![\"https://i.ibb.co/wzp4YHd/lsbm01.png\"](\"https://i.ibb.co/wzp4YHd/lsbm01.png\")
\nI chose features without looking at public score (only cv).\nThrew everything that did not have a sawtooth structure. About 250 features are left for model training.Well the competition has ended - many will be disappointed by the shakeup but don't be. Now is the time to really learn from the competition - look at Grand Masters insights. Look at your submissions and see what worked and what didn't and try to figure out why.
\n\nOver the next few days kernels and discussion topics will be released that will be pure gold for you and me going forward.
\n\nWe knew early on how tiny the data was in terms of earthquakes and we knew that the error on Public LB was way smaller than what our CVs were telling us.
\n\nFor me my biggest take aways was learning and using MFCC and LSTM which is enough for me.
", "messages": [ { "id": "542949", "postDate": "06/04/2019 08:15:58", "content": "Well the competition has ended - many will be disappointed by the shakeup but don't be. Now is the time to really learn from the competition - look at Grand Masters insights. Look at your submissions and see what worked and what didn't and try to figure out why.
\n\nOver the next few days kernels and discussion topics will be released that will be pure gold for you and me going forward.
\n\nWe knew early on how tiny the data was in terms of earthquakes and we knew that the error on Public LB was way smaller than what our CVs were telling us.
\n\nFor me my biggest take aways was learning and using MFCC and LSTM which is enough for me.
", "rawMarkdown": "Well the competition has ended - many will be disappointed by the shakeup but don't be. Now is the time to really learn from the competition - look at Grand Masters insights. Look at your submissions and see what worked and what didn't and try to figure out why.\n\nOver the next few days kernels and discussion topics will be released that will be pure gold for you and me going forward.\n\nWe knew early on how tiny the data was in terms of earthquakes and we knew that the error on Public LB was way smaller than what our CVs were telling us.\n\nFor me my biggest take aways was learning and using MFCC and LSTM which is enough for me.", "votes": null }, { "id": "542963", "postDate": "06/04/2019 08:33:24", "content": "hi Scirpus, many thanks for your (and Andrew @artgor ) kernel. Actually it works. As a novice i was fail to adapt to information leak from the paper. But, i try to use one of my model (with 109 features i select best from your kernel - with public LB 1459 and private LB 2224 with LGB), with slight modification for late submission, i normalize with standard scaler for each training and testing set, then i multiply the prediction result by factor (testing mean(from paper)/myprediction mean) and boom it give me 2.38 for late submission (gold range). So i think your/Andrew kernel is very nice.
\n\nTo be concluded : \n1. use your/Andrew feature exctraction\n2. normalize\n3. scale up to testing mean
", "rawMarkdown": "hi Scirpus, many thanks for your (and Andrew @artgor ) kernel. Actually it works. As a novice i was fail to adapt to information leak from the paper. But, i try to use one of my model (with 109 features i select best from your kernel - with public LB 1459 and private LB 2224 with LGB), with slight modification for late submission, i normalize with standard scaler for each training and testing set, then i multiply the prediction result by factor (testing mean(from paper)/myprediction mean) and boom it give me 2.38 for late submission (gold range). So i think your/Andrew kernel is very nice.\n\nTo be concluded : \n1. use your/Andrew feature exctraction\n2. normalize\n3. scale up to testing mean", "votes": null }, { "id": "542969", "postDate": "06/04/2019 08:42:17", "content": "\n\n\nmany will be disappointed by the shakeup but don't be
\n
Yeah, I'm disappointed. I've never had such a shake-down. But you've shaked down even more. :(
", "rawMarkdown": "> many will be disappointed by the shakeup but don't be\n\nYeah, I'm disappointed. I've never had such a shake-down. But you've shaked down even more. :(", "votes": null }, { "id": "543001", "postDate": "06/04/2019 09:03:45", "content": "I would have been gutted if the Public mean was similar to the Private mean! The fact you can do extremely well by scaling your outputs by 1.1 based on the academic paper says it all!
", "rawMarkdown": "I would have been gutted if the Public mean was similar to the Private mean! The fact you can do extremely well by scaling your outputs by 1.1 based on the academic paper says it all!", "votes": null }, { "id": "543026", "postDate": "06/04/2019 09:25:42", "content": "Damn, you're right! Multiplying my final submission by 1.1, I get 2.38934 on Private test set (gold zone).\nMultiplying by 1.2 is even better: 2.36674\nThe coefficient can be calculated as @arisukma wrote: testing mean(from paper)/myprediction mean.\nGiba wrote: Private LB mean is 6.7. (I've not looked in details yet).\nIn my case: 6.7 / 5.567 = 1.2
", "rawMarkdown": "Damn, you're right! Multiplying my final submission by 1.1, I get 2.38934 on Private test set (gold zone).\nMultiplying by 1.2 is even better: 2.36674\nThe coefficient can be calculated as @arisukma wrote: testing mean(from paper)/myprediction mean.\nGiba wrote: Private LB mean is 6.7. (I've not looked in details yet).\nIn my case: 6.7 / 5.567 = 1.2", "votes": null }, { "id": "543054", "postDate": "06/04/2019 09:55:16", "content": "LOL - so now we can predict earthquakes as long as the results are published in an academic paper before hand- yay!! Seriously though at least now they can use all of their data to find out the real mean value and just scale there models which may be useful to them. MFCC seems pretty good too!
", "rawMarkdown": "LOL - so now we can predict earthquakes as long as the results are published in an academic paper before hand- yay!! Seriously though at least now they can use all of their data to find out the real mean value and just scale there models which may be useful to them. MFCC seems pretty good too!", "votes": null }, { "id": "543072", "postDate": "06/04/2019 10:09:52", "content": "You're right, learning is what lasts. On my side I learned about MFCC and general additive models.
", "rawMarkdown": "You're right, learning is what lasts. On my side I learned about MFCC and general additive models.", "votes": null }, { "id": "543170", "postDate": "06/04/2019 11:07:57", "content": "I have learn a lot and lot has to be learned in these few days from shared solutions. Kaggle community is great.\nThanks for sharing.
", "rawMarkdown": "I have learn a lot and lot has to be learned in these few days from shared solutions. Kaggle community is great.\nThanks for sharing.", "votes": null }, { "id": "543569", "postDate": "06/04/2019 15:45:15", "content": "Almost the same story:\nIn my case, the multiplier is 1.296314.\nAs a result, private score is changed from 2.58290 (1588 place) to 2.35601 (gold)\n\nI chose features without looking at public score (only cv).\nThrew everything that did not have a sawtooth structure. About 250 features are left for model training.
hi Scirpus, many thanks for your (and Andrew @artgor ) kernel. Actually it works. As a novice i was fail to adapt to information leak from the paper. But, i try to use one of my model (with 109 features i select best from your kernel - with public LB 1459 and private LB 2224 with LGB), with slight modification for late submission, i normalize with standard scaler for each training and testing set, then i multiply the prediction result by factor (testing mean(from paper)/myprediction mean) and boom it give me 2.38 for late submission (gold range). So i think your/Andrew kernel is very nice.
\n\nTo be concluded : \n1. use your/Andrew feature exctraction\n2. normalize\n3. scale up to testing mean
", "votes": null, "replies": [] }, { "id": 542969, "author_name": "sergeyzlobin", "author_url": "", "post_date": "06/04/2019 08:42:17", "content": "\n\n\nmany will be disappointed by the shakeup but don't be
\n
Yeah, I'm disappointed. I've never had such a shake-down. But you've shaked down even more. :(
", "votes": null, "replies": [ { "id": 543001, "author_name": "scirpus", "author_url": "", "post_date": "06/04/2019 09:03:45", "content": "I would have been gutted if the Public mean was similar to the Private mean! The fact you can do extremely well by scaling your outputs by 1.1 based on the academic paper says it all!
", "votes": null, "replies": [] }, { "id": 543026, "author_name": "sergeyzlobin", "author_url": "", "post_date": "06/04/2019 09:25:42", "content": "Damn, you're right! Multiplying my final submission by 1.1, I get 2.38934 on Private test set (gold zone).\nMultiplying by 1.2 is even better: 2.36674\nThe coefficient can be calculated as @arisukma wrote: testing mean(from paper)/myprediction mean.\nGiba wrote: Private LB mean is 6.7. (I've not looked in details yet).\nIn my case: 6.7 / 5.567 = 1.2
", "votes": null, "replies": [] }, { "id": 543054, "author_name": "scirpus", "author_url": "", "post_date": "06/04/2019 09:55:16", "content": "LOL - so now we can predict earthquakes as long as the results are published in an academic paper before hand- yay!! Seriously though at least now they can use all of their data to find out the real mean value and just scale there models which may be useful to them. MFCC seems pretty good too!
", "votes": null, "replies": [] }, { "id": 543569, "author_name": "parushkin", "author_url": "", "post_date": "06/04/2019 15:45:15", "content": "Almost the same story:\nIn my case, the multiplier is 1.296314.\nAs a result, private score is changed from 2.58290 (1588 place) to 2.35601 (gold)\n\nI chose features without looking at public score (only cv).\nThrew everything that did not have a sawtooth structure. About 250 features are left for model training.
You're right, learning is what lasts. On my side I learned about MFCC and general additive models.
", "votes": null, "replies": [] }, { "id": 543170, "author_name": "karanjakhar", "author_url": "", "post_date": "06/04/2019 11:07:57", "content": "I have learn a lot and lot has to be learned in these few days from shared solutions. Kaggle community is great.\nThanks for sharing.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "542949": "Well the competition has ended - many will be disappointed by the shakeup but don't be. Now is the time to really learn from the competition - look at Grand Masters insights. Look at your submissions and see what worked and what didn't and try to figure out why.\n\nOver the next few days kernels and discussion topics will be released that will be pure gold for you and me going forward.\n\nWe knew early on how tiny the data was in terms of earthquakes and we knew that the error on Public LB was way smaller than what our CVs were telling us.\n\nFor me my biggest take aways was learning and using MFCC and LSTM which is enough for me.", "542963": "hi Scirpus, many thanks for your (and Andrew @artgor ) kernel. Actually it works. As a novice i was fail to adapt to information leak from the paper. But, i try to use one of my model (with 109 features i select best from your kernel - with public LB 1459 and private LB 2224 with LGB), with slight modification for late submission, i normalize with standard scaler for each training and testing set, then i multiply the prediction result by factor (testing mean(from paper)/myprediction mean) and boom it give me 2.38 for late submission (gold range). So i think your/Andrew kernel is very nice.\n\nTo be concluded : \n1. use your/Andrew feature exctraction\n2. normalize\n3. scale up to testing mean", "542969": "> many will be disappointed by the shakeup but don't be\n\nYeah, I'm disappointed. I've never had such a shake-down. But you've shaked down even more. :(", "543001": "I would have been gutted if the Public mean was similar to the Private mean! The fact you can do extremely well by scaling your outputs by 1.1 based on the academic paper says it all!", "543026": "Damn, you're right! Multiplying my final submission by 1.1, I get 2.38934 on Private test set (gold zone).\nMultiplying by 1.2 is even better: 2.36674\nThe coefficient can be calculated as @arisukma wrote: testing mean(from paper)/myprediction mean.\nGiba wrote: Private LB mean is 6.7. (I've not looked in details yet).\nIn my case: 6.7 / 5.567 = 1.2", "543054": "LOL - so now we can predict earthquakes as long as the results are published in an academic paper before hand- yay!! Seriously though at least now they can use all of their data to find out the real mean value and just scale there models which may be useful to them. MFCC seems pretty good too!", "543072": "You're right, learning is what lasts. On my side I learned about MFCC and general additive models.", "543170": "I have learn a lot and lot has to be learned in these few days from shared solutions. Kaggle community is great.\nThanks for sharing.", "543569": "Almost the same story:\nIn my case, the multiplier is 1.296314.\nAs a result, private score is changed from 2.58290 (1588 place) to 2.35601 (gold)\n\nI chose features without looking at public score (only cv).\nThrew everything that did not have a sawtooth structure. About 250 features are left for model training." }, "source": "meta" }