{
  "id": 75245,
  "title": "GP or curve fitting as input to RNN/CNN",
  "url": "/competitions/PLAsTiCC-2018/discussion/75245",
  "author_name": "",
  "post_date": "2018-12-19T22:02:08.614971500Z",
  "votes": 5,
  "comment_count": 1,
  "views": 0,
  "content": "<p>As many people stated out, time gaps make this problem quite hard for deep learning models, may be because of that LightGBM was better for most of the competitors.</p>\n\n<p>Have anybody tried to use fit from Gaussian Process (or from curve fitting methods) as input to RNN/CNN model instead of raw time series? \nPictures with GP fits are pretty good, and you can get regular time series, it looks to me like a good option. But I didn't find anything about such approach in write-ups.</p>",
  "messages": [
    {
      "id": "442380",
      "postDate": "12/19/2018 22:02:08",
      "content": "<p>As many people stated out, time gaps make this problem quite hard for deep learning models, may be because of that LightGBM was better for most of the competitors.</p>\n\n<p>Have anybody tried to use fit from Gaussian Process (or from curve fitting methods) as input to RNN/CNN model instead of raw time series? \nPictures with GP fits are pretty good, and you can get regular time series, it looks to me like a good option. But I didn't find anything about such approach in write-ups.</p>",
      "rawMarkdown": "As many people stated out, time gaps make this problem quite hard for deep learning models, may be because of that LightGBM was better for most of the competitors.\n\nHave anybody tried to use fit from Gaussian Process (or from curve fitting methods) as input to RNN/CNN model instead of raw time series? \nPictures with GP fits are pretty good, and you can get regular time series, it looks to me like a good option. But I didn't find anything about such approach in write-ups.",
      "votes": null
    },
    {
      "id": "442465",
      "postDate": "12/20/2018 02:01:07",
      "content": "<p>This could be a great approach inded.  We had this in mind but mastering GP enough took us too much time.  There are research papers where this was used successfully, eg: <a href=\"https://arxiv.org/pdf/1706.03811.pdf\">https://arxiv.org/pdf/1706.03811.pdf</a></p>\n\n<p>We discovered that paper late during the competition.</p>\n\n<p>I think that replacing the linear interpolation Yuvalr used by GP interpolations would be really good.</p>",
      "rawMarkdown": "This could be a great approach inded.  We had this in mind but mastering GP enough took us too much time.  There are research papers where this was used successfully, eg: https://arxiv.org/pdf/1706.03811.pdf\n\nWe discovered that paper late during the competition.\n\nI think that replacing the linear interpolation Yuvalr used by GP interpolations would be really good.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 442465,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "12/20/2018 02:01:07",
      "content": "<p>This could be a great approach inded.  We had this in mind but mastering GP enough took us too much time.  There are research papers where this was used successfully, eg: <a href=\"https://arxiv.org/pdf/1706.03811.pdf\">https://arxiv.org/pdf/1706.03811.pdf</a></p>\n\n<p>We discovered that paper late during the competition.</p>\n\n<p>I think that replacing the linear interpolation Yuvalr used by GP interpolations would be really good.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "442380": "As many people stated out, time gaps make this problem quite hard for deep learning models, may be because of that LightGBM was better for most of the competitors.\n\nHave anybody tried to use fit from Gaussian Process (or from curve fitting methods) as input to RNN/CNN model instead of raw time series? \nPictures with GP fits are pretty good, and you can get regular time series, it looks to me like a good option. But I didn't find anything about such approach in write-ups.",
    "442465": "This could be a great approach inded.  We had this in mind but mastering GP enough took us too much time.  There are research papers where this was used successfully, eg: https://arxiv.org/pdf/1706.03811.pdf\n\nWe discovered that paper late during the competition.\n\nI think that replacing the linear interpolation Yuvalr used by GP interpolations would be really good."
  },
  "source": "meta"
}