{
  "id": 361560,
  "title": "Stacking with control - unseen donor and day in validation",
  "url": "/competitions/open-problems-multimodal/discussion/361560",
  "author_name": "",
  "post_date": "2022-10-22T10:15:04.323086500Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Possible stacking scheme - your comments are welcome. </p>\n<p>Problem:<br>\nprivate LB - unseen DAY + donor<br>\npublic LB  - unseen donor ONLY</p>\n<p>We think of private LB, so for CITE-seq:</p>\n<p>Model 1:<br>\nTrain: Day 2  &amp; female<br>\nTest: Day 2 &amp; male 1 (day is the same but it is preliminary model - so do not bother)<br>\nPredict: Days 3,4 - it will be used as features.</p>\n<p>Model 2 (use predictions of Model1 AS FEATURES)<br>\nTrain: Day 3 + female + male1 <br>\nTest (private like):  Day 4 and all donors  (in particular unseen male 2).   <br>\n(Note: it can be used to for blending coefficients search)<br>\nTest (public like): Day 3 &amp; male2<br>\nPredict:  we can train on the whole Days 2,3 and predict to submission. (Can be other options with average - train day 2 , train day 3  and make averages for submission)<br>\n(I think it is simpler and safer exclude Day 1 here)</p>",
  "messages": [
    {
      "id": "1999412",
      "postDate": "10/22/2022 10:15:04",
      "content": "<p>Possible stacking scheme - your comments are welcome. </p>\n<p>Problem:<br>\nprivate LB - unseen DAY + donor<br>\npublic LB  - unseen donor ONLY</p>\n<p>We think of private LB, so for CITE-seq:</p>\n<p>Model 1:<br>\nTrain: Day 2  &amp; female<br>\nTest: Day 2 &amp; male 1 (day is the same but it is preliminary model - so do not bother)<br>\nPredict: Days 3,4 - it will be used as features.</p>\n<p>Model 2 (use predictions of Model1 AS FEATURES)<br>\nTrain: Day 3 + female + male1 <br>\nTest (private like):  Day 4 and all donors  (in particular unseen male 2).   <br>\n(Note: it can be used to for blending coefficients search)<br>\nTest (public like): Day 3 &amp; male2<br>\nPredict:  we can train on the whole Days 2,3 and predict to submission. (Can be other options with average - train day 2 , train day 3  and make averages for submission)<br>\n(I think it is simpler and safer exclude Day 1 here)</p>",
      "rawMarkdown": "Possible stacking scheme - your comments are welcome. \n\nProblem:\nprivate LB - unseen DAY + donor\npublic LB  - unseen donor ONLY\n\nWe think of private LB, so for CITE-seq:\n\nModel 1:\nTrain: Day 2  & female\nTest: Day 2 & male 1 (day is the same but it is preliminary model - so do not bother)\nPredict: Days 3,4 - it will be used as features.\n\nModel 2 (use predictions of Model1 AS FEATURES)\nTrain: Day 3 + female + male1 \nTest (private like):  Day 4 and all donors  (in particular unseen male 2).   \n(Note: it can be used to for blending coefficients search)\nTest (public like): Day 3 & male2\nPredict:  we can train on the whole Days 2,3 and predict to submission. (Can be other options with average - train day 2 , train day 3  and make averages for submission)\n(I think it is simpler and safer exclude Day 1 here)",
      "votes": null
    },
    {
      "id": "2006525",
      "postDate": "10/27/2022 16:18:50",
      "content": "<p>[deleted comment]</p>",
      "rawMarkdown": "[deleted comment]",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2006525,
      "author_name": "jinyang18",
      "author_url": "",
      "post_date": "10/27/2022 16:18:50",
      "content": "<p>[deleted comment]</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1999412": "Possible stacking scheme - your comments are welcome. \n\nProblem:\nprivate LB - unseen DAY + donor\npublic LB  - unseen donor ONLY\n\nWe think of private LB, so for CITE-seq:\n\nModel 1:\nTrain: Day 2  & female\nTest: Day 2 & male 1 (day is the same but it is preliminary model - so do not bother)\nPredict: Days 3,4 - it will be used as features.\n\nModel 2 (use predictions of Model1 AS FEATURES)\nTrain: Day 3 + female + male1 \nTest (private like):  Day 4 and all donors  (in particular unseen male 2).   \n(Note: it can be used to for blending coefficients search)\nTest (public like): Day 3 & male2\nPredict:  we can train on the whole Days 2,3 and predict to submission. (Can be other options with average - train day 2 , train day 3  and make averages for submission)\n(I think it is simpler and safer exclude Day 1 here)",
    "2006525": "[deleted comment]"
  },
  "source": "meta"
}