{
  "id": 12503,
  "title": "Anyone wants to team up?",
  "url": "/competitions/inria-bci-challenge/discussion/12503",
  "author_name": "",
  "post_date": "2015-02-13T00:13:52.543Z",
  "votes": 1,
  "comment_count": 5,
  "views": 1589,
  "content": "<p>I wrote an ensemble pipeline that I believe is tolerant&nbsp;to overfitting. I use 5 fold cross validation, each fold takes 20% of every&nbsp;subject as validate set. And I did some research on eeg features and include some features that I think might be useful, like correlation matrix, fisher information, etc.</p>\n<p>My current cv for one best single model is about 76%.</p>\n<p>I am looking for some team that does something&nbsp;on finding features.</p>\n<p>Please reply here or contact me &nbsp;by message.&nbsp;</p>\n<p>Thanks :)</p>",
  "messages": [
    {
      "id": "64153",
      "postDate": "02/13/2015 00:13:52",
      "content": "<p>I wrote an ensemble pipeline that I believe is tolerant&nbsp;to overfitting. I use 5 fold cross validation, each fold takes 20% of every&nbsp;subject as validate set. And I did some research on eeg features and include some features that I think might be useful, like correlation matrix, fisher information, etc.</p>\n<p>My current cv for one best single model is about 76%.</p>\n<p>I am looking for some team that does something&nbsp;on finding features.</p>\n<p>Please reply here or contact me &nbsp;by message.&nbsp;</p>\n<p>Thanks :)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64243",
      "postDate": "02/14/2015 20:42:26",
      "content": "<p>found one&#65281;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64283",
      "postDate": "02/15/2015 19:43:46",
      "content": "<p>I achieved 78% cv myself but got 65% upon submission. &nbsp;Try submitting and see what you get.</p>\n\n<p>Are you using subject / session id in your model?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64305",
      "postDate": "02/16/2015 02:49:43",
      "content": "<p>Hi&nbsp;Phillip Chilton Adkins,</p>\n<p>My current score is the one with 76% cv result.&nbsp;</p>\n<p>I didn't include subject or session ID, all the features are extracted from the channels. Although the LB is not trustful here, I suspect you did something inappropriate to get such a gap. But I might be wrong.</p>\n<p>Good luck!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64345",
      "postDate": "02/16/2015 15:48:39",
      "content": "<p>Public LB is only 20% and notoriously fickle in these types of competitions (see past history of Kaggle transfer learning, ECG data based, etc.). Expect a massive shakeup. Models in which I've obtained well over 0.8+ leave one subject out CV have not been scoring well on the&nbsp;public&nbsp;LB. &nbsp;</p>\n<p>But overfitting is a lot more fun for the YOLO types.&nbsp;</p>\n\n<p>YOLO.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64346",
      "postDate": "02/16/2015 15:52:30",
      "content": "<p>i think this will be like africa challenge or MLSP challenge.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 64243,
      "author_name": "xiaozhouwang",
      "author_url": "",
      "post_date": "02/14/2015 20:42:26",
      "content": "<p>found one&#65281;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 64283,
      "author_name": "phillipadkins",
      "author_url": "",
      "post_date": "02/15/2015 19:43:46",
      "content": "<p>I achieved 78% cv myself but got 65% upon submission. &nbsp;Try submitting and see what you get.</p>\n\n<p>Are you using subject / session id in your model?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 64305,
      "author_name": "xiaozhouwang",
      "author_url": "",
      "post_date": "02/16/2015 02:49:43",
      "content": "<p>Hi&nbsp;Phillip Chilton Adkins,</p>\n<p>My current score is the one with 76% cv result.&nbsp;</p>\n<p>I didn't include subject or session ID, all the features are extracted from the channels. Although the LB is not trustful here, I suspect you did something inappropriate to get such a gap. But I might be wrong.</p>\n<p>Good luck!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 64345,
      "author_name": "mikeskim",
      "author_url": "",
      "post_date": "02/16/2015 15:48:39",
      "content": "<p>Public LB is only 20% and notoriously fickle in these types of competitions (see past history of Kaggle transfer learning, ECG data based, etc.). Expect a massive shakeup. Models in which I've obtained well over 0.8+ leave one subject out CV have not been scoring well on the&nbsp;public&nbsp;LB. &nbsp;</p>\n<p>But overfitting is a lot more fun for the YOLO types.&nbsp;</p>\n\n<p>YOLO.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 64346,
      "author_name": "abhishek",
      "author_url": "",
      "post_date": "02/16/2015 15:52:30",
      "content": "<p>i think this will be like africa challenge or MLSP challenge.&nbsp;</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "64153": "",
    "64243": "",
    "64283": "",
    "64305": "",
    "64345": "",
    "64346": ""
  },
  "source": "meta"
}