{
  "id": 67153,
  "title": "Teaching Lighgbm to sum predictions",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/67153",
  "author_name": "",
  "post_date": "2018-09-29T10:59:29.766981300Z",
  "votes": 12,
  "comment_count": 2,
  "views": 0,
  "content": "<p>This is the last kernel of a series about what we can do to gain advantage of session and visitor level predictions.</p>\n\n<p>The idea here is to make all prices predicted at session level available at the visitor level and let LighGBM learn how to best sum them to get even better predictions.</p>\n\n<p><a href=\"https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions\">https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions</a></p>\n\n<p>Hope you'll enjoy.</p>",
  "messages": [
    {
      "id": "395784",
      "postDate": "09/29/2018 10:59:29",
      "content": "<p>This is the last kernel of a series about what we can do to gain advantage of session and visitor level predictions.</p>\n\n<p>The idea here is to make all prices predicted at session level available at the visitor level and let LighGBM learn how to best sum them to get even better predictions.</p>\n\n<p><a href=\"https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions\">https://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions</a></p>\n\n<p>Hope you'll enjoy.</p>",
      "rawMarkdown": "This is the last kernel of a series about what we can do to gain advantage of session and visitor level predictions.\n\nThe idea here is to make all prices predicted at session level available at the visitor level and let LighGBM learn how to best sum them to get even better predictions.\n\nhttps://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions\n\nHope you'll enjoy.",
      "votes": null
    },
    {
      "id": "396008",
      "postDate": "09/29/2018 20:18:27",
      "content": "<p>Interesting idea! What kind of LB/CV improvement do you observe if you compare to just taking the sum of the session predictions (i.e. log1p(sum(exp(log_predictions))) )?</p>",
      "rawMarkdown": "Interesting idea! What kind of LB/CV improvement do you observe if you compare to just taking the sum of the session predictions (i.e. log1p(sum(exp(log_predictions))) )?",
      "votes": null
    },
    {
      "id": "396012",
      "postDate": "09/29/2018 20:23:16",
      "content": "<p>Thanks S D. Something about 0.005 improvement but there was a silly bug in the code found by <a href=\"/fatihozturk\">@fatihozturk</a> . I'm short of submissions so I'll check that tomorrow.</p>",
      "rawMarkdown": "Thanks S D. Something about 0.005 improvement but there was a silly bug in the code found by @fatihozturk . I'm short of submissions so I'll check that tomorrow.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 396008,
      "author_name": "sdoria",
      "author_url": "",
      "post_date": "09/29/2018 20:18:27",
      "content": "<p>Interesting idea! What kind of LB/CV improvement do you observe if you compare to just taking the sum of the session predictions (i.e. log1p(sum(exp(log_predictions))) )?</p>",
      "votes": null,
      "replies": [
        {
          "id": 396012,
          "author_name": "ogrellier",
          "author_url": "",
          "post_date": "09/29/2018 20:23:16",
          "content": "<p>Thanks S D. Something about 0.005 improvement but there was a silly bug in the code found by <a href=\"/fatihozturk\">@fatihozturk</a> . I'm short of submissions so I'll check that tomorrow.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "395784": "This is the last kernel of a series about what we can do to gain advantage of session and visitor level predictions.\n\nThe idea here is to make all prices predicted at session level available at the visitor level and let LighGBM learn how to best sum them to get even better predictions.\n\nhttps://www.kaggle.com/ogrellier/teach-lightgbm-to-sum-predictions\n\nHope you'll enjoy.",
    "396008": "Interesting idea! What kind of LB/CV improvement do you observe if you compare to just taking the sum of the session predictions (i.e. log1p(sum(exp(log_predictions))) )?",
    "396012": "Thanks S D. Something about 0.005 improvement but there was a silly bug in the code found by @fatihozturk . I'm short of submissions so I'll check that tomorrow."
  },
  "source": "meta"
}