the importance of the original over the blending kernels, the effect would be the opposite - the results of a little bit improved original method would be rejected, and blending kernel results, as less correlated to the original submissions, would be accepted.", "votes": null }, { "id": "304195", "postDate": "03/27/2018 06:44:37", "content": "@Kishore M, Your proposal makes sense, however the blending race starts usually few weeks before the end of the competition, when there is no new idea how to improve original methods, so in practice your proposal would simply states that blending kernels are private during the competition.
", "rawMarkdown": "Kishore M, Your proposal makes sense, however the blending race starts usually few weeks before the end of the competition, when there is no new idea how to improve original methods, so in practice your proposal would simply states that blending kernels are private during the competition.", "votes": null }, { "id": "304214", "postDate": "03/27/2018 07:19:42", "content": "Sure, nothing wrong with blending models! There is a lot of nice ideas and tricks to be applied there as well. Blindly blending as you say, is just useless though. In my understanding kernels have the option to be made public in order to stimulate learning, and not to stimulate blindly copy-pasting or even worse download-uploading.
", "rawMarkdown": "Sure, nothing wrong with blending models! There is a lot of nice ideas and tricks to be applied there as well. Blindly blending as you say, is just useless though. In my understanding kernels have the option to be made public in order to stimulate learning, and not to stimulate blindly copy-pasting or even worse download-uploading.", "votes": null }, { "id": "304365", "postDate": "03/27/2018 13:12:32", "content": "@Grzegorz Sionkowski thanks for your reply
\nYes .. ! it serves Following purposes by keeping a dead line for public kernels
\n1. The blending or any kernels after certain date will remain private. Only best models published publicly before this date can be blended by individuals.
\n2. Any original diverse solution/(best solution) after certain date also remains private so who worked for diverse solution will get rewarded accordingly for their work.
\n3. For Learning purposes and knowledge sharing these private kernels can be marked public post competition
OK, I've done the \"make blends educational\" thing.
", "rawMarkdown": "OK, I've done the \"[make blends educational][1]\" thing.\n\n [1]: https://www.kaggle.com/aharless/simple-linear-stacking-lb-9704", "votes": null }, { "id": "308135", "postDate": "04/03/2018 02:06:38", "content": "Thanks a bunch. A perfect educational thing!
", "rawMarkdown": "Thanks a bunch. A perfect educational thing!", "votes": null } ], "comments": [ { "id": 301931, "author_name": "wythhh", "author_url": "", "post_date": "03/23/2018 13:33:51", "content": "Totally agree, posting blindly blending to public is not useful, but we could learn a lot if it is evidence based
", "votes": null, "replies": [ { "id": 304214, "author_name": "asparuhhristov", "author_url": "", "post_date": "03/27/2018 07:19:42", "content": "Sure, nothing wrong with blending models! There is a lot of nice ideas and tricks to be applied there as well. Blindly blending as you say, is just useless though. In my understanding kernels have the option to be made public in order to stimulate learning, and not to stimulate blindly copy-pasting or even worse download-uploading.
", "votes": null, "replies": [] } ] }, { "id": 302195, "author_name": "msf908", "author_url": "", "post_date": "03/23/2018 19:29:38", "content": "At the end of the day it really just creates lots of leaderboard noise, since typically the blends are overfit to the leaderboard so in the private leaderboard update there'll be a huge upset.
", "votes": null, "replies": [ { "id": 302411, "author_name": "asparuhhristov", "author_url": "", "post_date": "03/24/2018 02:14:33", "content": "Unfortunatelly, just because in theory it is true and we hope it to be true for all cases, it doesn't mean that it is always true. As an example -> the toxic comp
", "votes": null, "replies": [] } ] }, { "id": 302470, "author_name": "liubingchen", "author_url": "", "post_date": "03/24/2018 05:11:49", "content": "Actually I totally agree with that, I do a lot of work on structure to improve my nn scores,and I made top 3% which I was proud of.Howerver after a blending kernel submission which is higher than me by 0.001 point and 1 hour later,I was going down from top 30 to top150,I think there must be people just submit the blending submission.
", "votes": null, "replies": [] }, { "id": 302502, "author_name": "reachkishore", "author_url": "", "post_date": "03/24/2018 07:20:21", "content": "I have posted my view in spongebobs thread as below. I respect everyone's views.
I agree Blending is also an art and a piece of data science solutions framework.
\nBut we should also encourage the Diverse solutions and the effort put in them.
My suggestion to @Kaggle is
\nto make an intermediate dead line like ( one/two month after launch of competition) ,
Agreed! Wonderful idea
", "votes": null, "replies": [] }, { "id": 304195, "author_name": "sionek", "author_url": "", "post_date": "03/27/2018 06:44:37", "content": "@Kishore M, Your proposal makes sense, however the blending race starts usually few weeks before the end of the competition, when there is no new idea how to improve original methods, so in practice your proposal would simply states that blending kernels are private during the competition.
", "votes": null, "replies": [] }, { "id": 304365, "author_name": "reachkishore", "author_url": "", "post_date": "03/27/2018 13:12:32", "content": "@Grzegorz Sionkowski thanks for your reply
\nYes .. ! it serves Following purposes by keeping a dead line for public kernels
\n1. The blending or any kernels after certain date will remain private. Only best models published publicly before this date can be blended by individuals.
\n2. Any original diverse solution/(best solution) after certain date also remains private so who worked for diverse solution will get rewarded accordingly for their work.
\n3. For Learning purposes and knowledge sharing these private kernels can be marked public post competition
I vote for totally blocking out kernal submissions that achieves bronze or higher on public LB, either do this completely throughout the competition or do this near to the ending or middle of the competition, this will prevent actual capable bronze achievers to be degraded by fake achievers that just submits kernal submissions. If not I feel it will ultimately degrade the actual value and skillset needed to achieve a bronze medal or higher. Or maybe kaggle can just clean them up after the competitions lol
", "votes": null, "replies": [ { "id": 303783, "author_name": "stimakov", "author_url": "", "post_date": "03/26/2018 16:54:51", "content": "Ok so people will clone kernels into private kernels, change a few insignificant lines and submit that instead. If you do exact submission comparison then somebody will add tiny amount of noise so you checker doesn't trigger.\nOn the other hand, if you have a good model , nothing is stopping you from blending it with all the other models that appear - you should end up above them.
\n\nI think end of competition release of information seems to be incentivised by kaggle. I don't like it, and clearly many other people don't but it is now a definite part of the game.
", "votes": null, "replies": [] } ] }, { "id": 304053, "author_name": "paulorzp", "author_url": "", "post_date": "03/27/2018 00:09:06", "content": "I propose a solution that does not prejudice the dissemination of knowledge through kernel sharing and that could be useful to avoid several distortions of the competitions in Kaggle, besides benefiting the development of original kernels.
\n\nMy proposal would simply block or disqualify submissions that have high correlation with other submissions already made in a competition. In order for the processing of this new criterion not to consume many computational resources, it would be sufficient to compare a small part of the rows of the new submission with the submissions of the public kernels.
", "votes": null, "replies": [ { "id": 304167, "author_name": "cpmpml", "author_url": "", "post_date": "03/27/2018 05:55:14", "content": "Some of my submissions are highly correlated with each other. I'm not sure I would like them t be removed.
", "votes": null, "replies": [] }, { "id": 304185, "author_name": "sionek", "author_url": "", "post_date": "03/27/2018 06:26:00", "content": "\n\n\nMy proposal would simply block or disqualify submissions that have high correlation with other submissions already made in a competition.
\n
@PaoloPinto, IMHO, in this kind of competition, it is not a good idea. In the matter of improving the importance of the original over the blending kernels, the effect would be the opposite - the results of a little bit improved original method would be rejected, and blending kernel results, as less correlated to the original submissions, would be accepted.
", "votes": null, "replies": [] } ] }, { "id": 307988, "author_name": "aharless", "author_url": "", "post_date": "04/02/2018 18:57:42", "content": "OK, I've done the \"make blends educational\" thing.
", "votes": null, "replies": [ { "id": 308135, "author_name": "shaz13", "author_url": "", "post_date": "04/03/2018 02:06:38", "content": "Thanks a bunch. A perfect educational thing!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "301800": "Kaggle is a good platform and I appreciate the experts in the fields. But blending actually leads to mega submissions which discourage the bronze position candidates who are intermediate in machine learning. \n\nFurther, it defeats the purpose of learning and committing the kernel without prior knowledge can actually land you higher than someone who is attempting the competition from scratch. This isnt ethical \n\n## Make Blends educational instead\n- Why do these weights work?\n- How you got that weights?\n\nOr, if some are very keen showing the blending expertise please do post an EDA on weights. Avoid submit a kernel submission leading to false leaderboard positions without a minute's effort. This might turn down **enthusiast** candidates and encourage just a numbers competition.\n\nI request Kaggle to avoid kernel public if it surpasses some x% of bronze position. Please do share your opinions.", "301931": "Totally agree, posting blindly blending to public is not useful, but we could learn a lot if it is evidence based", "302195": "At the end of the day it really just creates lots of leaderboard noise, since typically the blends are overfit to the leaderboard so in the private leaderboard update there'll be a huge upset.", "302411": "Unfortunatelly, just because in theory it is true and we hope it to be true for all cases, it doesn't mean that it is always true. As an example -> the toxic comp", "302470": "Actually I totally agree with that, I do a lot of work on structure to improve my nn scores,and I made top 3% which I was proud of.Howerver after a blending kernel submission which is higher than me by 0.001 point and 1 hour later,I was going down from top 30 to top150,I think there must be people just submit the blending submission.", "302502": "I have posted my view in spongebobs thread as below. I respect everyone's views.