{
  "id": 66799,
  "title": "Our models predictive?",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/66799",
  "author_name": "Maxwell",
  "post_date": "2018-09-25T17:04:29.988000",
  "votes": 10,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Just one coin. <br>\nThanks to <a href=\"https://www.kaggle.com/ogrellier/user-level-lightgbm-lb-1-4480/code\">olivier's kernel</a>, I trained LGBM and got a scatter plot of actual vs predicted. <br>\nAt a glance of it, I'm very surprised and disappointed by the fact our models might be not enough to foretell user level revenue. <br>\nI think this plot shows some hints to improve our models.  </p>\n\n<ol>\n<li>How do we improve non-zero revenue predictions?  </li>\n<li>How do we predict non-zero / zero revenue?</li>\n</ol>\n\n<p><img src=\"https://cdn-ak.f.st-hatena.com/images/fotolife/g/greenwind120170/20180926/20180926014938.png\" alt=\"fig1\"></p>",
  "messages": [
    {
      "id": 393601,
      "postDate": "2018-09-25T17:04:29.987Z",
      "content": "<p>Just one coin. <br>\nThanks to <a href=\"https://www.kaggle.com/ogrellier/user-level-lightgbm-lb-1-4480/code\">olivier's kernel</a>, I trained LGBM and got a scatter plot of actual vs predicted. <br>\nAt a glance of it, I'm very surprised and disappointed by the fact our models might be not enough to foretell user level revenue. <br>\nI think this plot shows some hints to improve our models.  </p>\n\n<ol>\n<li>How do we improve non-zero revenue predictions?  </li>\n<li>How do we predict non-zero / zero revenue?</li>\n</ol>\n\n<p><img src=\"https://cdn-ak.f.st-hatena.com/images/fotolife/g/greenwind120170/20180926/20180926014938.png\" alt=\"fig1\"></p>",
      "rawMarkdown": "Just one coin.  \nThanks to [olivier's kernel][1], I trained LGBM and got a scatter plot of actual vs predicted.  \nAt a glance of it, I'm very surprised and disappointed by the fact our models might be not enough to foretell user level revenue.  \nI think this plot shows some hints to improve our models.  \n  \n\n 1. How do we improve non-zero revenue predictions?  \n 2. How do we predict non-zero / zero revenue?\n\n![fig1][2]\n\n\n  [1]: https://www.kaggle.com/ogrellier/user-level-lightgbm-lb-1-4480/code\n  [2]: https://cdn-ak.f.st-hatena.com/images/fotolife/g/greenwind120170/20180926/20180926014938.png",
      "votes": 9
    },
    {
      "id": 393717,
      "postDate": "2018-09-25T19:42:16.163Z",
      "content": "<p>Pretty sure the 0/nonzero split is the most important angle here - so far I've had limited success ...</p>",
      "rawMarkdown": "Pretty sure the 0/nonzero split is the most important angle here - so far I've had limited success ...",
      "votes": 3,
      "replies": [
        {
          "id": 393835,
          "postDate": "2018-09-26T00:43:52.037Z",
          "content": "<p>Yes. <br>\nI have also tried, but no success.\nThe number of false positive was large in my case. And it made score worse.</p>",
          "rawMarkdown": "Yes.  \nI have also tried, but no success.\nThe number of false positive was large in my case. And it made score worse."
        },
        {
          "id": 395027,
          "postDate": "2018-09-27T20:51:02.397Z",
          "content": "<p>Me as well.  Nothing seems to work on that.  </p>",
          "rawMarkdown": "Me as well.  Nothing seems to work on that.  ",
          "votes": 1
        }
      ]
    },
    {
      "id": 393704,
      "postDate": "2018-09-25T19:24:19.823Z",
      "content": "<p>Don't forget it is log+1 - try look at it as expm1 - LOL</p>",
      "rawMarkdown": "Don't forget it is log+1 - try look at it as expm1 - LOL",
      "votes": 1,
      "replies": [
        {
          "id": 393706,
          "postDate": "2018-09-25T19:26:24.287Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 395031,
      "postDate": "2018-09-27T20:54:39.123Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true,
      "replies": [
        {
          "id": 395060,
          "postDate": "2018-09-27T22:11:14.863Z",
          "content": "<p><a href=\"/wenshihao1993\">@wenshihao1993</a> <br>\nThey announced that they have fixed. <br>\nBut someone pointed out there was still a problem.  </p>\n\n<p>Check this thread to catch those.  </p>\n\n<p><a href=\"https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66737\">https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66737</a>  </p>",
          "rawMarkdown": "@wenshihao1993  \nThey announced that they have fixed.  \nBut someone pointed out there was still a problem.  \n  \nCheck this thread to catch those.  \n  \n[https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66737][1]  \n  \n\n  [1]: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66737"
        },
        {
          "id": 395074,
          "postDate": "2018-09-27T22:59:26.200Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 393717,
      "author_name": "Konrad Banachewicz",
      "author_url": "",
      "post_date": "2018-09-25T19:42:16.163000",
      "content": "<p>Pretty sure the 0/nonzero split is the most important angle here - so far I've had limited success ...</p>",
      "votes": 3,
      "replies": [
        {
          "id": 393835,
          "author_name": "Maxwell",
          "author_url": "",
          "post_date": "2018-09-26T00:43:52.037000",
          "content": "<p>Yes. <br>\nI have also tried, but no success.\nThe number of false positive was large in my case. And it made score worse.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 395027,
          "author_name": "Ozzy Campos",
          "author_url": "",
          "post_date": "2018-09-27T20:51:02.397000",
          "content": "<p>Me as well.  Nothing seems to work on that.  </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 393704,
      "author_name": "Scirpus",
      "author_url": "",
      "post_date": "2018-09-25T19:24:19.823000",
      "content": "<p>Don't forget it is log+1 - try look at it as expm1 - LOL</p>",
      "votes": 1,
      "replies": [
        {
          "id": 393706,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-09-25T19:26:24.287000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 395031,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-09-27T20:54:39.123000",
      "content": "",
      "votes": 1,
      "replies": [
        {
          "id": 395060,
          "author_name": "Maxwell",
          "author_url": "",
          "post_date": "2018-09-27T22:11:14.863000",
          "content": "<p><a href=\"/wenshihao1993\">@wenshihao1993</a> <br>\nThey announced that they have fixed. <br>\nBut someone pointed out there was still a problem.  </p>\n\n<p>Check this thread to catch those.  </p>\n\n<p><a href=\"https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66737\">https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/66737</a>  </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 395074,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-09-27T22:59:26.200000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "393601": "Just one coin.  \nThanks to [olivier's kernel][1], I trained LGBM and got a scatter plot of actual vs predicted.  \nAt a glance of it, I'm very surprised and disappointed by the fact our models might be not enough to foretell user level revenue.  \nI think this plot shows some hints to improve our models.  \n  \n\n 1. How do we improve non-zero revenue predictions?  \n 2. How do we predict non-zero / zero revenue?\n\n![fig1][2]\n\n\n  [1]: https://www.kaggle.com/ogrellier/user-level-lightgbm-lb-1-4480/code\n  [2]: https://cdn-ak.f.st-hatena.com/images/fotolife/g/greenwind120170/20180926/20180926014938.png",
    "393717": "Pretty sure the 0/nonzero split is the most important angle here - so far I've had limited success ...",
    "393704": "Don't forget it is log+1 - try look at it as expm1 - LOL",
    "395031": ""
  }
}