{ "id": 96436, "title": "Neighbors are your friends", "url": "/competitions/champs-scalar-coupling/discussion/96436", "author_name": "", "post_date": "2019-06-20T13:40:34.939802600Z", "votes": 39, "comment_count": 27, "views": 0, "content": "

\"neighbors\"

\n\n

As you can see you definitely want to think about adding information from nearest neighbors!

", "messages": [ { "id": "556645", "postDate": "06/20/2019 13:40:34", "content": "

\"neighbors\"

\n\n

As you can see you definitely want to think about adding information from nearest neighbors!

", "rawMarkdown": "![neighbors](https://storage.googleapis.com/kaggle-forum-message-attachments/556645/13602/neighbors.png)\n\nAs you can see you definitely want to think about adding information from nearest neighbors!", "votes": null }, { "id": "556843", "postDate": "06/20/2019 18:35:41", "content": "

I´m exploring KNN too. It´s seem promising.

", "rawMarkdown": "I´m exploring KNN too. It´s seem promising.", "votes": null }, { "id": "556860", "postDate": "06/20/2019 19:06:50", "content": "

I was planning to do the same as well. What type of algorithm are you using if I may ask? I am going to use the following one from sklearn: https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.NearestNeighbors.html.

", "rawMarkdown": "I was planning to do the same as well. What type of algorithm are you using if I may ask? I am going to use the following one from sklearn: https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.NearestNeighbors.html.", "votes": null }, { "id": "556875", "postDate": "06/20/2019 19:24:51", "content": "

This has nothing to do with KNN - just ensure if you have a bond with atom_index_0 of 3 and atom_index_1 of 4 ensure that you add info for atom_index_0 of 1,2,(3),4,5 and atom_index_1 of 2,3,(4,)5,6 if you are using two nearest neighbours (I use over 20).

", "rawMarkdown": "This has nothing to do with KNN - just ensure if you have a bond with atom_index_0 of 3 and atom_index_1 of 4 ensure that you add info for atom_index_0 of 1,2,(3),4,5 and atom_index_1 of 2,3,(4,)5,6 if you are using two nearest neighbours (I use over 20).", "votes": null }, { "id": "556884", "postDate": "06/20/2019 19:48:14", "content": "

Thanks for the clarification. 👍

", "rawMarkdown": "Thanks for the clarification. 👍", "votes": null }, { "id": "556924", "postDate": "06/20/2019 22:16:27", "content": "

Somehow based on Andrew's adversarial validation, topological features are superior to geometry-based features at least for 1JHC (haven't tried for other types yet due to long running time).

", "rawMarkdown": "Somehow based on [Andrew's adversarial validation](https://www.kaggle.com/artgor/validation-feature-selection-interpretation-etc), topological features are superior to geometry-based features at least for 1JHC (haven't tried for other types yet due to long running time).", "votes": null }, { "id": "556951", "postDate": "06/20/2019 23:04:43", "content": "

What's the difference between topological and geometry-based features? I don't quite understand because to me, they mean basically the same thing.

", "rawMarkdown": "What's the difference between topological and geometry-based features? I don't quite understand because to me, they mean basically the same thing.", "votes": null }, { "id": "557175", "postDate": "06/21/2019 01:33:29", "content": "

Maybe I have abused the notion a bit. To me, for example, angles, distances, and their variants like mean or stds are geometric based features. Features like Fiedler eigenvalue I shall refer as topological (invariant with respect to spacial configs).

", "rawMarkdown": "Maybe I have abused the notion a bit. To me, for example, angles, distances, and their variants like mean or stds are geometric based features. Features like Fiedler eigenvalue I shall refer as topological (invariant with respect to spacial configs).", "votes": null }, { "id": "557500", "postDate": "06/21/2019 08:14:59", "content": "

Thanks for the explanation.

\n\n

I've found that angular features perform better for 2JXX and 3JXX couplings.

", "rawMarkdown": "Thanks for the explanation. \n\nI've found that angular features perform better for 2JXX and 3JXX couplings.", "votes": null }, { "id": "557609", "postDate": "06/21/2019 10:05:11", "content": "

@joshxsarah - topology is global structure and geometry is local structure in a mathematical sense

", "rawMarkdown": "joshxsarah - topology is global structure and geometry is local structure in a mathematical sense", "votes": null }, { "id": "558023", "postDate": "06/21/2019 17:47:53", "content": "

Hi, what are the neighbors in this context? neighboring atoms?

", "rawMarkdown": "Hi, what are the neighbors in this context? neighboring atoms?", "votes": null }, { "id": "558165", "postDate": "06/21/2019 22:56:09", "content": "

Hi, are the distance values between atom_0 and atom_1 in your feature list also?

", "rawMarkdown": "Hi, are the distance values between atom_0 and atom_1 in your feature list also?", "votes": null }, { "id": "558338", "postDate": "06/22/2019 06:56:09", "content": "

Yep

", "rawMarkdown": "Yep", "votes": null }, { "id": "558339", "postDate": "06/22/2019 06:56:21", "content": "

Look below

", "rawMarkdown": "Look below", "votes": null }, { "id": "558827", "postDate": "06/23/2019 02:23:59", "content": "

You mention \"bond with atomindex0 of 3 and atomindex1 of 4.\" What do you mean by \"3\" and \"4,\" do you mean molecule 3 and 4? If so, how would you have a bond between two different molecules? Do you mean that you should look at you should look at the atom with the same index in different molecules??

", "rawMarkdown": "You mention \"bond with atomindex0 of 3 and atomindex1 of 4.\" What do you mean by \"3\" and \"4,\" do you mean molecule 3 and 4? If so, how would you have a bond between two different molecules? Do you mean that you should look at you should look at the atom with the same index in different molecules??", "votes": null }, { "id": "558910", "postDate": "06/23/2019 06:47:59", "content": "

Oh dear!\nNo, all indices pertain to each individual molecule

", "rawMarkdown": "Oh dear!\nNo, all indices pertain to each individual molecule", "votes": null }, { "id": "558999", "postDate": "06/23/2019 10:20:47", "content": "

negative and positive distance?\ndo I understand right that you look at the nearest neighbors? but what is negative distance?

", "rawMarkdown": "negative and positive distance?\ndo I understand right that you look at the nearest neighbors? but what is negative distance?", "votes": null }, { "id": "559101", "postDate": "06/23/2019 13:36:07", "content": "

Distances cannot be negative can they?

\n\n

neg just means atom index x is less than atom index y\npos just means atom index x is greater than atom index y

", "rawMarkdown": "Distances cannot be negative can they?\n\nneg just means atom index x is less than atom index y\npos just means atom index x is greater than atom index y", "votes": null }, { "id": "559597", "postDate": "06/24/2019 10:10:36", "content": "

Not only distance, other attributes is also helpful.

", "rawMarkdown": "Not only distance, other attributes is also helpful.", "votes": null }, { "id": "559620", "postDate": "06/24/2019 10:34:37", "content": "

Yep ;)

", "rawMarkdown": "Yep ;)", "votes": null }, { "id": "559699", "postDate": "06/24/2019 13:16:10", "content": "

Thanks for sharing this, Scirpus.

\n\n

I am a beginner and I tend to give reason to things.

\n\n

I have two questions:

\n\n

1) If we are including neighbor information, then, you must decide which feature each neighbor atom's information goes to. To my understanding, you look at the atom index directly in the structures.csv table above and below so you have dist_pos_1, dist_pos_2,... and dist_neg_1,...,

\n\n

Then, did you assume how the atom ordered in each molecule is ordered? I looked at a few molecules, only found that H atoms are after heavy molecules. For example, 'dist_pos_1' always has the distance to the atom above it, but could the atom above it sometimes be far a way from the atom itself?

\n\n

2) How do you compare your way of ordering the features to order the feature by atom distance? i.e., include features like, closest_atom_dist, 2nd_closest_atom_dist, 3rd_closest_atom_dist...

\n\n

Thanks!

", "rawMarkdown": "Thanks for sharing this, Scirpus.\n\nI am a beginner and I tend to give reason to things.\n\nI have two questions:\n\n1) If we are including neighbor information, then, you must decide which feature each neighbor atom's information goes to. To my understanding, you look at the atom index directly in the structures.csv table above and below so you have dist\\_pos\\_1, dist\\_pos\\_2,... and dist\\_neg\\_1,..., \n\nThen, did you assume how the atom ordered in each molecule is ordered? I looked at a few molecules, only found that H atoms are after heavy molecules. For example, 'dist\\_pos\\_1' always has the distance to the atom above it, but could the atom above it sometimes be far a way from the atom itself? \n\n2) How do you compare your way of ordering the features to order the feature by atom distance? i.e., include features like, closest\\_atom\\_dist, 2nd\\_closest\\_atom\\_dist, 3rd\\_closest\\_atom\\_dist...\n\nThanks!", "votes": null }, { "id": "559823", "postDate": "06/24/2019 16:28:39", "content": "

You are indeed correct finding the closest atoms by distance seems sensible - I just used the raw indices because a.) I am lazy and b.) just to illustrate you won't do well if you just use the direct bond only.

\n\n

You will probably get better piece of mind using the distances rather than the indices

", "rawMarkdown": "You are indeed correct finding the closest atoms by distance seems sensible - I just used the raw indices because a.) I am lazy and b.) just to illustrate you won't do well if you just use the direct bond only.\n\nYou will probably get better piece of mind using the distances rather than the indices", "votes": null }, { "id": "560345", "postDate": "06/25/2019 10:26:53", "content": "

This might be of use https://www.kaggle.com/scirpus/spatial-structure

", "rawMarkdown": "This might be of use https://www.kaggle.com/scirpus/spatial-structure", "votes": null }, { "id": "563061", "postDate": "06/27/2019 20:15:05", "content": "

Makes sense, since nearby electronegative atoms decrease J-coupling, or so I've heard.

", "rawMarkdown": "Makes sense, since nearby electronegative atoms decrease J-coupling, [or so I've heard](https://chemistry.stackexchange.com/questions/75664/why-do-electronegative-atoms-depress-coupling-constants).", "votes": null }, { "id": "563596", "postDate": "06/28/2019 13:22:29", "content": "

I'm trying to understand how it works and how to implement this features in my model for 2 days already, but it seems I'm too dumb for it :D

", "rawMarkdown": "I'm trying to understand how it works and how to implement this features in my model for 2 days already, but it seems I'm too dumb for it :D", "votes": null }, { "id": "563632", "postDate": "06/28/2019 14:17:13", "content": "

Download this to get an idea!

\n\n

https://www.kaggle.com/scirpus/stanislav-blinov

", "rawMarkdown": "Download this to get an idea!\n\nhttps://www.kaggle.com/scirpus/stanislav-blinov", "votes": null }, { "id": "563747", "postDate": "06/28/2019 16:23:56", "content": "

Wow thanks for personalized kernel ❤️ \nDefinitely gonna look it through tonight!

", "rawMarkdown": "Wow thanks for personalized kernel ❤️ \nDefinitely gonna look it through tonight!", "votes": null }, { "id": "565548", "postDate": "07/01/2019 05:24:18", "content": "

Hi, good information to know. Can you tell us how much CV score improvement you achieve by generating neighbor features? I have seen a couple of features with high importance but actually lower my CV score in this competition.

", "rawMarkdown": "Hi, good information to know. Can you tell us how much CV score improvement you achieve by generating neighbor features? I have seen a couple of features with high importance but actually lower my CV score in this competition.", "votes": null } ], "comments": [ { "id": 556843, "author_name": "mddiaz", "author_url": "", "post_date": "06/20/2019 18:35:41", "content": "

I´m exploring KNN too. It´s seem promising.

", "votes": null, "replies": [] }, { "id": 556860, "author_name": "yassinealouini", "author_url": "", "post_date": "06/20/2019 19:06:50", "content": "

I was planning to do the same as well. What type of algorithm are you using if I may ask? I am going to use the following one from sklearn: https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.NearestNeighbors.html.

", "votes": null, "replies": [ { "id": 556875, "author_name": "scirpus", "author_url": "", "post_date": "06/20/2019 19:24:51", "content": "

This has nothing to do with KNN - just ensure if you have a bond with atom_index_0 of 3 and atom_index_1 of 4 ensure that you add info for atom_index_0 of 1,2,(3),4,5 and atom_index_1 of 2,3,(4,)5,6 if you are using two nearest neighbours (I use over 20).

", "votes": null, "replies": [] }, { "id": 556884, "author_name": "yassinealouini", "author_url": "", "post_date": "06/20/2019 19:48:14", "content": "

Thanks for the clarification. 👍

", "votes": null, "replies": [] } ] }, { "id": 556924, "author_name": "scaomath", "author_url": "", "post_date": "06/20/2019 22:16:27", "content": "

Somehow based on Andrew's adversarial validation, topological features are superior to geometry-based features at least for 1JHC (haven't tried for other types yet due to long running time).

", "votes": null, "replies": [ { "id": 556951, "author_name": "joshxsarah", "author_url": "", "post_date": "06/20/2019 23:04:43", "content": "

What's the difference between topological and geometry-based features? I don't quite understand because to me, they mean basically the same thing.

", "votes": null, "replies": [] }, { "id": 557175, "author_name": "scaomath", "author_url": "", "post_date": "06/21/2019 01:33:29", "content": "

Maybe I have abused the notion a bit. To me, for example, angles, distances, and their variants like mean or stds are geometric based features. Features like Fiedler eigenvalue I shall refer as topological (invariant with respect to spacial configs).

", "votes": null, "replies": [] }, { "id": 557500, "author_name": "joshxsarah", "author_url": "", "post_date": "06/21/2019 08:14:59", "content": "

Thanks for the explanation.

\n\n

I've found that angular features perform better for 2JXX and 3JXX couplings.

", "votes": null, "replies": [] }, { "id": 557609, "author_name": "scirpus", "author_url": "", "post_date": "06/21/2019 10:05:11", "content": "

@joshxsarah - topology is global structure and geometry is local structure in a mathematical sense

", "votes": null, "replies": [] } ] }, { "id": 558023, "author_name": "sidhanthholalkere", "author_url": "", "post_date": "06/21/2019 17:47:53", "content": "

Hi, what are the neighbors in this context? neighboring atoms?

", "votes": null, "replies": [ { "id": 558339, "author_name": "scirpus", "author_url": "", "post_date": "06/22/2019 06:56:21", "content": "

Look below

", "votes": null, "replies": [] }, { "id": 558827, "author_name": "sidhanthholalkere", "author_url": "", "post_date": "06/23/2019 02:23:59", "content": "

You mention \"bond with atomindex0 of 3 and atomindex1 of 4.\" What do you mean by \"3\" and \"4,\" do you mean molecule 3 and 4? If so, how would you have a bond between two different molecules? Do you mean that you should look at you should look at the atom with the same index in different molecules??

", "votes": null, "replies": [] }, { "id": 558910, "author_name": "scirpus", "author_url": "", "post_date": "06/23/2019 06:47:59", "content": "

Oh dear!\nNo, all indices pertain to each individual molecule

", "votes": null, "replies": [] } ] }, { "id": 558165, "author_name": "everyitfrombit", "author_url": "", "post_date": "06/21/2019 22:56:09", "content": "

Hi, are the distance values between atom_0 and atom_1 in your feature list also?

", "votes": null, "replies": [ { "id": 558338, "author_name": "scirpus", "author_url": "", "post_date": "06/22/2019 06:56:09", "content": "

Yep

", "votes": null, "replies": [] } ] }, { "id": 558999, "author_name": "snovik1975", "author_url": "", "post_date": "06/23/2019 10:20:47", "content": "

negative and positive distance?\ndo I understand right that you look at the nearest neighbors? but what is negative distance?

", "votes": null, "replies": [ { "id": 559101, "author_name": "scirpus", "author_url": "", "post_date": "06/23/2019 13:36:07", "content": "

Distances cannot be negative can they?

\n\n

neg just means atom index x is less than atom index y\npos just means atom index x is greater than atom index y

", "votes": null, "replies": [] }, { "id": 559699, "author_name": "barnwellguy", "author_url": "", "post_date": "06/24/2019 13:16:10", "content": "

Thanks for sharing this, Scirpus.

\n\n

I am a beginner and I tend to give reason to things.

\n\n

I have two questions:

\n\n

1) If we are including neighbor information, then, you must decide which feature each neighbor atom's information goes to. To my understanding, you look at the atom index directly in the structures.csv table above and below so you have dist_pos_1, dist_pos_2,... and dist_neg_1,...,

\n\n

Then, did you assume how the atom ordered in each molecule is ordered? I looked at a few molecules, only found that H atoms are after heavy molecules. For example, 'dist_pos_1' always has the distance to the atom above it, but could the atom above it sometimes be far a way from the atom itself?

\n\n

2) How do you compare your way of ordering the features to order the feature by atom distance? i.e., include features like, closest_atom_dist, 2nd_closest_atom_dist, 3rd_closest_atom_dist...

\n\n

Thanks!

", "votes": null, "replies": [] }, { "id": 559823, "author_name": "scirpus", "author_url": "", "post_date": "06/24/2019 16:28:39", "content": "

You are indeed correct finding the closest atoms by distance seems sensible - I just used the raw indices because a.) I am lazy and b.) just to illustrate you won't do well if you just use the direct bond only.

\n\n

You will probably get better piece of mind using the distances rather than the indices

", "votes": null, "replies": [] }, { "id": 560345, "author_name": "scirpus", "author_url": "", "post_date": "06/25/2019 10:26:53", "content": "

This might be of use https://www.kaggle.com/scirpus/spatial-structure

", "votes": null, "replies": [] } ] }, { "id": 559597, "author_name": "lucas1347", "author_url": "", "post_date": "06/24/2019 10:10:36", "content": "

Not only distance, other attributes is also helpful.

", "votes": null, "replies": [ { "id": 559620, "author_name": "scirpus", "author_url": "", "post_date": "06/24/2019 10:34:37", "content": "

Yep ;)

", "votes": null, "replies": [] } ] }, { "id": 563061, "author_name": "gibbonfreeenergy", "author_url": "", "post_date": "06/27/2019 20:15:05", "content": "

Makes sense, since nearby electronegative atoms decrease J-coupling, or so I've heard.

", "votes": null, "replies": [] }, { "id": 563596, "author_name": "stanislavblinov", "author_url": "", "post_date": "06/28/2019 13:22:29", "content": "

I'm trying to understand how it works and how to implement this features in my model for 2 days already, but it seems I'm too dumb for it :D

", "votes": null, "replies": [ { "id": 563632, "author_name": "scirpus", "author_url": "", "post_date": "06/28/2019 14:17:13", "content": "

Download this to get an idea!

\n\n

https://www.kaggle.com/scirpus/stanislav-blinov

", "votes": null, "replies": [] }, { "id": 563747, "author_name": "stanislavblinov", "author_url": "", "post_date": "06/28/2019 16:23:56", "content": "

Wow thanks for personalized kernel ❤️ \nDefinitely gonna look it through tonight!

", "votes": null, "replies": [] } ] }, { "id": 565548, "author_name": "dondon2019", "author_url": "", "post_date": "07/01/2019 05:24:18", "content": "

Hi, good information to know. Can you tell us how much CV score improvement you achieve by generating neighbor features? I have seen a couple of features with high importance but actually lower my CV score in this competition.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "556645": "![neighbors](https://storage.googleapis.com/kaggle-forum-message-attachments/556645/13602/neighbors.png)\n\nAs you can see you definitely want to think about adding information from nearest neighbors!", "556843": "I´m exploring KNN too. It´s seem promising.", "556860": "I was planning to do the same as well. What type of algorithm are you using if I may ask? I am going to use the following one from sklearn: https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.NearestNeighbors.html.", "556875": "This has nothing to do with KNN - just ensure if you have a bond with atom_index_0 of 3 and atom_index_1 of 4 ensure that you add info for atom_index_0 of 1,2,(3),4,5 and atom_index_1 of 2,3,(4,)5,6 if you are using two nearest neighbours (I use over 20).", "556884": "Thanks for the clarification. 👍", "556924": "Somehow based on [Andrew's adversarial validation](https://www.kaggle.com/artgor/validation-feature-selection-interpretation-etc), topological features are superior to geometry-based features at least for 1JHC (haven't tried for other types yet due to long running time).", "556951": "What's the difference between topological and geometry-based features? I don't quite understand because to me, they mean basically the same thing.", "557175": "Maybe I have abused the notion a bit. To me, for example, angles, distances, and their variants like mean or stds are geometric based features. Features like Fiedler eigenvalue I shall refer as topological (invariant with respect to spacial configs).", "557500": "Thanks for the explanation. \n\nI've found that angular features perform better for 2JXX and 3JXX couplings.", "557609": "joshxsarah - topology is global structure and geometry is local structure in a mathematical sense", "558023": "Hi, what are the neighbors in this context? neighboring atoms?", "558165": "Hi, are the distance values between atom_0 and atom_1 in your feature list also?", "558338": "Yep", "558339": "Look below", "558827": "You mention \"bond with atomindex0 of 3 and atomindex1 of 4.\" What do you mean by \"3\" and \"4,\" do you mean molecule 3 and 4? If so, how would you have a bond between two different molecules? Do you mean that you should look at you should look at the atom with the same index in different molecules??", "558910": "Oh dear!\nNo, all indices pertain to each individual molecule", "558999": "negative and positive distance?\ndo I understand right that you look at the nearest neighbors? but what is negative distance?", "559101": "Distances cannot be negative can they?\n\nneg just means atom index x is less than atom index y\npos just means atom index x is greater than atom index y", "559597": "Not only distance, other attributes is also helpful.", "559620": "Yep ;)", "559699": "Thanks for sharing this, Scirpus.\n\nI am a beginner and I tend to give reason to things.\n\nI have two questions:\n\n1) If we are including neighbor information, then, you must decide which feature each neighbor atom's information goes to. To my understanding, you look at the atom index directly in the structures.csv table above and below so you have dist\\_pos\\_1, dist\\_pos\\_2,... and dist\\_neg\\_1,..., \n\nThen, did you assume how the atom ordered in each molecule is ordered? I looked at a few molecules, only found that H atoms are after heavy molecules. For example, 'dist\\_pos\\_1' always has the distance to the atom above it, but could the atom above it sometimes be far a way from the atom itself? \n\n2) How do you compare your way of ordering the features to order the feature by atom distance? i.e., include features like, closest\\_atom\\_dist, 2nd\\_closest\\_atom\\_dist, 3rd\\_closest\\_atom\\_dist...\n\nThanks!", "559823": "You are indeed correct finding the closest atoms by distance seems sensible - I just used the raw indices because a.) I am lazy and b.) just to illustrate you won't do well if you just use the direct bond only.\n\nYou will probably get better piece of mind using the distances rather than the indices", "560345": "This might be of use https://www.kaggle.com/scirpus/spatial-structure", "563061": "Makes sense, since nearby electronegative atoms decrease J-coupling, [or so I've heard](https://chemistry.stackexchange.com/questions/75664/why-do-electronegative-atoms-depress-coupling-constants).", "563596": "I'm trying to understand how it works and how to implement this features in my model for 2 days already, but it seems I'm too dumb for it :D", "563632": "Download this to get an idea!\n\nhttps://www.kaggle.com/scirpus/stanislav-blinov", "563747": "Wow thanks for personalized kernel ❤️ \nDefinitely gonna look it through tonight!", "565548": "Hi, good information to know. Can you tell us how much CV score improvement you achieve by generating neighbor features? I have seen a couple of features with high importance but actually lower my CV score in this competition." }, "source": "meta" }