Hello,
\n\nI have noticed, that people use even features like \"mean z distance\" between atoms etc. I just want to say that such features are unphysical/unchemical and I wouldn't say this generalize well. I have also noticed a question if the coupling between atom 1 and 2 is the same as between 2 and 1, because you can then mirror your data and have twice as many molecules. In the same way we could generate an arbitrary number of molecules by simple translation or rotation.\nHowever using such data and features, your models will have to learn that translation and rotation do not change the coupling constant. And if a model has to learn more things, the prediction performance is generally worse.\nI would recommend to stick to features like distances and angles etc. which are invariant to translation and rotation.
\n\nJust my two cents :)
", "messages": [ { "id": "548702", "postDate": "06/09/2019 18:23:49", "content": "Hello,
\n\nI have noticed, that people use even features like \"mean z distance\" between atoms etc. I just want to say that such features are unphysical/unchemical and I wouldn't say this generalize well. I have also noticed a question if the coupling between atom 1 and 2 is the same as between 2 and 1, because you can then mirror your data and have twice as many molecules. In the same way we could generate an arbitrary number of molecules by simple translation or rotation.\nHowever using such data and features, your models will have to learn that translation and rotation do not change the coupling constant. And if a model has to learn more things, the prediction performance is generally worse.\nI would recommend to stick to features like distances and angles etc. which are invariant to translation and rotation.
\n\nJust my two cents :)
", "rawMarkdown": "Hello,\n\nI have noticed, that people use even features like \"mean z distance\" between atoms etc. I just want to say that such features are unphysical/unchemical and I wouldn't say this generalize well. I have also noticed a question if the coupling between atom 1 and 2 is the same as between 2 and 1, because you can then mirror your data and have twice as many molecules. In the same way we could generate an arbitrary number of molecules by simple translation or rotation.\nHowever using such data and features, your models will have to learn that translation and rotation do not change the coupling constant. And if a model has to learn more things, the prediction performance is generally worse.\nI would recommend to stick to features like distances and angles etc. which are invariant to translation and rotation.\n\nJust my two cents :)", "votes": null }, { "id": "548704", "postDate": "06/09/2019 18:30:06", "content": "If you mean distance from origin, yeah those features probably aren't useful unless the dataset is biased (which honestly I wouldn't be surprised). But the distance between two atoms is rotation and translation invariant.
", "rawMarkdown": "If you mean distance from origin, yeah those features probably aren't useful unless the dataset is biased (which honestly I wouldn't be surprised). But the distance between two atoms is rotation and translation invariant.", "votes": null }, { "id": "548720", "postDate": "06/09/2019 19:05:25", "content": "Yes, exactly, this is what I meant. If we use distance between two atoms or angle between three atoms, these features are rotationally and translationally invariant, but differences between x/y/z coordinates are not.
", "rawMarkdown": "Yes, exactly, this is what I meant. If we use distance between two atoms or angle between three atoms, these features are rotationally and translationally invariant, but differences between x/y/z coordinates are not.", "votes": null } ], "comments": [ { "id": 548704, "author_name": "matthewmasters", "author_url": "", "post_date": "06/09/2019 18:30:06", "content": "If you mean distance from origin, yeah those features probably aren't useful unless the dataset is biased (which honestly I wouldn't be surprised). But the distance between two atoms is rotation and translation invariant.
", "votes": null, "replies": [ { "id": 548720, "author_name": "somnambwl", "author_url": "", "post_date": "06/09/2019 19:05:25", "content": "Yes, exactly, this is what I meant. If we use distance between two atoms or angle between three atoms, these features are rotationally and translationally invariant, but differences between x/y/z coordinates are not.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "548702": "Hello,\n\nI have noticed, that people use even features like \"mean z distance\" between atoms etc. I just want to say that such features are unphysical/unchemical and I wouldn't say this generalize well. I have also noticed a question if the coupling between atom 1 and 2 is the same as between 2 and 1, because you can then mirror your data and have twice as many molecules. In the same way we could generate an arbitrary number of molecules by simple translation or rotation.\nHowever using such data and features, your models will have to learn that translation and rotation do not change the coupling constant. And if a model has to learn more things, the prediction performance is generally worse.\nI would recommend to stick to features like distances and angles etc. which are invariant to translation and rotation.\n\nJust my two cents :)", "548704": "If you mean distance from origin, yeah those features probably aren't useful unless the dataset is biased (which honestly I wouldn't be surprised). But the distance between two atoms is rotation and translation invariant.", "548720": "Yes, exactly, this is what I meant. If we use distance between two atoms or angle between three atoms, these features are rotationally and translationally invariant, but differences between x/y/z coordinates are not." }, "source": "meta" }