{
  "id": 96605,
  "title": "Train only data",
  "url": "/competitions/champs-scalar-coupling/discussion/96605",
  "author_name": "",
  "post_date": "2019-06-21T11:44:01.250971Z",
  "votes": -2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>What exactly is the point of providing data which is only available for the training set ?\nIf you use it, surprise, you can get better results.   </p>",
  "messages": [
    {
      "id": "557683",
      "postDate": "06/21/2019 11:44:01",
      "content": "<p>What exactly is the point of providing data which is only available for the training set ?\nIf you use it, surprise, you can get better results.   </p>",
      "rawMarkdown": "What exactly is the point of providing data which is only available for the training set ?\nIf you use it, surprise, you can get better results.",
      "votes": null
    },
    {
      "id": "558052",
      "postDate": "06/21/2019 18:56:15",
      "content": "<p>I think the idea is that you could in principle train a proxy model that predicts these properties for the test set as well. These predictions can then be used to enhance your data.</p>",
      "rawMarkdown": "I think the idea is that you could in principle train a proxy model that predicts these properties for the test set as well. These predictions can then be used to enhance your data.",
      "votes": null
    },
    {
      "id": "558755",
      "postDate": "06/22/2019 21:28:05",
      "content": "<p>As a simple example compare my kernel that only uses distances that can be computed for the train and test sets and has an LB score of -0.306:\n<a href=\"https://www.kaggle.com/robertburbidge/distance-features\">https://www.kaggle.com/robertburbidge/distance-features</a>\nwith my kernel that uses estimated dipole moment magnitude, potential energy, and Mulliken charges for both the train and test set and has an LB score of -0.393:\n<a href=\"https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges\">https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges</a>\nIt is only a small improvement and is not a particularly good LB score but it proves the point. I expect that some feature engineering on the imputed features (brute force or guided by the physical chemistry) would improve the score further.</p>",
      "rawMarkdown": "As a simple example compare my kernel that only uses distances that can be computed for the train and test sets and has an LB score of -0.306:\n[https://www.kaggle.com/robertburbidge/distance-features](https://www.kaggle.com/robertburbidge/distance-features)\nwith my kernel that uses estimated dipole moment magnitude, potential energy, and Mulliken charges for both the train and test set and has an LB score of -0.393:\n[https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges](https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges)\nIt is only a small improvement and is not a particularly good LB score but it proves the point. I expect that some feature engineering on the imputed features (brute force or guided by the physical chemistry) would improve the score further.",
      "votes": null
    },
    {
      "id": "561048",
      "postDate": "06/26/2019 04:44:04",
      "content": "<p>This kernel by <a href=\"/artgor\">@artgor</a> could be another good starting point : <a href=\"https://www.kaggle.com/artgor/using-meta-features-to-improve-model\">https://www.kaggle.com/artgor/using-meta-features-to-improve-model</a></p>",
      "rawMarkdown": "This kernel by @artgor could be another good starting point : https://www.kaggle.com/artgor/using-meta-features-to-improve-model",
      "votes": null
    },
    {
      "id": "561126",
      "postDate": "06/26/2019 06:52:50",
      "content": "<p>What exactly is the point of creating another topic with this question when there's already 9000+ of them? Srsly?</p>",
      "rawMarkdown": "What exactly is the point of creating another topic with this question when there's already 9000+ of them? Srsly?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 558052,
      "author_name": "semodi",
      "author_url": "",
      "post_date": "06/21/2019 18:56:15",
      "content": "<p>I think the idea is that you could in principle train a proxy model that predicts these properties for the test set as well. These predictions can then be used to enhance your data.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 558755,
      "author_name": "robertburbidge",
      "author_url": "",
      "post_date": "06/22/2019 21:28:05",
      "content": "<p>As a simple example compare my kernel that only uses distances that can be computed for the train and test sets and has an LB score of -0.306:\n<a href=\"https://www.kaggle.com/robertburbidge/distance-features\">https://www.kaggle.com/robertburbidge/distance-features</a>\nwith my kernel that uses estimated dipole moment magnitude, potential energy, and Mulliken charges for both the train and test set and has an LB score of -0.393:\n<a href=\"https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges\">https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges</a>\nIt is only a small improvement and is not a particularly good LB score but it proves the point. I expect that some feature engineering on the imputed features (brute force or guided by the physical chemistry) would improve the score further.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 561048,
      "author_name": "arnabbiswas1",
      "author_url": "",
      "post_date": "06/26/2019 04:44:04",
      "content": "<p>This kernel by <a href=\"/artgor\">@artgor</a> could be another good starting point : <a href=\"https://www.kaggle.com/artgor/using-meta-features-to-improve-model\">https://www.kaggle.com/artgor/using-meta-features-to-improve-model</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 561126,
      "author_name": "stanislavblinov",
      "author_url": "",
      "post_date": "06/26/2019 06:52:50",
      "content": "<p>What exactly is the point of creating another topic with this question when there's already 9000+ of them? Srsly?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "557683": "What exactly is the point of providing data which is only available for the training set ?\nIf you use it, surprise, you can get better results.",
    "558052": "I think the idea is that you could in principle train a proxy model that predicts these properties for the test set as well. These predictions can then be used to enhance your data.",
    "558755": "As a simple example compare my kernel that only uses distances that can be computed for the train and test sets and has an LB score of -0.306:\n[https://www.kaggle.com/robertburbidge/distance-features](https://www.kaggle.com/robertburbidge/distance-features)\nwith my kernel that uses estimated dipole moment magnitude, potential energy, and Mulliken charges for both the train and test set and has an LB score of -0.393:\n[https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges](https://www.kaggle.com/robertburbidge/using-estimated-mulliken-charges)\nIt is only a small improvement and is not a particularly good LB score but it proves the point. I expect that some feature engineering on the imputed features (brute force or guided by the physical chemistry) would improve the score further.",
    "561048": "This kernel by @artgor could be another good starting point : https://www.kaggle.com/artgor/using-meta-features-to-improve-model",
    "561126": "What exactly is the point of creating another topic with this question when there's already 9000+ of them? Srsly?"
  },
  "source": "meta"
}