{
  "id": 104241,
  "title": "Is 1JHC really one class",
  "url": "/competitions/champs-scalar-coupling/discussion/104241",
  "author_name": "",
  "post_date": "2019-08-15T12:59:46.057222500Z",
  "votes": 18,
  "comment_count": 17,
  "views": 0,
  "content": "<p>Messing about with my GP program 1JHC doesn't really look like a single class\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F241438%2F480b4f3d892166a660833992aeb4d01e%2F1JHC2Classes.png?generation=1565873976854009&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "599919",
      "postDate": "08/15/2019 12:59:46",
      "content": "<p>Messing about with my GP program 1JHC doesn't really look like a single class\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F241438%2F480b4f3d892166a660833992aeb4d01e%2F1JHC2Classes.png?generation=1565873976854009&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "Messing about with my GP program 1JHC doesn't really look like a single class\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F241438%2F480b4f3d892166a660833992aeb4d01e%2F1JHC2Classes.png?generation=1565873976854009&amp;alt=media)",
      "votes": null
    },
    {
      "id": "599921",
      "postDate": "08/15/2019 13:01:34",
      "content": "<p>Anyone had luck at classifying and then training a regressor model?</p>",
      "rawMarkdown": "Anyone had luck at classifying and then training a regressor model?",
      "votes": null
    },
    {
      "id": "599934",
      "postDate": "08/15/2019 13:19:27",
      "content": "<p>I don;t really understand what I am looking at here, can you explain a bit? \nI would suspect you might be able to split by the \"type\" of the Carbon atom in the interaction.</p>",
      "rawMarkdown": "I don;t really understand what I am looking at here, can you explain a bit? \nI would suspect you might be able to split by the \"type\" of the Carbon atom in the interaction.",
      "votes": null
    },
    {
      "id": "599939",
      "postDate": "08/15/2019 13:24:52",
      "content": "<p>There is a clear split around 175 for scalar coupling (Outputs Graph) so if you add average scalar coupling per atom for example it is going to be sub optimal</p>",
      "rawMarkdown": "There is a clear split around 175 for scalar coupling (Outputs Graph) so if you add average scalar coupling per atom for example it is going to be sub optimal",
      "votes": null
    },
    {
      "id": "600094",
      "postDate": "08/15/2019 16:24:31",
      "content": "<p>I tried a stratifiedkfold split after \"binning\" (using pd.cut and pd.qcut) the y. \nI'd guess it simulates kind of a classification. But my nn didn't perform better than a random split.</p>\n\n<p>I also used the bins to blend my three models using xgb. But the feature importance of the bins is very low. </p>",
      "rawMarkdown": "I tried a stratifiedkfold split after \"binning\" (using pd.cut and pd.qcut) the y. \nI'd guess it simulates kind of a classification. But my nn didn't perform better than a random split.\n\nI also used the bins to blend my three models using xgb. But the feature importance of the bins is very low.",
      "votes": null
    },
    {
      "id": "600672",
      "postDate": "08/16/2019 12:47:30",
      "content": "<p>Carbons with sp/sp2 hybridization (connected to other atoms with 4 or 3 bonds) and other subclasses will have different ranges of coupling constants. If you adjust for this, I would assume that you won't see the same splitting in your model.</p>",
      "rawMarkdown": "Carbons with sp/sp2 hybridization (connected to other atoms with 4 or 3 bonds) and other subclasses will have different ranges of coupling constants. If you adjust for this, I would assume that you won't see the same splitting in your model.",
      "votes": null
    },
    {
      "id": "600962",
      "postDate": "08/16/2019 20:05:01",
      "content": "<p>Forgive me but the output I show here is from your files not from my model ;)</p>",
      "rawMarkdown": "Forgive me but the output I show here is from your files not from my model ;)",
      "votes": null
    },
    {
      "id": "600975",
      "postDate": "08/16/2019 20:15:30",
      "content": "<p>The two groups are easily splited setting a threshold in 1J coupling distance to 1.065.</p>",
      "rawMarkdown": "The two groups are easily splited setting a threshold in 1J coupling distance to 1.065.",
      "votes": null
    },
    {
      "id": "600977",
      "postDate": "08/16/2019 20:19:23",
      "content": "<p>Thanks - we really don't deserve you!</p>",
      "rawMarkdown": "Thanks - we really don't deserve you!",
      "votes": null
    },
    {
      "id": "601041",
      "postDate": "08/16/2019 23:48:55",
      "content": "<blockquote>\n  <p>The two groups are easily spited setting a threshold in 1J coupling distance to 1.075.</p>\n</blockquote>\n\n<p><a href=\"/titericz\">@titericz</a>, I am not sure I understand your post. If possible, explanation please?</p>",
      "rawMarkdown": "&gt; The two groups are easily spited setting a threshold in 1J coupling distance to 1.075.\n\n@titericz, I am not sure I understand your post. If possible, explanation please?",
      "votes": null
    },
    {
      "id": "601064",
      "postDate": "08/17/2019 01:02:38",
      "content": "<p>Calculate the distance of atom_index_0 and atom_index_1 for 1JHC couplings. If that distance is &lt; 1.065 then scalar_coupling_constant is on the top (200 value) of Scirpus plot. If distance &gt; 1.065 then scc is on the botton of the plot.</p>",
      "rawMarkdown": "Calculate the distance of atom_index_0 and atom_index_1 for 1JHC couplings. If that distance is &lt; 1.065 then scalar_coupling_constant is on the top (200 value) of Scirpus plot. If distance &gt; 1.065 then scc is on the botton of the plot.",
      "votes": null
    },
    {
      "id": "601075",
      "postDate": "08/17/2019 01:29:24",
      "content": "<p>Thanks <a href=\"/titericz\">@titericz</a>.</p>",
      "rawMarkdown": "Thanks @titericz.",
      "votes": null
    },
    {
      "id": "602857",
      "postDate": "08/19/2019 15:16:32",
      "content": "<p>Oh. The same reasoning applies though. <a href=\"/titericz\">@titericz</a>'s observation that you can split by bond distance agrees well with this as well.</p>",
      "rawMarkdown": "Oh. The same reasoning applies though. @titericz's observation that you can split by bond distance agrees well with this as well.",
      "votes": null
    },
    {
      "id": "605081",
      "postDate": "08/22/2019 03:24:26",
      "content": "<p>All the 1JHC bonds with <code>sc_constant</code> around 200 begin with hydrogen and carbon followed by 2 carbons in a straight line. Below are two examples.\n<img src=\"http://playagricola.com/Kaggle/ani7.gif\" alt=\"image\">\n<img src=\"http://playagricola.com/Kaggle/ani8.gif\" alt=\"image\"></p>",
      "rawMarkdown": "All the 1JHC bonds with `sc_constant` around 200 begin with hydrogen and carbon followed by 2 carbons in a straight line. Below are two examples.\n![image](http://playagricola.com/Kaggle/ani7.gif)\n![image](http://playagricola.com/Kaggle/ani8.gif)",
      "votes": null
    },
    {
      "id": "605239",
      "postDate": "08/22/2019 07:31:52",
      "content": "<p>Nice to see you jumping on board, Chris! Nice to see your observations again, I love 'em! Wish you luck in the upcoming week :)</p>",
      "rawMarkdown": "Nice to see you jumping on board, Chris! Nice to see your observations again, I love 'em! Wish you luck in the upcoming week :)",
      "votes": null
    },
    {
      "id": "605311",
      "postDate": "08/22/2019 08:41:34",
      "content": "<blockquote>\n  <p>Nice to see you jumping on board, Chris!</p>\n</blockquote>\n\n<p>His first sub is from a month ago here ;)</p>\n\n<p>But I guess he's been busy with other competitions in the meantime.</p>",
      "rawMarkdown": "&gt; Nice to see you jumping on board, Chris!\n\nHis first sub is from a month ago here ;)\n\nBut I guess he's been busy with other competitions in the meantime.",
      "votes": null
    },
    {
      "id": "605432",
      "postDate": "08/22/2019 11:22:20",
      "content": "<p>Terminal triple bond: H:C:::C:Some_atom. All four atoms are always on the same line.</p>",
      "rawMarkdown": "Terminal triple bond: H:C:::C:Some_atom. All four atoms are always on the same line.",
      "votes": null
    },
    {
      "id": "606224",
      "postDate": "08/23/2019 10:46:26",
      "content": "<p><a href=\"https://www.kaggle.com/somnambwl/1jhc-split\">Here</a> is the kernel with graphs.</p>",
      "rawMarkdown": "[Here](https://www.kaggle.com/somnambwl/1jhc-split) is the kernel with graphs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 599921,
      "author_name": "scirpus",
      "author_url": "",
      "post_date": "08/15/2019 13:01:34",
      "content": "<p>Anyone had luck at classifying and then training a regressor model?</p>",
      "votes": null,
      "replies": [
        {
          "id": 600094,
          "author_name": "joatom",
          "author_url": "",
          "post_date": "08/15/2019 16:24:31",
          "content": "<p>I tried a stratifiedkfold split after \"binning\" (using pd.cut and pd.qcut) the y. \nI'd guess it simulates kind of a classification. But my nn didn't perform better than a random split.</p>\n\n<p>I also used the bins to blend my three models using xgb. But the feature importance of the bins is very low. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 599934,
      "author_name": "fnands",
      "author_url": "",
      "post_date": "08/15/2019 13:19:27",
      "content": "<p>I don;t really understand what I am looking at here, can you explain a bit? \nI would suspect you might be able to split by the \"type\" of the Carbon atom in the interaction.</p>",
      "votes": null,
      "replies": [
        {
          "id": 599939,
          "author_name": "scirpus",
          "author_url": "",
          "post_date": "08/15/2019 13:24:52",
          "content": "<p>There is a clear split around 175 for scalar coupling (Outputs Graph) so if you add average scalar coupling per atom for example it is going to be sub optimal</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 600672,
          "author_name": "larsbratholm",
          "author_url": "",
          "post_date": "08/16/2019 12:47:30",
          "content": "<p>Carbons with sp/sp2 hybridization (connected to other atoms with 4 or 3 bonds) and other subclasses will have different ranges of coupling constants. If you adjust for this, I would assume that you won't see the same splitting in your model.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 600962,
          "author_name": "scirpus",
          "author_url": "",
          "post_date": "08/16/2019 20:05:01",
          "content": "<p>Forgive me but the output I show here is from your files not from my model ;)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 602857,
          "author_name": "larsbratholm",
          "author_url": "",
          "post_date": "08/19/2019 15:16:32",
          "content": "<p>Oh. The same reasoning applies though. <a href=\"/titericz\">@titericz</a>'s observation that you can split by bond distance agrees well with this as well.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 600975,
      "author_name": "titericz",
      "author_url": "",
      "post_date": "08/16/2019 20:15:30",
      "content": "<p>The two groups are easily splited setting a threshold in 1J coupling distance to 1.065.</p>",
      "votes": null,
      "replies": [
        {
          "id": 600977,
          "author_name": "scirpus",
          "author_url": "",
          "post_date": "08/16/2019 20:19:23",
          "content": "<p>Thanks - we really don't deserve you!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 601041,
          "author_name": "sheriytm",
          "author_url": "",
          "post_date": "08/16/2019 23:48:55",
          "content": "<blockquote>\n  <p>The two groups are easily spited setting a threshold in 1J coupling distance to 1.075.</p>\n</blockquote>\n\n<p><a href=\"/titericz\">@titericz</a>, I am not sure I understand your post. If possible, explanation please?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 601064,
          "author_name": "titericz",
          "author_url": "",
          "post_date": "08/17/2019 01:02:38",
          "content": "<p>Calculate the distance of atom_index_0 and atom_index_1 for 1JHC couplings. If that distance is &lt; 1.065 then scalar_coupling_constant is on the top (200 value) of Scirpus plot. If distance &gt; 1.065 then scc is on the botton of the plot.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 601075,
          "author_name": "sheriytm",
          "author_url": "",
          "post_date": "08/17/2019 01:29:24",
          "content": "<p>Thanks <a href=\"/titericz\">@titericz</a>.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 606224,
          "author_name": "somnambwl",
          "author_url": "",
          "post_date": "08/23/2019 10:46:26",
          "content": "<p><a href=\"https://www.kaggle.com/somnambwl/1jhc-split\">Here</a> is the kernel with graphs.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 605081,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "08/22/2019 03:24:26",
      "content": "<p>All the 1JHC bonds with <code>sc_constant</code> around 200 begin with hydrogen and carbon followed by 2 carbons in a straight line. Below are two examples.\n<img src=\"http://playagricola.com/Kaggle/ani7.gif\" alt=\"image\">\n<img src=\"http://playagricola.com/Kaggle/ani8.gif\" alt=\"image\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 605239,
          "author_name": "stanislavblinov",
          "author_url": "",
          "post_date": "08/22/2019 07:31:52",
          "content": "<p>Nice to see you jumping on board, Chris! Nice to see your observations again, I love 'em! Wish you luck in the upcoming week :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 605311,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "08/22/2019 08:41:34",
          "content": "<blockquote>\n  <p>Nice to see you jumping on board, Chris!</p>\n</blockquote>\n\n<p>His first sub is from a month ago here ;)</p>\n\n<p>But I guess he's been busy with other competitions in the meantime.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 605432,
          "author_name": "stocks",
          "author_url": "",
          "post_date": "08/22/2019 11:22:20",
          "content": "<p>Terminal triple bond: H:C:::C:Some_atom. All four atoms are always on the same line.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "599919": "Messing about with my GP program 1JHC doesn't really look like a single class\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F241438%2F480b4f3d892166a660833992aeb4d01e%2F1JHC2Classes.png?generation=1565873976854009&amp;alt=media)",
    "599921": "Anyone had luck at classifying and then training a regressor model?",
    "599934": "I don;t really understand what I am looking at here, can you explain a bit? \nI would suspect you might be able to split by the \"type\" of the Carbon atom in the interaction.",
    "599939": "There is a clear split around 175 for scalar coupling (Outputs Graph) so if you add average scalar coupling per atom for example it is going to be sub optimal",
    "600094": "I tried a stratifiedkfold split after \"binning\" (using pd.cut and pd.qcut) the y. \nI'd guess it simulates kind of a classification. But my nn didn't perform better than a random split.\n\nI also used the bins to blend my three models using xgb. But the feature importance of the bins is very low.",
    "600672": "Carbons with sp/sp2 hybridization (connected to other atoms with 4 or 3 bonds) and other subclasses will have different ranges of coupling constants. If you adjust for this, I would assume that you won't see the same splitting in your model.",
    "600962": "Forgive me but the output I show here is from your files not from my model ;)",
    "600975": "The two groups are easily splited setting a threshold in 1J coupling distance to 1.065.",
    "600977": "Thanks - we really don't deserve you!",
    "601041": "&gt; The two groups are easily spited setting a threshold in 1J coupling distance to 1.075.\n\n@titericz, I am not sure I understand your post. If possible, explanation please?",
    "601064": "Calculate the distance of atom_index_0 and atom_index_1 for 1JHC couplings. If that distance is &lt; 1.065 then scalar_coupling_constant is on the top (200 value) of Scirpus plot. If distance &gt; 1.065 then scc is on the botton of the plot.",
    "601075": "Thanks @titericz.",
    "602857": "Oh. The same reasoning applies though. @titericz's observation that you can split by bond distance agrees well with this as well.",
    "605081": "All the 1JHC bonds with `sc_constant` around 200 begin with hydrogen and carbon followed by 2 carbons in a straight line. Below are two examples.\n![image](http://playagricola.com/Kaggle/ani7.gif)\n![image](http://playagricola.com/Kaggle/ani8.gif)",
    "605239": "Nice to see you jumping on board, Chris! Nice to see your observations again, I love 'em! Wish you luck in the upcoming week :)",
    "605311": "&gt; Nice to see you jumping on board, Chris!\n\nHis first sub is from a month ago here ;)\n\nBut I guess he's been busy with other competitions in the meantime.",
    "605432": "Terminal triple bond: H:C:::C:Some_atom. All four atoms are always on the same line.",
    "606224": "[Here](https://www.kaggle.com/somnambwl/1jhc-split) is the kernel with graphs."
  },
  "source": "meta"
}