{
  "id": 96930,
  "title": "weak/semi-supervised regression ",
  "url": "/competitions/champs-scalar-coupling/discussion/96930",
  "author_name": "",
  "post_date": "2019-06-24T16:39:08.082706900Z",
  "votes": 19,
  "comment_count": 2,
  "views": 0,
  "content": "<p>this kaggle task is a bit unique in the sense that it is given:</p>\n\n<p>coupling_constant = fc + sd + pso + dso</p>\n\n<p>If you train a predictor separately for each of the terms,  then given a test sample, you can estimate the confidence of its prediction by considering difference = coupling_constant -( fc + sd + pso + dso). </p>\n\n<p>If you know the confidence of your prediction, you can plug into weak/semi-supervised  methods using both labelled train and unlabeled test  training.\n```\ne.g. total loss = L( train_coupling_constant, true_coupling_constant) + L( train_fc, true_fc)  + ... + L(test_coupling_constant, test_fc, test_sd, test_pso, test_dso)</p>\n\n<p>```</p>\n\n<p>There are also other several other terms like shielding, dipole, charges, etc ... I am not a chemistry expert but i think these terms are related to coupling_constant as well and can be used for self-supervision signals.</p>\n\n<p>To improve learning, train with drop task:\n```\ne.g. train sample1 : truth = (coupling_constant, fc, sd,pso, dso)\ne.g. train sample2 : truth = (coupling_constant, fc, None,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, None,pso, None)\ne.g. train sample4: truth = (None, fc, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, None, sd,pso, dso)</p>\n\n<p>next batch\ne.g. train sample1 : truth = (None, fc, sd,pso, dso)\ne.g. train sample2 : truth = (None, fc, sd,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, sd,pso, None)\ne.g. train sample4: truth = (coupling_constant, None, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, fc, sd,None, dso)\n```</p>",
  "messages": [
    {
      "id": "559830",
      "postDate": "06/24/2019 16:39:08",
      "content": "<p>this kaggle task is a bit unique in the sense that it is given:</p>\n\n<p>coupling_constant = fc + sd + pso + dso</p>\n\n<p>If you train a predictor separately for each of the terms,  then given a test sample, you can estimate the confidence of its prediction by considering difference = coupling_constant -( fc + sd + pso + dso). </p>\n\n<p>If you know the confidence of your prediction, you can plug into weak/semi-supervised  methods using both labelled train and unlabeled test  training.\n```\ne.g. total loss = L( train_coupling_constant, true_coupling_constant) + L( train_fc, true_fc)  + ... + L(test_coupling_constant, test_fc, test_sd, test_pso, test_dso)</p>\n\n<p>```</p>\n\n<p>There are also other several other terms like shielding, dipole, charges, etc ... I am not a chemistry expert but i think these terms are related to coupling_constant as well and can be used for self-supervision signals.</p>\n\n<p>To improve learning, train with drop task:\n```\ne.g. train sample1 : truth = (coupling_constant, fc, sd,pso, dso)\ne.g. train sample2 : truth = (coupling_constant, fc, None,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, None,pso, None)\ne.g. train sample4: truth = (None, fc, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, None, sd,pso, dso)</p>\n\n<p>next batch\ne.g. train sample1 : truth = (None, fc, sd,pso, dso)\ne.g. train sample2 : truth = (None, fc, sd,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, sd,pso, None)\ne.g. train sample4: truth = (coupling_constant, None, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, fc, sd,None, dso)\n```</p>",
      "rawMarkdown": "this kaggle task is a bit unique in the sense that it is given:\n\ncoupling_constant = fc + sd + pso + dso\n\nIf you train a predictor separately for each of the terms,  then given a test sample, you can estimate the confidence of its prediction by considering difference = coupling_constant -( fc + sd + pso + dso). \n\nIf you know the confidence of your prediction, you can plug into weak/semi-supervised  methods using both labelled train and unlabeled test  training.\n```\ne.g. total loss = L( train_coupling_constant, true_coupling_constant) + L( train_fc, true_fc)  + ... + L(test_coupling_constant, test_fc, test_sd, test_pso, test_dso)\n\n```\n\nThere are also other several other terms like shielding, dipole, charges, etc ... I am not a chemistry expert but i think these terms are related to coupling_constant as well and can be used for self-supervision signals.\n\nTo improve learning, train with drop task:\n```\ne.g. train sample1 : truth = (coupling_constant, fc, sd,pso, dso)\ne.g. train sample2 : truth = (coupling_constant, fc, None,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, None,pso, None)\ne.g. train sample4: truth = (None, fc, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, None, sd,pso, dso)\n\nnext batch\ne.g. train sample1 : truth = (None, fc, sd,pso, dso)\ne.g. train sample2 : truth = (None, fc, sd,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, sd,pso, None)\ne.g. train sample4: truth = (coupling_constant, None, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, fc, sd,None, dso)\n```",
      "votes": null
    },
    {
      "id": "563266",
      "postDate": "06/28/2019 04:41:46",
      "content": "<p>Do you add the oof estimates of the corresponding dropped target from the first batch as features for the next batch?</p>",
      "rawMarkdown": "Do you add the oof estimates of the corresponding dropped target from the first batch as features for the next batch?",
      "votes": null
    },
    {
      "id": "563275",
      "postDate": "06/28/2019 04:55:51",
      "content": "<p>that can be another  plausible strategy. i haven't tried yet</p>",
      "rawMarkdown": "that can be another  plausible strategy. i haven't tried yet",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 563266,
      "author_name": "scaomath",
      "author_url": "",
      "post_date": "06/28/2019 04:41:46",
      "content": "<p>Do you add the oof estimates of the corresponding dropped target from the first batch as features for the next batch?</p>",
      "votes": null,
      "replies": [
        {
          "id": 563275,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "06/28/2019 04:55:51",
          "content": "<p>that can be another  plausible strategy. i haven't tried yet</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "559830": "this kaggle task is a bit unique in the sense that it is given:\n\ncoupling_constant = fc + sd + pso + dso\n\nIf you train a predictor separately for each of the terms,  then given a test sample, you can estimate the confidence of its prediction by considering difference = coupling_constant -( fc + sd + pso + dso). \n\nIf you know the confidence of your prediction, you can plug into weak/semi-supervised  methods using both labelled train and unlabeled test  training.\n```\ne.g. total loss = L( train_coupling_constant, true_coupling_constant) + L( train_fc, true_fc)  + ... + L(test_coupling_constant, test_fc, test_sd, test_pso, test_dso)\n\n```\n\nThere are also other several other terms like shielding, dipole, charges, etc ... I am not a chemistry expert but i think these terms are related to coupling_constant as well and can be used for self-supervision signals.\n\nTo improve learning, train with drop task:\n```\ne.g. train sample1 : truth = (coupling_constant, fc, sd,pso, dso)\ne.g. train sample2 : truth = (coupling_constant, fc, None,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, None,pso, None)\ne.g. train sample4: truth = (None, fc, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, None, sd,pso, dso)\n\nnext batch\ne.g. train sample1 : truth = (None, fc, sd,pso, dso)\ne.g. train sample2 : truth = (None, fc, sd,pso, dso)\ne.g. train sample3 : truth = (coupling_constant, fc, sd,pso, None)\ne.g. train sample4: truth = (coupling_constant, None, sd,pso, dso)\ne.g. train sample5 : truth = (coupling_constant, fc, sd,None, dso)\n```",
    "563266": "Do you add the oof estimates of the corresponding dropped target from the first batch as features for the next batch?",
    "563275": "that can be another  plausible strategy. i haven't tried yet"
  },
  "source": "meta"
}