Any hint to get into -0.74 club will be appreciated with upvote straight away. \nI am overfitting to get a better score but hitting the kernel time limit. Both of the practices I am doing are not best and will not be helpful in a long run.
\n\nJust wondering if one could share any idea how they hit -0.74 or less.
Any hint to get into -0.74 club will be appreciated with upvote straight away. \nI am overfitting to get a better score but hitting the kernel time limit. Both of the practices I am doing are not best and will not be helpful in a long run.
\n\nJust wondering if one could share any idea how they hit -0.74 or less.
Haven't tried yet but Coulomb Matrix and Magpie Descriptors are often cited in similar tasks :)
", "rawMarkdown": "Haven't tried yet but Coulomb Matrix and Magpie Descriptors are often cited in similar tasks :)", "votes": null }, { "id": "552169", "postDate": "06/13/2019 15:35:33", "content": "-0.851 on leaderboard.\nSimply take my best kernel and increase number of estimators while training :) Though I doubt it can be done in one kernel due to time limitations.
", "rawMarkdown": "0.851 on leaderboard.\nSimply take my best kernel and increase number of estimators while training :) Though I doubt it can be done in one kernel due to time limitations.", "votes": null }, { "id": "552178", "postDate": "06/13/2019 15:41:54", "content": "Thanks a million Andrew. I am actually doing something similar. Your kernel is diamond. 😺
", "rawMarkdown": "Thanks a million Andrew. I am actually doing something similar. Your kernel is diamond. 😺", "votes": null }, { "id": "552182", "postDate": "06/13/2019 15:45:03", "content": "Thank you :) And that kernel won't be my last kernel for this competition ;)
", "rawMarkdown": "Thank you :) And that kernel won't be my last kernel for this competition ;)", "votes": null }, { "id": "552948", "postDate": "06/14/2019 20:59:04", "content": "@hamzabenchekroun \nThe problem with the Coulomb Matrix is that it's a 29x29 matrix for each atom in each molecule. It is usually flattened to be used as a descriptor but that still makes a vector of shape (1,841) for each atom.
\n\nAnd if you further average it over all atoms, it becomes less useful for such a local property as the scalar coupling.
\n\nWithout memory management and some other tricks, it's difficult to run on the Kaggle servers.
\n\nAnd I'm not so sure about magpie. Magpie descriptors are useful when you don't have any structural information, which is not our case. I think better descriptors can be built with the structural information that we have.
", "rawMarkdown": "hamzabenchekroun \nThe problem with the Coulomb Matrix is that it's a 29x29 matrix for each atom in each molecule. It is usually flattened to be used as a descriptor but that still makes a vector of shape (1,841) for each atom.\n\nAnd if you further average it over all atoms, it becomes less useful for such a local property as the scalar coupling. \n\nWithout memory management and some other tricks, it's difficult to run on the Kaggle servers.\n\nAnd I'm not so sure about magpie. Magpie descriptors are useful when you don't have any structural information, which is not our case. I think better descriptors can be built with the structural information that we have.", "votes": null }, { "id": "552951", "postDate": "06/14/2019 21:21:40", "content": "Have you tried using the eigeen value of the Coulomb Matrix as an input ? I tried it with no success. I also tried the bags of Bonds but it didn't help a lot.
", "rawMarkdown": "Have you tried using the eigeen value of the Coulomb Matrix as an input ? I tried it with no success. I also tried the bags of Bonds but it didn't help a lot.", "votes": null }, { "id": "552955", "postDate": "06/14/2019 21:32:43", "content": "\n\nI started to, yes! But I ended up using ACSF and SOAP. First, the Coulomb matrix is a bit of an old descriptor so that there are more accurate descriptors now, and second, ACSF and SOAP are more memory friendly : I get good accuracies with vectors of shape (1,240) for ACSF and (1,451) for SOAP.
", "rawMarkdown": "areveillon \n\nI started to, yes! But I ended up using ACSF and SOAP. First, the Coulomb matrix is a bit of an old descriptor so that there are more accurate descriptors now, and second, ACSF and SOAP are more memory friendly : I get good accuracies with vectors of shape (1,240) for ACSF and (1,451) for SOAP.", "votes": null }, { "id": "552984", "postDate": "06/14/2019 22:58:16", "content": "I see thanks !
", "rawMarkdown": "I see thanks !", "votes": null }, { "id": "553234", "postDate": "06/15/2019 10:53:01", "content": "Yep, a couple :)\nBut they're all just kinda bruteforcing.\nMy bet is that pure scientific approach will beat any ML-approach in this competition, but i'm just terrible in chemistry...
", "rawMarkdown": "Yep, a couple :)\nBut they're all just kinda bruteforcing.\nMy bet is that pure scientific approach will beat any ML-approach in this competition, but i'm just terrible in chemistry...", "votes": null }, { "id": "553472", "postDate": "06/15/2019 18:56:54", "content": "May I ask what are ACSF and SOAP standing for?
", "rawMarkdown": "May I ask what are ACSF and SOAP standing for?", "votes": null }, { "id": "553515", "postDate": "06/15/2019 21:02:25", "content": "@scaomath \nI wrote a Kernel to explain what ACSF are: https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors
\n\nAs for SOAP, it means Smooth Overlap of Atomic Positions, it is yet another molecular descriptor
", "rawMarkdown": "scaomath \nI wrote a Kernel to explain what ACSF are: https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors\n\nAs for SOAP, it means Smooth Overlap of Atomic Positions, it is yet another molecular descriptor", "votes": null }, { "id": "553529", "postDate": "06/15/2019 21:52:04", "content": "Yes. I did a quick Google, and found your kernel. Upvoted.
", "rawMarkdown": "Yes. I did a quick Google, and found your kernel. Upvoted.", "votes": null }, { "id": "553749", "postDate": "06/16/2019 09:06:03", "content": "I did that with more rounds, 1500 is apparently not enough.
", "rawMarkdown": "I did that with more rounds, 1500 is apparently not enough.", "votes": null }, { "id": "554131", "postDate": "06/17/2019 03:18:30", "content": "you can get -1.xxx by reading through public kernel
", "rawMarkdown": "you can get -1.xxx by reading through public kernel", "votes": null }, { "id": "554139", "postDate": "06/17/2019 04:05:37", "content": "50000 is good number to get early stopping
", "rawMarkdown": "50000 is good number to get early stopping", "votes": null }, { "id": "554455", "postDate": "06/17/2019 12:23:55", "content": "Which one? ;)
", "rawMarkdown": "Which one? ;)", "votes": null }, { "id": "554852", "postDate": "06/18/2019 05:22:07", "content": "with what LR? I never touched early stopping in this competition...
", "rawMarkdown": "with what LR? I never touched early stopping in this competition...", "votes": null }, { "id": "556040", "postDate": "06/19/2019 18:54:19", "content": "Hi @cyberia , check this kernel that I made. I've discovered that the bond calculation can improve the score significantly https://www.kaggle.com/adrianoavelar/bond-calculation-lb-0-82
", "rawMarkdown": "Hi @cyberia , check this kernel that I made. I've discovered that the bond calculation can improve the score significantly https://www.kaggle.com/adrianoavelar/bond-calculation-lb-0-82", "votes": null } ], "comments": [ { "id": 552167, "author_name": "hamzabenchekroun", "author_url": "", "post_date": "06/13/2019 15:30:42", "content": "Haven't tried yet but Coulomb Matrix and Magpie Descriptors are often cited in similar tasks :)
", "votes": null, "replies": [ { "id": 552948, "author_name": "borisdee", "author_url": "", "post_date": "06/14/2019 20:59:04", "content": "@hamzabenchekroun \nThe problem with the Coulomb Matrix is that it's a 29x29 matrix for each atom in each molecule. It is usually flattened to be used as a descriptor but that still makes a vector of shape (1,841) for each atom.
\n\nAnd if you further average it over all atoms, it becomes less useful for such a local property as the scalar coupling.
\n\nWithout memory management and some other tricks, it's difficult to run on the Kaggle servers.
\n\nAnd I'm not so sure about magpie. Magpie descriptors are useful when you don't have any structural information, which is not our case. I think better descriptors can be built with the structural information that we have.
", "votes": null, "replies": [] }, { "id": 552951, "author_name": "areveillon", "author_url": "", "post_date": "06/14/2019 21:21:40", "content": "Have you tried using the eigeen value of the Coulomb Matrix as an input ? I tried it with no success. I also tried the bags of Bonds but it didn't help a lot.
", "votes": null, "replies": [] }, { "id": 552955, "author_name": "borisdee", "author_url": "", "post_date": "06/14/2019 21:32:43", "content": "\n\nI started to, yes! But I ended up using ACSF and SOAP. First, the Coulomb matrix is a bit of an old descriptor so that there are more accurate descriptors now, and second, ACSF and SOAP are more memory friendly : I get good accuracies with vectors of shape (1,240) for ACSF and (1,451) for SOAP.
", "votes": null, "replies": [] }, { "id": 552984, "author_name": "areveillon", "author_url": "", "post_date": "06/14/2019 22:58:16", "content": "I see thanks !
", "votes": null, "replies": [] }, { "id": 553472, "author_name": "scaomath", "author_url": "", "post_date": "06/15/2019 18:56:54", "content": "May I ask what are ACSF and SOAP standing for?
", "votes": null, "replies": [] }, { "id": 553515, "author_name": "borisdee", "author_url": "", "post_date": "06/15/2019 21:02:25", "content": "@scaomath \nI wrote a Kernel to explain what ACSF are: https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors
\n\nAs for SOAP, it means Smooth Overlap of Atomic Positions, it is yet another molecular descriptor
", "votes": null, "replies": [] }, { "id": 553529, "author_name": "scaomath", "author_url": "", "post_date": "06/15/2019 21:52:04", "content": "Yes. I did a quick Google, and found your kernel. Upvoted.
", "votes": null, "replies": [] } ] }, { "id": 552169, "author_name": "artgor", "author_url": "", "post_date": "06/13/2019 15:35:33", "content": "-0.851 on leaderboard.\nSimply take my best kernel and increase number of estimators while training :) Though I doubt it can be done in one kernel due to time limitations.
", "votes": null, "replies": [ { "id": 552178, "author_name": "cyberia", "author_url": "", "post_date": "06/13/2019 15:41:54", "content": "Thanks a million Andrew. I am actually doing something similar. Your kernel is diamond. 😺
", "votes": null, "replies": [] }, { "id": 552182, "author_name": "artgor", "author_url": "", "post_date": "06/13/2019 15:45:03", "content": "Thank you :) And that kernel won't be my last kernel for this competition ;)
", "votes": null, "replies": [] } ] }, { "id": 553234, "author_name": "stanislavblinov", "author_url": "", "post_date": "06/15/2019 10:53:01", "content": "Yep, a couple :)\nBut they're all just kinda bruteforcing.\nMy bet is that pure scientific approach will beat any ML-approach in this competition, but i'm just terrible in chemistry...
", "votes": null, "replies": [] }, { "id": 553749, "author_name": "krismaaa", "author_url": "", "post_date": "06/16/2019 09:06:03", "content": "I did that with more rounds, 1500 is apparently not enough.
", "votes": null, "replies": [ { "id": 554139, "author_name": "pcjimmmy", "author_url": "", "post_date": "06/17/2019 04:05:37", "content": "50000 is good number to get early stopping
", "votes": null, "replies": [] }, { "id": 554852, "author_name": "krismaaa", "author_url": "", "post_date": "06/18/2019 05:22:07", "content": "with what LR? I never touched early stopping in this competition...
", "votes": null, "replies": [] } ] }, { "id": 554131, "author_name": "chauhuynh", "author_url": "", "post_date": "06/17/2019 03:18:30", "content": "you can get -1.xxx by reading through public kernel
", "votes": null, "replies": [ { "id": 554455, "author_name": "stanislavblinov", "author_url": "", "post_date": "06/17/2019 12:23:55", "content": "Which one? ;)
", "votes": null, "replies": [] } ] }, { "id": 556040, "author_name": "adrianoavelar", "author_url": "", "post_date": "06/19/2019 18:54:19", "content": "Hi @cyberia , check this kernel that I made. I've discovered that the bond calculation can improve the score significantly https://www.kaggle.com/adrianoavelar/bond-calculation-lb-0-82
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "552040": "Any hint to get into -0.74 club will be appreciated with upvote straight away. \nI am overfitting to get a better score but hitting the kernel time limit. Both of the practices I am doing are not best and will not be helpful in a long run. \n\nJust wondering if one could share any idea how they hit `-0.74` or less.", "552167": "Haven't tried yet but Coulomb Matrix and Magpie Descriptors are often cited in similar tasks :)", "552169": "0.851 on leaderboard.\nSimply take my best kernel and increase number of estimators while training :) Though I doubt it can be done in one kernel due to time limitations.", "552178": "Thanks a million Andrew. I am actually doing something similar. Your kernel is diamond. 😺", "552182": "Thank you :) And that kernel won't be my last kernel for this competition ;)", "552948": "hamzabenchekroun \nThe problem with the Coulomb Matrix is that it's a 29x29 matrix for each atom in each molecule. It is usually flattened to be used as a descriptor but that still makes a vector of shape (1,841) for each atom.\n\nAnd if you further average it over all atoms, it becomes less useful for such a local property as the scalar coupling. \n\nWithout memory management and some other tricks, it's difficult to run on the Kaggle servers.\n\nAnd I'm not so sure about magpie. Magpie descriptors are useful when you don't have any structural information, which is not our case. I think better descriptors can be built with the structural information that we have.", "552951": "Have you tried using the eigeen value of the Coulomb Matrix as an input ? I tried it with no success. I also tried the bags of Bonds but it didn't help a lot.", "552955": "areveillon \n\nI started to, yes! But I ended up using ACSF and SOAP. First, the Coulomb matrix is a bit of an old descriptor so that there are more accurate descriptors now, and second, ACSF and SOAP are more memory friendly : I get good accuracies with vectors of shape (1,240) for ACSF and (1,451) for SOAP.", "552984": "I see thanks !", "553234": "Yep, a couple :)\nBut they're all just kinda bruteforcing.\nMy bet is that pure scientific approach will beat any ML-approach in this competition, but i'm just terrible in chemistry...", "553472": "May I ask what are ACSF and SOAP standing for?", "553515": "scaomath \nI wrote a Kernel to explain what ACSF are: https://www.kaggle.com/borisdee/predicting-mulliken-charges-with-acsf-descriptors\n\nAs for SOAP, it means Smooth Overlap of Atomic Positions, it is yet another molecular descriptor", "553529": "Yes. I did a quick Google, and found your kernel. Upvoted.", "553749": "I did that with more rounds, 1500 is apparently not enough.", "554131": "you can get -1.xxx by reading through public kernel", "554139": "50000 is good number to get early stopping", "554455": "Which one? ;)", "554852": "with what LR? I never touched early stopping in this competition...", "556040": "Hi @cyberia , check this kernel that I made. I've discovered that the bond calculation can improve the score significantly https://www.kaggle.com/adrianoavelar/bond-calculation-lb-0-82" }, "source": "meta" }