Congrats to the prize winners, looking forward to reading about your solution.
\n\nOur solution is rather simple, and the result is incredibly better than I thought. Both Giba and I have the flu, and we had kind of left our hope somewhere given we could not break the 0.508 barrier on the public LB. But this is really a great surprise to us.
\n\nWe'll provide a writeup in the coming days.
", "messages": [ { "id": "268919", "postDate": "01/16/2018 00:08:33", "content": "Congrats to the prize winners, looking forward to reading about your solution.
\n\nOur solution is rather simple, and the result is incredibly better than I thought. Both Giba and I have the flu, and we had kind of left our hope somewhere given we could not break the 0.508 barrier on the public LB. But this is really a great surprise to us.
\n\nWe'll provide a writeup in the coming days.
", "rawMarkdown": "Congrats to the prize winners, looking forward to reading about your solution.\n\nOur solution is rather simple, and the result is incredibly better than I thought. Both Giba and I have the flu, and we had kind of left our hope somewhere given we could not break the 0.508 barrier on the public LB. But this is really a great surprise to us.\n\nWe'll provide a writeup in the coming days.", "votes": null }, { "id": "268922", "postDate": "01/16/2018 00:20:45", "content": "Anyone can explain the reason of such a big shake-up? Did this split have different zero sale promotion distributions?
", "rawMarkdown": "Anyone can explain the reason of such a big shake-up? Did this split have different zero sale promotion distributions?", "votes": null }, { "id": "268923", "postDate": "01/16/2018 00:20:52", "content": "Looking forward to hearing from you !
", "rawMarkdown": "Looking forward to hearing from you !", "votes": null }, { "id": "268925", "postDate": "01/16/2018 00:21:52", "content": "Congrats to you. As I said to my teammates previously, CPMP would definitely improve a lot in the private board
", "rawMarkdown": "Congrats to you. As I said to my teammates previously, CPMP would definitely improve a lot in the private board", "votes": null }, { "id": "268926", "postDate": "01/16/2018 00:22:55", "content": "I think it was easy to overfit to public LB. Remember WTF, how the ranking changed between the first update and the rest. Some models look great on the next few days, and they don't generalize further in the future. You seem to have found models that generalize well, looking forward to learning from what you did! Congrats on the result, very impressive.
", "rawMarkdown": "I think it was easy to overfit to public LB. Remember WTF, how the ranking changed between the first update and the rest. Some models look great on the next few days, and they don't generalize further in the future. You seem to have found models that generalize well, looking forward to learning from what you did! Congrats on the result, very impressive.", "votes": null }, { "id": "268927", "postDate": "01/16/2018 00:26:47", "content": "when you team up with giba, I know you two guys will \"shake up\" not \"shake down\"
", "rawMarkdown": "when you team up with giba, I know you two guys will \"shake up\" not \"shake down\"", "votes": null }, { "id": "268932", "postDate": "01/16/2018 00:34:56", "content": "@AhmetErdem
\n\nI tried some method to deal with the bias for promotion, which is useful for publicboard. But these method make me drop out of gold zone in privateboard
", "rawMarkdown": "AhmetErdem\n\nI tried some method to deal with the bias for promotion, which is useful for publicboard. But these method make me drop out of gold zone in privateboard", "votes": null }, { "id": "268939", "postDate": "01/16/2018 00:45:05", "content": "CPMP - you did really well. I got several good submission files but I missed two of them (got me at least 0.517 in private LB), which during the time I am a bit hesitate to select. Sad story, but I guess picking the good ones to use is a skill indeed...
", "rawMarkdown": "CPMP - you did really well. I got several good submission files but I missed two of them (got me at least 0.517 in private LB), which during the time I am a bit hesitate to select. Sad story, but I guess picking the good ones to use is a skill indeed...", "votes": null }, { "id": "268957", "postDate": "01/16/2018 01:30:43", "content": "I can confirm that trying to deal with missing onpromotion in training actually backfired on us. Doing nothing about it would've got us at least 0.513 private score...
", "rawMarkdown": "I can confirm that trying to deal with missing onpromotion in training actually backfired on us. Doing nothing about it would've got us at least 0.513 private score...", "votes": null }, { "id": "268976", "postDate": "01/16/2018 02:19:05", "content": "@Ceshine - Thank you for your huge contribution to this challenge.
\n\nHow did you try to deal with missing onpromotion information?
\nWhat was the public score for the 'do nothing about it' .513 private score?
I want to know why they overfitted. We can learn a lot from those insight. What they did? What they didn't?
\n\n(I gave up beginning of this competition)
", "rawMarkdown": "I want to know why they overfitted. We can learn a lot from those insight. What they did? What they didn't?\n\n(I gave up beginning of this competition)", "votes": null }, { "id": "269009", "postDate": "01/16/2018 03:47:16", "content": "@Ceshine - I also would like to thank you for the huge contribution. I have learned a bunch.
", "rawMarkdown": "Ceshine - I also would like to thank you for the huge contribution. I have learned a bunch.", "votes": null }, { "id": "269037", "postDate": "01/16/2018 04:52:41", "content": "Time based split will always lead to such shake-up. Especially with holidays.
", "rawMarkdown": "Time based split will always lead to such shake-up. Especially with holidays.", "votes": null }, { "id": "269059", "postDate": "01/16/2018 05:49:02", "content": "Looking forward to the writeup!
", "rawMarkdown": "Looking forward to the writeup!", "votes": null }, { "id": "269104", "postDate": "01/16/2018 07:56:07", "content": "When I looked at the public leader board, the last days, GiBa and You looks stay at 508 so calmly. I thought it was a\nmessage that 508 is enough. And our final best model's public score is also 508
", "rawMarkdown": "When I looked at the public leader board, the last days, GiBa and You looks stay at 508 so calmly. I thought it was a\nmessage that 508 is enough. And our final best model's public score is also 508", "votes": null }, { "id": "269105", "postDate": "01/16/2018 07:58:11", "content": "yep, really crazy - I'm really looking forward to reading the writeups from the top teams (especially how they did CV), since I have a few really funny submission (the \"best\" single model is .517 private and .908 public, have to check my logs what local score I got there; at least the ensemble helped me out)
", "rawMarkdown": "yep, really crazy - I'm really looking forward to reading the writeups from the top teams (especially how they did CV), since I have a few really funny submission (the \"best\" single model is .517 private and .908 public, have to check my logs what local score I got there; at least the ensemble helped me out)", "votes": null }, { "id": "269130", "postDate": "01/16/2018 09:21:31", "content": "@Nicola, the public score of \"do nothing\" model is 0.507.
\n\nThe algorithm was developed by my teammate. The basic idea is to find subset of store/item combos which tend to have the same onpromotion pattern. I did not find it very aggressive. In hindsight, one of the final submissions should be a \"do nothing\" ensemble. We should never 100% rely on a algorithm only verifiable via public score.
\n\nWe switched to onpromotion imputation algorithm a month ago, so it's hard to see how our validation works without it. Will need to do some postmortem analysis. Maybe our CV does not work, and we wouldn't be able to pick the correct submission anyway.
\n\nThank you both for your kind words. My goal for this competition is to practice building DNN models, so I contributed my (watered-down) LGBM baseline as I would't develop it much further anyway. I'm glad that it seems to be helpful to other people.
", "rawMarkdown": "Nicola, the public score of \"do nothing\" model is 0.507.\n\nThe algorithm was developed by my teammate. The basic idea is to find subset of store/item combos which tend to have the same onpromotion pattern. I did not find it very aggressive. In hindsight, one of the final submissions should be a \"do nothing\" ensemble. We should never 100% rely on a algorithm only verifiable via public score.\n\nWe switched to onpromotion imputation algorithm a month ago, so it's hard to see how our validation works without it. Will need to do some postmortem analysis. Maybe our CV does not work, and we wouldn't be able to pick the correct submission anyway.\n\nThank you both for your kind words. My goal for this competition is to practice building DNN models, so I contributed my (watered-down) LGBM baseline as I would't develop it much further anyway. I'm glad that it seems to be helpful to other people.", "votes": null }, { "id": "269140", "postDate": "01/16/2018 09:44:07", "content": "@Ceshine your LGB starter kernel, plus your early comments changed the course of the competition. I frankly was thinking you were sharing too much, and I'm sorry your final result isn't as good as what you could have hoped for. Don't give up, you can get gold anytime with your skills!
", "rawMarkdown": "Ceshine your LGB starter kernel, plus your early comments changed the course of the competition. I frankly was thinking you were sharing too much, and I'm sorry your final result isn't as good as what you could have hoped for. Don't give up, you can get gold anytime with your skills!", "votes": null }, { "id": "269824", "postDate": "01/17/2018 11:11:26", "content": "Thanks you @CPMP for the encouragement.
\n\nThe discussion and public kernel sections of this competition were relatively quite, so I decided to put myself out there for the first time. I tried not to share too much and deprive people of their chance to learn. My apology to anyone who found it otherwise and was affected. (I tend to believe I did not shoot myself in the foot, though.)
\n\nFinally, I'd like to give a shout-out to my teammate @alijs, a very nice and intelligent person, who chose to join my despite of me sharing (part of) my methodology publicly earlier. Learned a lot from him.
", "rawMarkdown": "Thanks you @CPMP for the encouragement. \n\nThe discussion and public kernel sections of this competition were relatively quite, so I decided to put myself out there for the first time. I tried not to share too much and deprive people of their chance to learn. My apology to anyone who found it otherwise and was affected. (I tend to believe I did not shoot myself in the foot, though.)\n\nFinally, I'd like to give a shout-out to my teammate @alijs, a very nice and intelligent person, who chose to join my despite of me sharing (part of) my methodology publicly earlier. Learned a lot from him.", "votes": null }, { "id": "269828", "postDate": "01/17/2018 11:17:19", "content": "I may not have been clear. I don't think you hurt anyone by sharing so much, except for yourself ;) You gave away some edge you had over others, that's all. Sharing, especially after competition end, is great, and one of the reason Kaggle is so popular.
", "rawMarkdown": "I may not have been clear. I don't think you hurt anyone by sharing so much, except for yourself ;) You gave away some edge you had over others, that's all. Sharing, especially after competition end, is great, and one of the reason Kaggle is so popular.", "votes": null } ], "comments": [ { "id": 268922, "author_name": "aerdem4", "author_url": "", "post_date": "01/16/2018 00:20:45", "content": "Anyone can explain the reason of such a big shake-up? Did this split have different zero sale promotion distributions?
", "votes": null, "replies": [ { "id": 268926, "author_name": "cpmpml", "author_url": "", "post_date": "01/16/2018 00:22:55", "content": "I think it was easy to overfit to public LB. Remember WTF, how the ranking changed between the first update and the rest. Some models look great on the next few days, and they don't generalize further in the future. You seem to have found models that generalize well, looking forward to learning from what you did! Congrats on the result, very impressive.
", "votes": null, "replies": [] }, { "id": 268932, "author_name": "h4211819", "author_url": "", "post_date": "01/16/2018 00:34:56", "content": "@AhmetErdem
\n\nI tried some method to deal with the bias for promotion, which is useful for publicboard. But these method make me drop out of gold zone in privateboard
", "votes": null, "replies": [] }, { "id": 268957, "author_name": "ceshine", "author_url": "", "post_date": "01/16/2018 01:30:43", "content": "I can confirm that trying to deal with missing onpromotion in training actually backfired on us. Doing nothing about it would've got us at least 0.513 private score...
", "votes": null, "replies": [] }, { "id": 268976, "author_name": "npa02012", "author_url": "", "post_date": "01/16/2018 02:19:05", "content": "@Ceshine - Thank you for your huge contribution to this challenge.
\n\nHow did you try to deal with missing onpromotion information?
\nWhat was the public score for the 'do nothing about it' .513 private score?
@Ceshine - I also would like to thank you for the huge contribution. I have learned a bunch.
", "votes": null, "replies": [] }, { "id": 269130, "author_name": "ceshine", "author_url": "", "post_date": "01/16/2018 09:21:31", "content": "@Nicola, the public score of \"do nothing\" model is 0.507.
\n\nThe algorithm was developed by my teammate. The basic idea is to find subset of store/item combos which tend to have the same onpromotion pattern. I did not find it very aggressive. In hindsight, one of the final submissions should be a \"do nothing\" ensemble. We should never 100% rely on a algorithm only verifiable via public score.
\n\nWe switched to onpromotion imputation algorithm a month ago, so it's hard to see how our validation works without it. Will need to do some postmortem analysis. Maybe our CV does not work, and we wouldn't be able to pick the correct submission anyway.
\n\nThank you both for your kind words. My goal for this competition is to practice building DNN models, so I contributed my (watered-down) LGBM baseline as I would't develop it much further anyway. I'm glad that it seems to be helpful to other people.
", "votes": null, "replies": [] }, { "id": 269140, "author_name": "cpmpml", "author_url": "", "post_date": "01/16/2018 09:44:07", "content": "@Ceshine your LGB starter kernel, plus your early comments changed the course of the competition. I frankly was thinking you were sharing too much, and I'm sorry your final result isn't as good as what you could have hoped for. Don't give up, you can get gold anytime with your skills!
", "votes": null, "replies": [] }, { "id": 269824, "author_name": "ceshine", "author_url": "", "post_date": "01/17/2018 11:11:26", "content": "Thanks you @CPMP for the encouragement.
\n\nThe discussion and public kernel sections of this competition were relatively quite, so I decided to put myself out there for the first time. I tried not to share too much and deprive people of their chance to learn. My apology to anyone who found it otherwise and was affected. (I tend to believe I did not shoot myself in the foot, though.)
\n\nFinally, I'd like to give a shout-out to my teammate @alijs, a very nice and intelligent person, who chose to join my despite of me sharing (part of) my methodology publicly earlier. Learned a lot from him.
", "votes": null, "replies": [] }, { "id": 269828, "author_name": "cpmpml", "author_url": "", "post_date": "01/17/2018 11:17:19", "content": "I may not have been clear. I don't think you hurt anyone by sharing so much, except for yourself ;) You gave away some edge you had over others, that's all. Sharing, especially after competition end, is great, and one of the reason Kaggle is so popular.
", "votes": null, "replies": [] } ] }, { "id": 268923, "author_name": "paulz11", "author_url": "", "post_date": "01/16/2018 00:20:52", "content": "Looking forward to hearing from you !
", "votes": null, "replies": [] }, { "id": 268925, "author_name": "yiheng", "author_url": "", "post_date": "01/16/2018 00:21:52", "content": "Congrats to you. As I said to my teammates previously, CPMP would definitely improve a lot in the private board
", "votes": null, "replies": [] }, { "id": 268927, "author_name": "nickycan", "author_url": "", "post_date": "01/16/2018 00:26:47", "content": "when you team up with giba, I know you two guys will \"shake up\" not \"shake down\"
", "votes": null, "replies": [] }, { "id": 268939, "author_name": "dongxu027", "author_url": "", "post_date": "01/16/2018 00:45:05", "content": "CPMP - you did really well. I got several good submission files but I missed two of them (got me at least 0.517 in private LB), which during the time I am a bit hesitate to select. Sad story, but I guess picking the good ones to use is a skill indeed...
", "votes": null, "replies": [] }, { "id": 268977, "author_name": "zhilangtaosha", "author_url": "", "post_date": "01/16/2018 02:19:53", "content": "", "votes": null, "replies": [] }, { "id": 268979, "author_name": "onodera", "author_url": "", "post_date": "01/16/2018 02:32:46", "content": "I want to know why they overfitted. We can learn a lot from those insight. What they did? What they didn't?
\n\n(I gave up beginning of this competition)
", "votes": null, "replies": [] }, { "id": 269037, "author_name": "zavodrobotov", "author_url": "", "post_date": "01/16/2018 04:52:41", "content": "Time based split will always lead to such shake-up. Especially with holidays.
", "votes": null, "replies": [] }, { "id": 269059, "author_name": "iolaire", "author_url": "", "post_date": "01/16/2018 05:49:02", "content": "Looking forward to the writeup!
", "votes": null, "replies": [] }, { "id": 269104, "author_name": "luciusyu", "author_url": "", "post_date": "01/16/2018 07:56:07", "content": "When I looked at the public leader board, the last days, GiBa and You looks stay at 508 so calmly. I thought it was a\nmessage that 508 is enough. And our final best model's public score is also 508
", "votes": null, "replies": [] }, { "id": 269105, "author_name": "steelrose", "author_url": "", "post_date": "01/16/2018 07:58:11", "content": "yep, really crazy - I'm really looking forward to reading the writeups from the top teams (especially how they did CV), since I have a few really funny submission (the \"best\" single model is .517 private and .908 public, have to check my logs what local score I got there; at least the ensemble helped me out)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "268919": "Congrats to the prize winners, looking forward to reading about your solution.\n\nOur solution is rather simple, and the result is incredibly better than I thought. Both Giba and I have the flu, and we had kind of left our hope somewhere given we could not break the 0.508 barrier on the public LB. But this is really a great surprise to us.\n\nWe'll provide a writeup in the coming days.", "268922": "Anyone can explain the reason of such a big shake-up? Did this split have different zero sale promotion distributions?", "268923": "Looking forward to hearing from you !", "268925": "Congrats to you. As I said to my teammates previously, CPMP would definitely improve a lot in the private board", "268926": "I think it was easy to overfit to public LB. Remember WTF, how the ranking changed between the first update and the rest. Some models look great on the next few days, and they don't generalize further in the future. You seem to have found models that generalize well, looking forward to learning from what you did! Congrats on the result, very impressive.", "268927": "when you team up with giba, I know you two guys will \"shake up\" not \"shake down\"", "268932": "AhmetErdem\n\nI tried some method to deal with the bias for promotion, which is useful for publicboard. But these method make me drop out of gold zone in privateboard", "268939": "CPMP - you did really well. I got several good submission files but I missed two of them (got me at least 0.517 in private LB), which during the time I am a bit hesitate to select. Sad story, but I guess picking the good ones to use is a skill indeed...", "268957": "I can confirm that trying to deal with missing onpromotion in training actually backfired on us. Doing nothing about it would've got us at least 0.513 private score...", "268976": "Ceshine - Thank you for your huge contribution to this challenge. \n \nHow did you try to deal with missing onpromotion information? \nWhat was the public score for the 'do nothing about it' .513 private score?", "268977": "", "268979": "I want to know why they overfitted. We can learn a lot from those insight. What they did? What they didn't?\n\n(I gave up beginning of this competition)", "269009": "Ceshine - I also would like to thank you for the huge contribution. I have learned a bunch.", "269037": "Time based split will always lead to such shake-up. Especially with holidays.", "269059": "Looking forward to the writeup!", "269104": "When I looked at the public leader board, the last days, GiBa and You looks stay at 508 so calmly. I thought it was a\nmessage that 508 is enough. And our final best model's public score is also 508", "269105": "yep, really crazy - I'm really looking forward to reading the writeups from the top teams (especially how they did CV), since I have a few really funny submission (the \"best\" single model is .517 private and .908 public, have to check my logs what local score I got there; at least the ensemble helped me out)", "269130": "Nicola, the public score of \"do nothing\" model is 0.507.\n\nThe algorithm was developed by my teammate. The basic idea is to find subset of store/item combos which tend to have the same onpromotion pattern. I did not find it very aggressive. In hindsight, one of the final submissions should be a \"do nothing\" ensemble. We should never 100% rely on a algorithm only verifiable via public score.\n\nWe switched to onpromotion imputation algorithm a month ago, so it's hard to see how our validation works without it. Will need to do some postmortem analysis. Maybe our CV does not work, and we wouldn't be able to pick the correct submission anyway.\n\nThank you both for your kind words. My goal for this competition is to practice building DNN models, so I contributed my (watered-down) LGBM baseline as I would't develop it much further anyway. I'm glad that it seems to be helpful to other people.", "269140": "Ceshine your LGB starter kernel, plus your early comments changed the course of the competition. I frankly was thinking you were sharing too much, and I'm sorry your final result isn't as good as what you could have hoped for. Don't give up, you can get gold anytime with your skills!", "269824": "Thanks you @CPMP for the encouragement. \n\nThe discussion and public kernel sections of this competition were relatively quite, so I decided to put myself out there for the first time. I tried not to share too much and deprive people of their chance to learn. My apology to anyone who found it otherwise and was affected. (I tend to believe I did not shoot myself in the foot, though.)\n\nFinally, I'd like to give a shout-out to my teammate @alijs, a very nice and intelligent person, who chose to join my despite of me sharing (part of) my methodology publicly earlier. Learned a lot from him.", "269828": "I may not have been clear. I don't think you hurt anyone by sharing so much, except for yourself ;) You gave away some edge you had over others, that's all. Sharing, especially after competition end, is great, and one of the reason Kaggle is so popular." }, "source": "meta" }