{
  "id": 85139,
  "title": "Huge shakeup and waiting for share solutions",
  "url": "/competitions/vsb-power-line-fault-detection/discussion/85139",
  "author_name": "Strideradu",
  "post_date": "2019-03-22T00:07:15.559000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Another competition with a huge shake. </p>\n\n<p>Just hope someone could share some ideas to avoid shakeup in such competition?</p>",
  "messages": [
    {
      "id": 496168,
      "postDate": "2019-03-22T00:24:57.780Z",
      "content": "<p>I did not take all time for this competition. It is a little supprised for me at the end.  Here is my lesson learn. \n- In the beginning, after probing LB and EDA, I saw that it is better to split data into 4 or 8 KFOLD even it hurts the LB. So I should trust CV\n- I dont like searching threshold idea for the competition like this.  You wouldnt know the performance comes from your new features or new threshold. I often use fix threshold (0.3) and trust the stable.  It means that with a fixed threshold, your CV increases and LB increases too. </p>\n\n<p>Finally, I use all the code of this <a href=\"https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694\">kernel</a>, change the KFOLD and threshold then submit. </p>",
      "rawMarkdown": "I did not take all time for this competition. It is a little supprised for me at the end.  Here is my lesson learn. \n- In the beginning, after probing LB and EDA, I saw that it is better to split data into 4 or 8 KFOLD even it hurts the LB. So I should trust CV\n- I dont like searching threshold idea for the competition like this.  You wouldnt know the performance comes from your new features or new threshold. I often use fix threshold (0.3) and trust the stable.  It means that with a fixed threshold, your CV increases and LB increases too. \n\nFinally, I use all the code of this [kernel](https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694), change the KFOLD and threshold then submit. \n",
      "votes": 1
    },
    {
      "id": 496162,
      "postDate": "2019-03-22T00:11:23.907Z",
      "content": "<p>This was bigger than the last shake. But had a good experience using RNN's</p>",
      "rawMarkdown": "This was bigger than the last shake. But had a good experience using RNN's",
      "votes": 2
    },
    {
      "id": 496154,
      "postDate": "2019-03-22T00:07:15.560Z",
      "content": "<p>Another competition with a huge shake. </p>\n\n<p>Just hope someone could share some ideas to avoid shakeup in such competition?</p>",
      "rawMarkdown": "Another competition with a huge shake. \n\nJust hope someone could share some ideas to avoid shakeup in such competition?",
      "votes": 2
    },
    {
      "id": 496170,
      "postDate": "2019-03-22T00:27:27.303Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 496168,
      "author_name": "cab",
      "author_url": "",
      "post_date": "2019-03-22T00:24:57.780000",
      "content": "<p>I did not take all time for this competition. It is a little supprised for me at the end.  Here is my lesson learn. \n- In the beginning, after probing LB and EDA, I saw that it is better to split data into 4 or 8 KFOLD even it hurts the LB. So I should trust CV\n- I dont like searching threshold idea for the competition like this.  You wouldnt know the performance comes from your new features or new threshold. I often use fix threshold (0.3) and trust the stable.  It means that with a fixed threshold, your CV increases and LB increases too. </p>\n\n<p>Finally, I use all the code of this <a href=\"https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694\">kernel</a>, change the KFOLD and threshold then submit. </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 496162,
      "author_name": "Ashish Gupta",
      "author_url": "",
      "post_date": "2019-03-22T00:11:23.907000",
      "content": "<p>This was bigger than the last shake. But had a good experience using RNN's</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 496170,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-03-22T00:27:27.303000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "496168": "I did not take all time for this competition. It is a little supprised for me at the end.  Here is my lesson learn. \n- In the beginning, after probing LB and EDA, I saw that it is better to split data into 4 or 8 KFOLD even it hurts the LB. So I should trust CV\n- I dont like searching threshold idea for the competition like this.  You wouldnt know the performance comes from your new features or new threshold. I often use fix threshold (0.3) and trust the stable.  It means that with a fixed threshold, your CV increases and LB increases too. \n\nFinally, I use all the code of this [kernel](https://www.kaggle.com/braquino/5-fold-lstm-attention-fully-commented-0-694), change the KFOLD and threshold then submit. \n",
    "496162": "This was bigger than the last shake. But had a good experience using RNN's",
    "496154": "Another competition with a huge shake. \n\nJust hope someone could share some ideas to avoid shakeup in such competition?",
    "496170": ""
  }
}