{ "id": 93861, "title": "leak or not?", "url": "/competitions/champs-scalar-coupling/discussion/93861", "author_name": "", "post_date": "2019-05-30T16:15:00.221485100Z", "votes": 8, "comment_count": 19, "views": 0, "content": "

0.059?

", "messages": [ { "id": "539863", "postDate": "05/30/2019 16:15:00", "content": "

0.059?

", "rawMarkdown": "0.059?", "votes": null }, { "id": "539874", "postDate": "05/30/2019 16:27:29", "content": "

According to the Evaluation description

\n\n
\n

For this metric, the MAE for any group has a floor of 1e-9, so that the minimum (best) possible score for perfect predictions is approximately -20.7232

\n
\n\n

I don't think it is a leak (yet).

\n\n

IMHO, you should add 1e-9 to your metric kernel\nnp.log(p.max([metrics.mean_absolute_error(y_true, y_pred), 1e-9]))

", "rawMarkdown": "According to the Evaluation description\n> For this metric, the MAE for any group has a floor of 1e-9, so that the minimum (best) possible score for perfect predictions is approximately -20.7232\n\nI don't think it is a leak (yet).\n\nIMHO, you should add 1e-9 to your metric kernel\n`np.log(p.max([metrics.mean_absolute_error(y_true, y_pred), 1e-9]))`", "votes": null }, { "id": "539888", "postDate": "05/30/2019 16:55:37", "content": "

Thats what confuses me. I think the score is too low and I would love to know more form the competitor or admins :)

", "rawMarkdown": "Thats what confuses me. I think the score is too low and I would love to know more form the competitor or admins :)", "votes": null }, { "id": "539949", "postDate": "05/30/2019 18:56:38", "content": "

After looking through his google scholar account, it seems like this competition is his research field :D

", "rawMarkdown": "After looking through his google scholar account, it seems like this competition is his research field :D", "votes": null }, { "id": "539953", "postDate": "05/30/2019 19:03:09", "content": "

looks like we have a domain expert\nhttps://scholar.google.com/citations?user=6ekxr3cAAAAJ&hl=en\nhttps://arxiv.org/abs/1806.05805

", "rawMarkdown": "looks like we have a domain expert\nhttps://scholar.google.com/citations?user=6ekxr3cAAAAJ&hl=en\nhttps://arxiv.org/abs/1806.05805", "votes": null }, { "id": "539964", "postDate": "05/30/2019 19:38:48", "content": "

damn, its time to go back to features

", "rawMarkdown": "damn, its time to go back to features", "votes": null }, { "id": "539967", "postDate": "05/30/2019 19:49:48", "content": "

looking through his papers seems he must be having some prewritten snippets from his previous works

", "rawMarkdown": "looking through his papers seems he must be having some prewritten snippets from his previous works", "votes": null }, { "id": "540057", "postDate": "05/31/2019 00:30:13", "content": "

Honestly, I'm not sure we can rely on the leaderboard for this competition. I expect some people, intentionally or not, will break the rules and use QM9 or some quantum mechanical computations. I get this score is from a domain expert but I am just saying in general.

", "rawMarkdown": "Honestly, I'm not sure we can rely on the leaderboard for this competition. I expect some people, intentionally or not, will break the rules and use QM9 or some quantum mechanical computations. I get this score is from a domain expert but I am just saying in general.", "votes": null }, { "id": "540143", "postDate": "05/31/2019 05:10:03", "content": "

Looks like these papers deal with molecular structures represented as graphs. That's not what we have in this competition - though it should be possible to derive bonds from Cartesian coordinates (it's not trivial, I think, as lenghts of bonds can vary based on various factors, some of them non-local). Any paper describing that part?

", "rawMarkdown": "Looks like these papers deal with molecular structures represented as graphs. That's not what we have in this competition - though it should be possible to derive bonds from Cartesian coordinates (it's not trivial, I think, as lenghts of bonds can vary based on various factors, some of them non-local). Any paper describing that part?", "votes": null }, { "id": "540145", "postDate": "05/31/2019 05:12:29", "content": "

Thanks Andris, both seems alien to me.

", "rawMarkdown": "Thanks Andris, both seems alien to me.", "votes": null }, { "id": "540173", "postDate": "05/31/2019 06:09:21", "content": "

@Andris, I wrote about how you can find bonding information here https://www.kaggle.com/c/champs-scalar-coupling/discussion/93770#539449\nThis should make it easier for you to represent the atoms as graphs

", "rawMarkdown": "Andris, I wrote about how you can find bonding information here https://www.kaggle.com/c/champs-scalar-coupling/discussion/93770#539449\nThis should make it easier for you to represent the atoms as graphs", "votes": null }, { "id": "540289", "postDate": "05/31/2019 09:53:01", "content": "

I think the name of atoms should have been anonymous making it hard to identify the actual molecules so that nobody could use theoretical knowledge IMHO.

", "rawMarkdown": "I think the name of atoms should have been anonymous making it hard to identify the actual molecules so that nobody could use theoretical knowledge IMHO.", "votes": null }, { "id": "540377", "postDate": "05/31/2019 12:28:38", "content": "

I don't think it's a leak. I'm at -.093, and I feel like I'm just barely scratching the surface of the feature engineering possibilities...

", "rawMarkdown": "I don't think it's a leak. I'm at -.093, and I feel like I'm just barely scratching the surface of the feature engineering possibilities...", "votes": null }, { "id": "540399", "postDate": "05/31/2019 13:00:41", "content": "

Thank you very much! Would you like to provide some information on what should we be thinking/doing differently ?

", "rawMarkdown": "Thank you very much! Would you like to provide some information on what should we be thinking/doing differently ?", "votes": null }, { "id": "540413", "postDate": "05/31/2019 13:22:04", "content": "

I may put together a kernel this weekend if I have time. But I'm building in fairly obvious ways on the features published so far in kernels. Also, I have about 60 features I haven't been able to test yet becsuse I've reached the limits of my desktop. Will give them a whirl this weekend when I have time to play in the cloud.

", "rawMarkdown": "I may put together a kernel this weekend if I have time. But I'm building in fairly obvious ways on the features published so far in kernels. Also, I have about 60 features I haven't been able to test yet becsuse I've reached the limits of my desktop. Will give them a whirl this weekend when I have time to play in the cloud.", "votes": null }, { "id": "540456", "postDate": "05/31/2019 14:42:55", "content": "

It is way too easy to determine which atoms they are by the average length of bonds, anyway.

", "rawMarkdown": "It is way too easy to determine which atoms they are by the average length of bonds, anyway.", "votes": null }, { "id": "540460", "postDate": "05/31/2019 14:50:14", "content": "

Thanks, openbabel popped up as the first hit for me as well when I tried to look around! Must be the tool.\nI also tried very crude bonding by comparing distances to sums of [covalent radii] (https://en.wikipedia.org/wiki/Covalent_radius) of both atoms, looked like working, though I am sure there are corner cases, e.g., when some atoms are squeezed by their non-bonded neighbors.

\n\n

openbabel has the benefit of adding much more metadata to the structure, like hybridization (which I guess can influence prediction). One inconvenience is interfacing to it, it's not just Pandas and Numpy anymore... Also not sure how fast that would be, but hey, this is not a kernel-only competition!

", "rawMarkdown": "Thanks, [openbabel](http://openbabel.org/wiki/Main_Page) popped up as the first hit for me as well when I tried to look around! Must be the tool.\nI also tried very crude bonding by comparing distances to sums of [covalent radii] (https://en.wikipedia.org/wiki/Covalent_radius) of both atoms, looked like working, though I am sure there are corner cases, e.g., when some atoms are squeezed by their non-bonded neighbors.\n\nopenbabel has the benefit of adding much more metadata to the structure, like hybridization (which I guess can influence prediction). One inconvenience is interfacing to it, it's not just Pandas and Numpy anymore... Also not sure how fast that would be, but hey, this is not a kernel-only competition!", "votes": null }, { "id": "540622", "postDate": "05/31/2019 18:46:03", "content": "

I'm new to Kaggle and trying data science models. Would using something like openbabel, or other opensource software be considered acceptable? I'm want to make sure I stay in the rules of the game.

", "rawMarkdown": "I'm new to Kaggle and trying data science models. Would using something like openbabel, or other opensource software be considered acceptable? I'm want to make sure I stay in the rules of the game.", "votes": null }, { "id": "540634", "postDate": "05/31/2019 19:22:40", "content": "

@Andris, I have a basic script to read the mol2 file atom and bond info into a pandas dataframe. Anyone, feel free to message me if you'd like it. Apparently discussion posts can only be so long and I don't want to make a kernel just for it.

", "rawMarkdown": "Andris, I have a basic script to read the mol2 file atom and bond info into a pandas dataframe. Anyone, feel free to message me if you'd like it. Apparently discussion posts can only be so long and I don't want to make a kernel just for it.", "votes": null }, { "id": "543905", "postDate": "06/04/2019 23:34:39", "content": "

Hah, it feels like I got involved in a street fight but with a professional boxer...(imagine how exciting it could be if I win the fight!)

", "rawMarkdown": "Hah, it feels like I got involved in a street fight but with a professional boxer...(imagine how exciting it could be if I win the fight!)", "votes": null } ], "comments": [ { "id": 539874, "author_name": "pestipeti", "author_url": "", "post_date": "05/30/2019 16:27:29", "content": "

According to the Evaluation description

\n\n
\n

For this metric, the MAE for any group has a floor of 1e-9, so that the minimum (best) possible score for perfect predictions is approximately -20.7232

\n
\n\n

I don't think it is a leak (yet).

\n\n

IMHO, you should add 1e-9 to your metric kernel\nnp.log(p.max([metrics.mean_absolute_error(y_true, y_pred), 1e-9]))

", "votes": null, "replies": [ { "id": 539888, "author_name": "abhishek", "author_url": "", "post_date": "05/30/2019 16:55:37", "content": "

Thats what confuses me. I think the score is too low and I would love to know more form the competitor or admins :)

", "votes": null, "replies": [] }, { "id": 540377, "author_name": "abendorf", "author_url": "", "post_date": "05/31/2019 12:28:38", "content": "

I don't think it's a leak. I'm at -.093, and I feel like I'm just barely scratching the surface of the feature engineering possibilities...

", "votes": null, "replies": [] }, { "id": 540399, "author_name": "abhishek", "author_url": "", "post_date": "05/31/2019 13:00:41", "content": "

Thank you very much! Would you like to provide some information on what should we be thinking/doing differently ?

", "votes": null, "replies": [] }, { "id": 540413, "author_name": "abendorf", "author_url": "", "post_date": "05/31/2019 13:22:04", "content": "

I may put together a kernel this weekend if I have time. But I'm building in fairly obvious ways on the features published so far in kernels. Also, I have about 60 features I haven't been able to test yet becsuse I've reached the limits of my desktop. Will give them a whirl this weekend when I have time to play in the cloud.

", "votes": null, "replies": [] } ] }, { "id": 539949, "author_name": "ilhamfp31", "author_url": "", "post_date": "05/30/2019 18:56:38", "content": "

After looking through his google scholar account, it seems like this competition is his research field :D

", "votes": null, "replies": [ { "id": 543905, "author_name": "segovia", "author_url": "", "post_date": "06/04/2019 23:34:39", "content": "

Hah, it feels like I got involved in a street fight but with a professional boxer...(imagine how exciting it could be if I win the fight!)

", "votes": null, "replies": [] } ] }, { "id": 539953, "author_name": "mks2192", "author_url": "", "post_date": "05/30/2019 19:03:09", "content": "

looks like we have a domain expert\nhttps://scholar.google.com/citations?user=6ekxr3cAAAAJ&hl=en\nhttps://arxiv.org/abs/1806.05805

", "votes": null, "replies": [ { "id": 539964, "author_name": "abhishek", "author_url": "", "post_date": "05/30/2019 19:38:48", "content": "

damn, its time to go back to features

", "votes": null, "replies": [] }, { "id": 539967, "author_name": "mks2192", "author_url": "", "post_date": "05/30/2019 19:49:48", "content": "

looking through his papers seems he must be having some prewritten snippets from his previous works

", "votes": null, "replies": [] }, { "id": 540143, "author_name": "andris", "author_url": "", "post_date": "05/31/2019 05:10:03", "content": "

Looks like these papers deal with molecular structures represented as graphs. That's not what we have in this competition - though it should be possible to derive bonds from Cartesian coordinates (it's not trivial, I think, as lenghts of bonds can vary based on various factors, some of them non-local). Any paper describing that part?

", "votes": null, "replies": [] }, { "id": 540145, "author_name": "mks2192", "author_url": "", "post_date": "05/31/2019 05:12:29", "content": "

Thanks Andris, both seems alien to me.

", "votes": null, "replies": [] }, { "id": 540173, "author_name": "matthewmasters", "author_url": "", "post_date": "05/31/2019 06:09:21", "content": "

@Andris, I wrote about how you can find bonding information here https://www.kaggle.com/c/champs-scalar-coupling/discussion/93770#539449\nThis should make it easier for you to represent the atoms as graphs

", "votes": null, "replies": [] }, { "id": 540460, "author_name": "andris", "author_url": "", "post_date": "05/31/2019 14:50:14", "content": "

Thanks, openbabel popped up as the first hit for me as well when I tried to look around! Must be the tool.\nI also tried very crude bonding by comparing distances to sums of [covalent radii] (https://en.wikipedia.org/wiki/Covalent_radius) of both atoms, looked like working, though I am sure there are corner cases, e.g., when some atoms are squeezed by their non-bonded neighbors.

\n\n

openbabel has the benefit of adding much more metadata to the structure, like hybridization (which I guess can influence prediction). One inconvenience is interfacing to it, it's not just Pandas and Numpy anymore... Also not sure how fast that would be, but hey, this is not a kernel-only competition!

", "votes": null, "replies": [] }, { "id": 540622, "author_name": "jjohnsumn", "author_url": "", "post_date": "05/31/2019 18:46:03", "content": "

I'm new to Kaggle and trying data science models. Would using something like openbabel, or other opensource software be considered acceptable? I'm want to make sure I stay in the rules of the game.

", "votes": null, "replies": [] }, { "id": 540634, "author_name": "matthewmasters", "author_url": "", "post_date": "05/31/2019 19:22:40", "content": "

@Andris, I have a basic script to read the mol2 file atom and bond info into a pandas dataframe. Anyone, feel free to message me if you'd like it. Apparently discussion posts can only be so long and I don't want to make a kernel just for it.

", "votes": null, "replies": [] } ] }, { "id": 540057, "author_name": "matthewmasters", "author_url": "", "post_date": "05/31/2019 00:30:13", "content": "

Honestly, I'm not sure we can rely on the leaderboard for this competition. I expect some people, intentionally or not, will break the rules and use QM9 or some quantum mechanical computations. I get this score is from a domain expert but I am just saying in general.

", "votes": null, "replies": [ { "id": 540289, "author_name": "mks2192", "author_url": "", "post_date": "05/31/2019 09:53:01", "content": "

I think the name of atoms should have been anonymous making it hard to identify the actual molecules so that nobody could use theoretical knowledge IMHO.

", "votes": null, "replies": [] }, { "id": 540456, "author_name": "andris", "author_url": "", "post_date": "05/31/2019 14:42:55", "content": "

It is way too easy to determine which atoms they are by the average length of bonds, anyway.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "539863": "0.059?", "539874": "According to the Evaluation description\n> For this metric, the MAE for any group has a floor of 1e-9, so that the minimum (best) possible score for perfect predictions is approximately -20.7232\n\nI don't think it is a leak (yet).\n\nIMHO, you should add 1e-9 to your metric kernel\n`np.log(p.max([metrics.mean_absolute_error(y_true, y_pred), 1e-9]))`", "539888": "Thats what confuses me. I think the score is too low and I would love to know more form the competitor or admins :)", "539949": "After looking through his google scholar account, it seems like this competition is his research field :D", "539953": "looks like we have a domain expert\nhttps://scholar.google.com/citations?user=6ekxr3cAAAAJ&hl=en\nhttps://arxiv.org/abs/1806.05805", "539964": "damn, its time to go back to features", "539967": "looking through his papers seems he must be having some prewritten snippets from his previous works", "540057": "Honestly, I'm not sure we can rely on the leaderboard for this competition. I expect some people, intentionally or not, will break the rules and use QM9 or some quantum mechanical computations. I get this score is from a domain expert but I am just saying in general.", "540143": "Looks like these papers deal with molecular structures represented as graphs. That's not what we have in this competition - though it should be possible to derive bonds from Cartesian coordinates (it's not trivial, I think, as lenghts of bonds can vary based on various factors, some of them non-local). Any paper describing that part?", "540145": "Thanks Andris, both seems alien to me.", "540173": "Andris, I wrote about how you can find bonding information here https://www.kaggle.com/c/champs-scalar-coupling/discussion/93770#539449\nThis should make it easier for you to represent the atoms as graphs", "540289": "I think the name of atoms should have been anonymous making it hard to identify the actual molecules so that nobody could use theoretical knowledge IMHO.", "540377": "I don't think it's a leak. I'm at -.093, and I feel like I'm just barely scratching the surface of the feature engineering possibilities...", "540399": "Thank you very much! Would you like to provide some information on what should we be thinking/doing differently ?", "540413": "I may put together a kernel this weekend if I have time. But I'm building in fairly obvious ways on the features published so far in kernels. Also, I have about 60 features I haven't been able to test yet becsuse I've reached the limits of my desktop. Will give them a whirl this weekend when I have time to play in the cloud.", "540456": "It is way too easy to determine which atoms they are by the average length of bonds, anyway.", "540460": "Thanks, [openbabel](http://openbabel.org/wiki/Main_Page) popped up as the first hit for me as well when I tried to look around! Must be the tool.\nI also tried very crude bonding by comparing distances to sums of [covalent radii] (https://en.wikipedia.org/wiki/Covalent_radius) of both atoms, looked like working, though I am sure there are corner cases, e.g., when some atoms are squeezed by their non-bonded neighbors.\n\nopenbabel has the benefit of adding much more metadata to the structure, like hybridization (which I guess can influence prediction). One inconvenience is interfacing to it, it's not just Pandas and Numpy anymore... Also not sure how fast that would be, but hey, this is not a kernel-only competition!", "540622": "I'm new to Kaggle and trying data science models. Would using something like openbabel, or other opensource software be considered acceptable? I'm want to make sure I stay in the rules of the game.", "540634": "Andris, I have a basic script to read the mol2 file atom and bond info into a pandas dataframe. Anyone, feel free to message me if you'd like it. Apparently discussion posts can only be so long and I don't want to make a kernel just for it.", "543905": "Hah, it feels like I got involved in a street fight but with a professional boxer...(imagine how exciting it could be if I win the fight!)" }, "source": "meta" }