I find it so silly the amount of people submitting blending kernels or republishing kernels with very small changes or no relevant findings. I know that people are free to do whatever they want, and I endorse it, but what is the motivation behind doing it? Easy medals?
", "messages": [ { "id": 598599, "postDate": "2019-08-13T19:06:50.180Z", "content": "Blending gives kernel votes for next to zero work.\nPeople who vote for blends also want to do zero work.\nThis is called synergy - sadly ;)
", "rawMarkdown": "Blending gives kernel votes for next to zero work.\nPeople who vote for blends also want to do zero work.\nThis is called synergy - sadly ;)", "votes": 25 }, { "id": 598596, "postDate": "2019-08-13T19:02:25.210Z", "content": "I find it so silly the amount of people submitting blending kernels or republishing kernels with very small changes or no relevant findings. I know that people are free to do whatever they want, and I endorse it, but what is the motivation behind doing it? Easy medals?
", "rawMarkdown": "I find it so silly the amount of people submitting blending kernels or republishing kernels with very small changes or no relevant findings. I know that people are free to do whatever they want, and I endorse it, but what is the motivation behind doing it? Easy medals?", "votes": 17 }, { "id": 600266, "postDate": "2019-08-15T21:38:54.133Z", "content": "The really true when they do that is MEDALS -> Go to better rank -> Go for Linkedin -> #### print(\"I am a Master on Kaggle\")
\n\nand when they have a REAL PROJECT in Data Science.
\n\n---------------------------ERRORs IN RuntimeError, StopIteration, SyntaxError, etc ------------------------------\nZeroDivisionError : division by zero
", "rawMarkdown": "The really true when they do that is MEDALS -> Go to better rank -> Go for Linkedin -> #### print(\"I am a Master on Kaggle\") \n\nand when they have a REAL PROJECT in Data Science. \n\n---------------------------ERRORs IN RuntimeError, StopIteration, SyntaxError, etc ------------------------------\nZeroDivisionError : division by zero", "votes": 10 }, { "id": 599930, "postDate": "2019-08-15T13:12:52.243Z", "content": "Public blends, and more generally public kernels, aren't an issue as long as everybody can reuse them. The only issue is if someone published a very high quality kernel close to competition end, say during last week, in which case some people may not be able to reuse. It happened in Talking Data competition where someone posted a solid silver medal kernel few hours before end. Many did not have any sub left and saw 600 people pass them. Most silver medals in that competition were illegitimate.
", "rawMarkdown": "Public blends, and more generally public kernels, aren't an issue as long as everybody can reuse them. The only issue is if someone published a very high quality kernel close to competition end, say during last week, in which case some people may not be able to reuse. It happened in Talking Data competition where someone posted a solid silver medal kernel few hours before end. Many did not have any sub left and saw 600 people pass them. Most silver medals in that competition were illegitimate.", "votes": 10, "replies": [ { "id": 599980, "postDate": "2019-08-15T13:52:58.417Z", "content": "Hey @cpmpml , I started seeing a warning from kaggle not to publish high scoring kernels towards the competition ends, perhaps they started freezing new kernel publishing ?
", "rawMarkdown": "Hey @cpmpml , I started seeing a warning from kaggle not to publish high scoring kernels towards the competition ends, perhaps they started freezing new kernel publishing ? " }, { "id": 599988, "postDate": "2019-08-15T13:56:17.453Z", "content": "Yes, they have installed this warning after that incident.
", "rawMarkdown": "Yes, they have installed this warning after that incident.", "votes": 2 }, { "id": 600003, "postDate": "2019-08-15T14:26:40.990Z", "content": "but just a warning ?
", "rawMarkdown": "but just a warning ?" }, { "id": 600028, "postDate": "2019-08-15T14:49:51.633Z", "content": "Yes, just a warning.
", "rawMarkdown": "Yes, just a warning.\n" }, { "id": 600990, "postDate": "2019-08-16T20:43:05.833Z", "content": "a downvote just for this convo, wow :)
", "rawMarkdown": "a downvote just for this convo, wow :) " } ] }, { "id": 598933, "postDate": "2019-08-14T08:39:27.717Z", "content": "One can make good output with high score using blending with public kernels. But, Public kernels are really public, so Others can make similar output. I guess It is not easy to get medal by using blended result with public kernels.
\n\nTo get a medal, one need to make own solution.\nCombining the solution and other public kernels, one can get a medal.
", "rawMarkdown": "One can make good output with high score using blending with public kernels. But, Public kernels are really public, so Others can make similar output. I guess It is not easy to get medal by using blended result with public kernels.\n\nTo get a medal, one need to make own solution.\nCombining the solution and other public kernels, one can get a medal. ", "votes": 6 }, { "id": 598864, "postDate": "2019-08-14T06:37:14.240Z", "content": "Yeah. I think if you can't get past the blenders you don't deserve the medals. \nGuess I'll do some more workout in the playground and come back to medal challenges later :)
", "rawMarkdown": "Yeah. I think if you can't get past the blenders you don't deserve the medals. \nGuess I'll do some more workout in the playground and come back to medal challenges later :)", "votes": 3 }, { "id": 603503, "postDate": "2019-08-20T11:26:42.087Z", "content": "Wow, my SchNet Kernel is also incorporated into blending kernel ๐\n- Chemistry of Best Models (-1.822)
https://www.kaggle.com/roydatascience/chemistry-of-best-models-1-822
\n\n\nWow, my SchNet Kernel is also incorporated into blending kernel ยต
\n
that's ultimate recognition, you can drop from competition now ;)
\n\nJust kidding of course.
", "rawMarkdown": "> Wow, my SchNet Kernel is also incorporated into blending kernel ยต\n\nthat's ultimate recognition, you can drop from competition now ;)\n\nJust kidding of course.", "votes": 1 } ] }, { "id": 602859, "postDate": "2019-08-19T15:18:02.617Z", "content": "There's some merit to them since blending can be more complex than simply averaging a bunch of different predictions. If I'm blending my own predictions I like to use correlation matrices to identify models that have made distinctly different outputs for example, and select them accordingly while removing clear outliers.
\n\nIt's a dry area of learning frankly, I doubt it's what we come to Kaggle for! But there are certainly easy medals up for grabs - I posted a kernel in the LANL competition with a new approach I'd thought of and it got a gold medal, but looking back I'm certain that most of the votes were simply because the output had a very high Public LB score.
", "rawMarkdown": "There's some merit to them since blending can be more complex than simply averaging a bunch of different predictions. If I'm blending my own predictions I like to use correlation matrices to identify models that have made distinctly different outputs for example, and select them accordingly while removing clear outliers.\n\nIt's a dry area of learning frankly, I doubt it's what we come to Kaggle for! But there are certainly easy medals up for grabs - I posted a kernel in the LANL competition with a new approach I'd thought of and it got a gold medal, but looking back I'm certain that most of the votes were simply because the output had a very high Public LB score.", "votes": 1 }, { "id": 600550, "postDate": "2019-08-16T09:03:19.900Z", "content": "Even if I accept blending kernels, blend of blends are useless. They spoil the reproductivity of kernels and make weights unclear. Authors should use original outputs predicted by ML models as least.
", "rawMarkdown": "Even if I accept blending kernels, blend of blends are useless. They spoil the reproductivity of kernels and make weights unclear. Authors should use original outputs predicted by ML models as least.", "votes": 1 }, { "id": 599748, "postDate": "2019-08-15T10:06:31.617Z", "content": "I usually blend my models, LGM + NN etc. and sometimes it works sometimes it doesn't . Doing it so, finding the proper weights are important.
\n\nI understand your frustration @felipemello , but as @youhanlee said \"I guess It is not easy to get medal by using blended result with public kernels\"
", "rawMarkdown": "I usually blend my models, LGM + NN etc. and sometimes it works sometimes it doesn't . Doing it so, finding the proper weights are important.\n\nI understand your frustration @felipemello , but as @youhanlee said \"I guess It is not easy to get medal by using blended result with public kernels\"", "votes": 1 }, { "id": 598925, "postDate": "2019-08-14T08:25:07.663Z", "content": "I think blending is also an important technology.\nHowever, this competition has been blending since one month before the deadline. I think it โs too much blending.
", "rawMarkdown": "I think blending is also an important technology.\nHowever, this competition has been blending since one month before the deadline. I think it โs too much blending.", "votes": 1 }, { "id": 599009, "postDate": "2019-08-14T11:19:38.977Z", "content": "It's same in IEEE-CIS Fraud Detection. I guess every tabular data competition is like that.
", "rawMarkdown": "It's same in IEEE-CIS Fraud Detection. I guess every tabular data competition is like that.", "votes": 2 }, { "id": 602873, "postDate": "2019-08-19T15:42:34.200Z", "content": "Once I improved my own model, then another blend model appears OMG... I wish I could get out of the rat race right away.
", "rawMarkdown": "Once I improved my own model, then another blend model appears OMG... I wish I could get out of the rat race right away." }, { "id": 600528, "postDate": "2019-08-16T08:33:42.747Z", "rawMarkdown": "", "isDeleted": true }, { "id": 599765, "postDate": "2019-08-15T10:33:56.277Z", "rawMarkdown": "", "votes": -2, "isDeleted": true } ], "comments": [ { "id": 598599, "author_name": "Scirpus", "author_url": "", "post_date": "2019-08-13T19:06:50.180000", "content": "Blending gives kernel votes for next to zero work.\nPeople who vote for blends also want to do zero work.\nThis is called synergy - sadly ;)
", "votes": 25, "replies": [] }, { "id": 600266, "author_name": "Marco Vasquez E", "author_url": "", "post_date": "2019-08-15T21:38:54.133000", "content": "The really true when they do that is MEDALS -> Go to better rank -> Go for Linkedin -> #### print(\"I am a Master on Kaggle\")
\n\nand when they have a REAL PROJECT in Data Science.
\n\n---------------------------ERRORs IN RuntimeError, StopIteration, SyntaxError, etc ------------------------------\nZeroDivisionError : division by zero
", "votes": 10, "replies": [] }, { "id": 599930, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-15T13:12:52.243000", "content": "Public blends, and more generally public kernels, aren't an issue as long as everybody can reuse them. The only issue is if someone published a very high quality kernel close to competition end, say during last week, in which case some people may not be able to reuse. It happened in Talking Data competition where someone posted a solid silver medal kernel few hours before end. Many did not have any sub left and saw 600 people pass them. Most silver medals in that competition were illegitimate.
", "votes": 10, "replies": [ { "id": 599980, "author_name": "Firat Gonen", "author_url": "", "post_date": "2019-08-15T13:52:58.417000", "content": "Hey @cpmpml , I started seeing a warning from kaggle not to publish high scoring kernels towards the competition ends, perhaps they started freezing new kernel publishing ?
", "votes": 0, "replies": [] }, { "id": 599988, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-15T13:56:17.453000", "content": "Yes, they have installed this warning after that incident.
", "votes": 2, "replies": [] }, { "id": 600003, "author_name": "Firat Gonen", "author_url": "", "post_date": "2019-08-15T14:26:40.990000", "content": "but just a warning ?
", "votes": 0, "replies": [] }, { "id": 600028, "author_name": "CPMP", "author_url": "", "post_date": "2019-08-15T14:49:51.633000", "content": "Yes, just a warning.
", "votes": 0, "replies": [] }, { "id": 600990, "author_name": "Firat Gonen", "author_url": "", "post_date": "2019-08-16T20:43:05.833000", "content": "a downvote just for this convo, wow :)
", "votes": 0, "replies": [] } ] }, { "id": 598933, "author_name": "Youhan Lee", "author_url": "", "post_date": "2019-08-14T08:39:27.717000", "content": "One can make good output with high score using blending with public kernels. But, Public kernels are really public, so Others can make similar output. I guess It is not easy to get medal by using blended result with public kernels.
\n\nTo get a medal, one need to make own solution.\nCombining the solution and other public kernels, one can get a medal.
", "votes": 6, "replies": [] }, { "id": 598864, "author_name": "Jo Tom", "author_url": "", "post_date": "2019-08-14T06:37:14.240000", "content": "Yeah. I think if you can't get past the blenders you don't deserve the medals. \nGuess I'll do some more workout in the playground and come back to medal challenges later :)
", "votes": 3, "replies": [] }, { "id": 603503, "author_name": "toshi_k", "author_url": "", "post_date": "2019-08-20T11:26:42.087000", "content": "Wow, my SchNet Kernel is also incorporated into blending kernel ๐\n- Chemistry of Best Models (-1.822)
https://www.kaggle.com/roydatascience/chemistry-of-best-models-1-822
\n\n\nWow, my SchNet Kernel is also incorporated into blending kernel ยต
\n
that's ultimate recognition, you can drop from competition now ;)
\n\nJust kidding of course.
", "votes": 1, "replies": [] } ] }, { "id": 602859, "author_name": "RNA", "author_url": "", "post_date": "2019-08-19T15:18:02.617000", "content": "There's some merit to them since blending can be more complex than simply averaging a bunch of different predictions. If I'm blending my own predictions I like to use correlation matrices to identify models that have made distinctly different outputs for example, and select them accordingly while removing clear outliers.
\n\nIt's a dry area of learning frankly, I doubt it's what we come to Kaggle for! But there are certainly easy medals up for grabs - I posted a kernel in the LANL competition with a new approach I'd thought of and it got a gold medal, but looking back I'm certain that most of the votes were simply because the output had a very high Public LB score.
", "votes": 1, "replies": [] }, { "id": 600550, "author_name": "toshi_k", "author_url": "", "post_date": "2019-08-16T09:03:19.900000", "content": "Even if I accept blending kernels, blend of blends are useless. They spoil the reproductivity of kernels and make weights unclear. Authors should use original outputs predicted by ML models as least.
", "votes": 1, "replies": [] }, { "id": 599748, "author_name": "Firat Gonen", "author_url": "", "post_date": "2019-08-15T10:06:31.617000", "content": "I usually blend my models, LGM + NN etc. and sometimes it works sometimes it doesn't . Doing it so, finding the proper weights are important.
\n\nI understand your frustration @felipemello , but as @youhanlee said \"I guess It is not easy to get medal by using blended result with public kernels\"
", "votes": 1, "replies": [] }, { "id": 598925, "author_name": "mochidan", "author_url": "", "post_date": "2019-08-14T08:25:07.663000", "content": "I think blending is also an important technology.\nHowever, this competition has been blending since one month before the deadline. I think it โs too much blending.
", "votes": 1, "replies": [] }, { "id": 599009, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2019-08-14T11:19:38.977000", "content": "It's same in IEEE-CIS Fraud Detection. I guess every tabular data competition is like that.
", "votes": 2, "replies": [] }, { "id": 602873, "author_name": "Shinsei66", "author_url": "", "post_date": "2019-08-19T15:42:34.200000", "content": "Once I improved my own model, then another blend model appears OMG... I wish I could get out of the rat race right away.
", "votes": 0, "replies": [] }, { "id": 600528, "author_name": "", "author_url": "", "post_date": "2019-08-16T08:33:42.747000", "content": "", "votes": 0, "replies": [] }, { "id": 599765, "author_name": "", "author_url": "", "post_date": "2019-08-15T10:33:56.277000", "content": "", "votes": -2, "replies": [] } ], "raw_markdown_by_id": { "598599": "Blending gives kernel votes for next to zero work.\nPeople who vote for blends also want to do zero work.\nThis is called synergy - sadly ;)", "598596": "I find it so silly the amount of people submitting blending kernels or republishing kernels with very small changes or no relevant findings. I know that people are free to do whatever they want, and I endorse it, but what is the motivation behind doing it? Easy medals?", "600266": "The really true when they do that is MEDALS -> Go to better rank -> Go for Linkedin -> #### print(\"I am a Master on Kaggle\") \n\nand when they have a REAL PROJECT in Data Science. \n\n---------------------------ERRORs IN RuntimeError, StopIteration, SyntaxError, etc ------------------------------\nZeroDivisionError : division by zero", "599930": "Public blends, and more generally public kernels, aren't an issue as long as everybody can reuse them. The only issue is if someone published a very high quality kernel close to competition end, say during last week, in which case some people may not be able to reuse. It happened in Talking Data competition where someone posted a solid silver medal kernel few hours before end. Many did not have any sub left and saw 600 people pass them. Most silver medals in that competition were illegitimate.", "598933": "One can make good output with high score using blending with public kernels. But, Public kernels are really public, so Others can make similar output. I guess It is not easy to get medal by using blended result with public kernels.\n\nTo get a medal, one need to make own solution.\nCombining the solution and other public kernels, one can get a medal. ", "598864": "Yeah. I think if you can't get past the blenders you don't deserve the medals. \nGuess I'll do some more workout in the playground and come back to medal challenges later :)", "603503": "Wow, my SchNet Kernel is also incorporated into blending kernel ๐\n- Chemistry of Best Models (-1.822)