As the competition started to mature, there was a strong movement of the top players to some form of neural network. We committed to LGBM early and never gained any traction with our NN efforts. We managed a decent -2.152, good for 82nd. However, there were many who push LGBM models much farther. Please reply to this if you have a top LGBM or other non-NN score. Also if you have any insight as to how you did it, please comment on your approach.
\n\nThanks!
", "messages": [ { "id": "614045", "postDate": "08/31/2019 03:03:40", "content": "As the competition started to mature, there was a strong movement of the top players to some form of neural network. We committed to LGBM early and never gained any traction with our NN efforts. We managed a decent -2.152, good for 82nd. However, there were many who push LGBM models much farther. Please reply to this if you have a top LGBM or other non-NN score. Also if you have any insight as to how you did it, please comment on your approach.
\n\nThanks!
", "rawMarkdown": "As the competition started to mature, there was a strong movement of the top players to some form of neural network. We committed to LGBM early and never gained any traction with our NN efforts. We managed a decent -2.152, good for 82nd. However, there were many who push LGBM models much farther. Please reply to this if you have a top LGBM or other non-NN score. Also if you have any insight as to how you did it, please comment on your approach.\n\nThanks!", "votes": null }, { "id": "616824", "postDate": "09/03/2019 14:08:09", "content": "@asauve has a nice post competition kernel that looks at how various models performed overall in the competition.\nIt shows how eventually the top of leaderboard became dominated by various NN models.\nhttps://www.kaggle.com/asauve/leaderbord-scores-and-models/notebook
", "rawMarkdown": "asauve has a nice post competition kernel that looks at how various models performed overall in the competition.\nIt shows how eventually the top of leaderboard became dominated by various NN models.\nhttps://www.kaggle.com/asauve/leaderbord-scores-and-models/notebook", "votes": null }, { "id": "616832", "postDate": "09/03/2019 14:14:43", "content": "Thanks to Rob Mulla @robikscube for his excellent discussion on his tabular model and write-up of his solution. His problem approach is great reading for all.
", "rawMarkdown": "Thanks to Rob Mulla @robikscube for his excellent discussion on his tabular model and write-up of his solution. His problem approach is great reading for all.", "votes": null }, { "id": "617328", "postDate": "09/04/2019 02:29:26", "content": "Here is a pointer to another LGBM kernel which scored 54th, with a -2.242\nhttps://www.kaggle.com/jesucristo/single-lgbm-2-242-top54
", "rawMarkdown": "Here is a pointer to another LGBM kernel which scored 54th, with a -2.242\nhttps://www.kaggle.com/jesucristo/single-lgbm-2-242-top54", "votes": null }, { "id": "617857", "postDate": "09/04/2019 14:43:32", "content": "Kernels published after competition end are not to be compared with kernels published before deadline. Indeed, much more info is available after deadline, including private LB score, unlimited number of submissions, and solution share by other teams.
", "rawMarkdown": "Kernels published after competition end are not to be compared with kernels published before deadline. Indeed, much more info is available after deadline, including private LB score, unlimited number of submissions, and solution share by other teams.", "votes": null }, { "id": "617911", "postDate": "09/04/2019 15:53:45", "content": "I intended for people to include the competition results, not post competition.
", "rawMarkdown": "I intended for people to include the competition results, not post competition.", "votes": null }, { "id": "617914", "postDate": "09/04/2019 16:00:29", "content": "That's why I wrote the above. The notebook you point to was published 2 days after competition deadline.
", "rawMarkdown": "That's why I wrote the above. The notebook you point to was published 2 days after competition deadline.", "votes": null }, { "id": "619082", "postDate": "09/05/2019 20:06:05", "content": "Our team used LGBM only. We are at 51-st place.
", "rawMarkdown": "Our team used LGBM only. We are at 51-st place.", "votes": null }, { "id": "619190", "postDate": "09/06/2019 00:11:56", "content": "I didn't realize that. I thought it represented competition results.
", "rawMarkdown": "I didn't realize that. I thought it represented competition results.", "votes": null }, { "id": "619195", "postDate": "09/06/2019 00:16:23", "content": "@sergeyzlobin Nice! -2.268 is the best non-NN score I've seen. What did you use for features?
", "rawMarkdown": "sergeyzlobin Nice! -2.268 is the best non-NN score I've seen. What did you use for features?", "votes": null }, { "id": "619363", "postDate": "09/06/2019 05:56:42", "content": "\n\n\nI thought it represented competition results.
\n
Maybe, maybe not.
\n\nA lot of people publish amazing solutions few days after competition end. The only way to show it represents competition result is to show a screen shot of the submission result during the competition.
\n\nI am not saying he did not submit this during competition. I am just saying we don't know that for sure.
", "rawMarkdown": "> I thought it represented competition results.\n\nMaybe, maybe not.\n\nA lot of people publish amazing solutions few days after competition end. The only way to show it represents competition result is to show a screen shot of the submission result during the competition. \n\nI am not saying he did not submit this during competition. I am just saying we don't know that for sure.", "votes": null }, { "id": "619728", "postDate": "09/06/2019 14:17:40", "content": "@wjholst \nThere are a lot of features actually (my and the partner's). For some types we used ~100 features, and for some types about 250 features. It will take some time to describe them all. :)
\n\nBesides features from public Notebooks we both independently use features like that. Consider the nearest atoms from atom0 or atom1 (my approach) or from their center (Kirill's approach).
\n\nFor example, for type 1JHC one of important features is 'b3_A_bonds'. It is the number of bonds between atom0 and atom that is the third by distance from atom1. I have many geometrical features like angles (e.g. 'a2_BAx_angle') and distances ('b3_B_dist'). Also I used OpenBabel and chemical features, e.g. the ring size for an atom (e.g. 'a1_max_ring_size'), the type of an atom (Open Babel gives the extended type like C2 with hybridization) and mulliken charges from QM9.
\n\nKirill's approach is a little bit different. For example, '1H42_dist' means the distance from the nearest atom to the center of [atom0, atom1] (except both) and that is of 'H' type; '4' and '2' in the end of the feature name are the types of J-couplings (8 known types, zero if other types). He also has cosines of angles and arithmetical combinations of dist and cos, like (1 - 3 * cos**2) - dist.
", "rawMarkdown": "wjholst \nThere are a lot of features actually (my and the partner's). For some types we used ~100 features, and for some types about 250 features. It will take some time to describe them all. :)\n\nBesides features from public Notebooks we both independently use features like that. Consider the nearest atoms from atom0 or atom1 (my approach) or from their center (Kirill's approach). \n\nFor example, for type 1JHC one of important features is 'b3\\_A\\_bonds'. It is the number of bonds between atom0 and atom that is the third by distance from atom1. I have many geometrical features like angles (e.g. 'a2\\_BAx\\_angle') and distances ('b3\\_B\\_dist'). Also I used OpenBabel and chemical features, e.g. the ring size for an atom (e.g. 'a1\\_max\\_ring\\_size'), the type of an atom (Open Babel gives the extended type like C2 with hybridization) and mulliken charges from QM9.\n\nKirill's approach is a little bit different. For example, '1H42\\_dist' means the distance from the nearest atom to the center of [atom0, atom1] (except both) and that is of 'H' type; '4' and '2' in the end of the feature name are the types of J-couplings (8 known types, zero if other types). He also has cosines of angles and arithmetical combinations of dist and cos, like (1 - 3 * cos**2) - dist.", "votes": null }, { "id": "619773", "postDate": "09/06/2019 15:00:48", "content": "Thanks @wjholst - Glad it was helpful for you!
", "rawMarkdown": "Thanks @wjholst - Glad it was helpful for you!", "votes": null }, { "id": "619799", "postDate": "09/06/2019 15:26:33", "content": "The main thing from my features was the generalized representation of all atoms in the molecule respectively to the center of the pair. Each atom was represented as row in dataframe including a name (like ‘3H24’) and a set of features like distance from the center of pair and cosine of angle between radial and line between pair. After pivot by molecule name, atom0, atom1 the name+featurename become a column with value like 3H24dist with the value of distance from 3H24 to the center of couple. Then fillna(0) and the representation is ready.
", "rawMarkdown": "The main thing from my features was the generalized representation of all atoms in the molecule respectively to the center of the pair. Each atom was represented as row in dataframe including a name (like ‘3H24’) and a set of features like distance from the center of pair and cosine of angle between radial and line between pair. After pivot by molecule name, atom0, atom1 the name+featurename become a column with value like 3H24dist with the value of distance from 3H24 to the center of couple. Then fillna(0) and the representation is ready.", "votes": null } ], "comments": [ { "id": 616824, "author_name": "wjholst", "author_url": "", "post_date": "09/03/2019 14:08:09", "content": "@asauve has a nice post competition kernel that looks at how various models performed overall in the competition.\nIt shows how eventually the top of leaderboard became dominated by various NN models.\nhttps://www.kaggle.com/asauve/leaderbord-scores-and-models/notebook
", "votes": null, "replies": [] }, { "id": 616832, "author_name": "wjholst", "author_url": "", "post_date": "09/03/2019 14:14:43", "content": "Thanks to Rob Mulla @robikscube for his excellent discussion on his tabular model and write-up of his solution. His problem approach is great reading for all.
", "votes": null, "replies": [ { "id": 619773, "author_name": "robikscube", "author_url": "", "post_date": "09/06/2019 15:00:48", "content": "Thanks @wjholst - Glad it was helpful for you!
", "votes": null, "replies": [] } ] }, { "id": 617328, "author_name": "wjholst", "author_url": "", "post_date": "09/04/2019 02:29:26", "content": "Here is a pointer to another LGBM kernel which scored 54th, with a -2.242\nhttps://www.kaggle.com/jesucristo/single-lgbm-2-242-top54
", "votes": null, "replies": [ { "id": 617857, "author_name": "cpmpml", "author_url": "", "post_date": "09/04/2019 14:43:32", "content": "Kernels published after competition end are not to be compared with kernels published before deadline. Indeed, much more info is available after deadline, including private LB score, unlimited number of submissions, and solution share by other teams.
", "votes": null, "replies": [] }, { "id": 617911, "author_name": "wjholst", "author_url": "", "post_date": "09/04/2019 15:53:45", "content": "I intended for people to include the competition results, not post competition.
", "votes": null, "replies": [] }, { "id": 617914, "author_name": "cpmpml", "author_url": "", "post_date": "09/04/2019 16:00:29", "content": "That's why I wrote the above. The notebook you point to was published 2 days after competition deadline.
", "votes": null, "replies": [] }, { "id": 619190, "author_name": "wjholst", "author_url": "", "post_date": "09/06/2019 00:11:56", "content": "I didn't realize that. I thought it represented competition results.
", "votes": null, "replies": [] }, { "id": 619363, "author_name": "cpmpml", "author_url": "", "post_date": "09/06/2019 05:56:42", "content": "\n\n\nI thought it represented competition results.
\n
Maybe, maybe not.
\n\nA lot of people publish amazing solutions few days after competition end. The only way to show it represents competition result is to show a screen shot of the submission result during the competition.
\n\nI am not saying he did not submit this during competition. I am just saying we don't know that for sure.
", "votes": null, "replies": [] } ] }, { "id": 619082, "author_name": "sergeyzlobin", "author_url": "", "post_date": "09/05/2019 20:06:05", "content": "Our team used LGBM only. We are at 51-st place.
", "votes": null, "replies": [ { "id": 619195, "author_name": "wjholst", "author_url": "", "post_date": "09/06/2019 00:16:23", "content": "@sergeyzlobin Nice! -2.268 is the best non-NN score I've seen. What did you use for features?
", "votes": null, "replies": [] }, { "id": 619728, "author_name": "sergeyzlobin", "author_url": "", "post_date": "09/06/2019 14:17:40", "content": "@wjholst \nThere are a lot of features actually (my and the partner's). For some types we used ~100 features, and for some types about 250 features. It will take some time to describe them all. :)
\n\nBesides features from public Notebooks we both independently use features like that. Consider the nearest atoms from atom0 or atom1 (my approach) or from their center (Kirill's approach).
\n\nFor example, for type 1JHC one of important features is 'b3_A_bonds'. It is the number of bonds between atom0 and atom that is the third by distance from atom1. I have many geometrical features like angles (e.g. 'a2_BAx_angle') and distances ('b3_B_dist'). Also I used OpenBabel and chemical features, e.g. the ring size for an atom (e.g. 'a1_max_ring_size'), the type of an atom (Open Babel gives the extended type like C2 with hybridization) and mulliken charges from QM9.
\n\nKirill's approach is a little bit different. For example, '1H42_dist' means the distance from the nearest atom to the center of [atom0, atom1] (except both) and that is of 'H' type; '4' and '2' in the end of the feature name are the types of J-couplings (8 known types, zero if other types). He also has cosines of angles and arithmetical combinations of dist and cos, like (1 - 3 * cos**2) - dist.
", "votes": null, "replies": [] }, { "id": 619799, "author_name": "kirill702b", "author_url": "", "post_date": "09/06/2019 15:26:33", "content": "The main thing from my features was the generalized representation of all atoms in the molecule respectively to the center of the pair. Each atom was represented as row in dataframe including a name (like ‘3H24’) and a set of features like distance from the center of pair and cosine of angle between radial and line between pair. After pivot by molecule name, atom0, atom1 the name+featurename become a column with value like 3H24dist with the value of distance from 3H24 to the center of couple. Then fillna(0) and the representation is ready.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "614045": "As the competition started to mature, there was a strong movement of the top players to some form of neural network. We committed to LGBM early and never gained any traction with our NN efforts. We managed a decent -2.152, good for 82nd. However, there were many who push LGBM models much farther. Please reply to this if you have a top LGBM or other non-NN score. Also if you have any insight as to how you did it, please comment on your approach.\n\nThanks!", "616824": "asauve has a nice post competition kernel that looks at how various models performed overall in the competition.\nIt shows how eventually the top of leaderboard became dominated by various NN models.\nhttps://www.kaggle.com/asauve/leaderbord-scores-and-models/notebook", "616832": "Thanks to Rob Mulla @robikscube for his excellent discussion on his tabular model and write-up of his solution. His problem approach is great reading for all.", "617328": "Here is a pointer to another LGBM kernel which scored 54th, with a -2.242\nhttps://www.kaggle.com/jesucristo/single-lgbm-2-242-top54", "617857": "Kernels published after competition end are not to be compared with kernels published before deadline. Indeed, much more info is available after deadline, including private LB score, unlimited number of submissions, and solution share by other teams.", "617911": "I intended for people to include the competition results, not post competition.", "617914": "That's why I wrote the above. The notebook you point to was published 2 days after competition deadline.", "619082": "Our team used LGBM only. We are at 51-st place.", "619190": "I didn't realize that. I thought it represented competition results.", "619195": "sergeyzlobin Nice! -2.268 is the best non-NN score I've seen. What did you use for features?", "619363": "> I thought it represented competition results.\n\nMaybe, maybe not.\n\nA lot of people publish amazing solutions few days after competition end. The only way to show it represents competition result is to show a screen shot of the submission result during the competition. \n\nI am not saying he did not submit this during competition. I am just saying we don't know that for sure.", "619728": "wjholst \nThere are a lot of features actually (my and the partner's). For some types we used ~100 features, and for some types about 250 features. It will take some time to describe them all. :)\n\nBesides features from public Notebooks we both independently use features like that. Consider the nearest atoms from atom0 or atom1 (my approach) or from their center (Kirill's approach). \n\nFor example, for type 1JHC one of important features is 'b3\\_A\\_bonds'. It is the number of bonds between atom0 and atom that is the third by distance from atom1. I have many geometrical features like angles (e.g. 'a2\\_BAx\\_angle') and distances ('b3\\_B\\_dist'). Also I used OpenBabel and chemical features, e.g. the ring size for an atom (e.g. 'a1\\_max\\_ring\\_size'), the type of an atom (Open Babel gives the extended type like C2 with hybridization) and mulliken charges from QM9.\n\nKirill's approach is a little bit different. For example, '1H42\\_dist' means the distance from the nearest atom to the center of [atom0, atom1] (except both) and that is of 'H' type; '4' and '2' in the end of the feature name are the types of J-couplings (8 known types, zero if other types). He also has cosines of angles and arithmetical combinations of dist and cos, like (1 - 3 * cos**2) - dist.", "619773": "Thanks @wjholst - Glad it was helpful for you!", "619799": "The main thing from my features was the generalized representation of all atoms in the molecule respectively to the center of the pair. Each atom was represented as row in dataframe including a name (like ‘3H24’) and a set of features like distance from the center of pair and cosine of angle between radial and line between pair. After pivot by molecule name, atom0, atom1 the name+featurename become a column with value like 3H24dist with the value of distance from 3H24 to the center of couple. Then fillna(0) and the representation is ready." }, "source": "meta" }