{
  "id": 81038,
  "title": "Overfitting the Public Leaderboard",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/81038",
  "author_name": "",
  "post_date": "2019-02-18T19:02:30.782468800Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Due to feature engineering and stacking leading to minimal gains, I believe that most people have used outlier detection to improve their score.  If this is true, then most people could have overfit their predictions trying to cater toward outliers.  What are you thoughts on this?  Do you believe that you overfit?\nYou can also select two predictions for the final test set, so I would select one prediction from a model with outlier detection and one from a model without it.</p>",
  "messages": [
    {
      "id": "473965",
      "postDate": "02/18/2019 19:02:30",
      "content": "<p>Due to feature engineering and stacking leading to minimal gains, I believe that most people have used outlier detection to improve their score.  If this is true, then most people could have overfit their predictions trying to cater toward outliers.  What are you thoughts on this?  Do you believe that you overfit?\nYou can also select two predictions for the final test set, so I would select one prediction from a model with outlier detection and one from a model without it.</p>",
      "rawMarkdown": "Due to feature engineering and stacking leading to minimal gains, I believe that most people have used outlier detection to improve their score.  If this is true, then most people could have overfit their predictions trying to cater toward outliers.  What are you thoughts on this?  Do you believe that you overfit?\nYou can also select two predictions for the final test set, so I would select one prediction from a model with outlier detection and one from a model without it.",
      "votes": null
    },
    {
      "id": "474019",
      "postDate": "02/18/2019 20:26:45",
      "content": "<p>You are absolutely right - everyone has just been doing massive outlier overrides based on LB feedback. But this is a recipe for disaster.</p>\n\n<p>And your submission selection strategy is valid - this is something we are strongly considering as well</p>",
      "rawMarkdown": "You are absolutely right - everyone has just been doing massive outlier overrides based on LB feedback. But this is a recipe for disaster.\n\nAnd your submission selection strategy is valid - this is something we are strongly considering as well",
      "votes": null
    },
    {
      "id": "474107",
      "postDate": "02/19/2019 00:19:33",
      "content": "<p>Outlier detection doesn't seem consistent to me, but I will go with both anyway. </p>",
      "rawMarkdown": "Outlier detection doesn't seem consistent to me, but I will go with both anyway.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 474019,
      "author_name": "raddar",
      "author_url": "",
      "post_date": "02/18/2019 20:26:45",
      "content": "<p>You are absolutely right - everyone has just been doing massive outlier overrides based on LB feedback. But this is a recipe for disaster.</p>\n\n<p>And your submission selection strategy is valid - this is something we are strongly considering as well</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 474107,
      "author_name": "klloke",
      "author_url": "",
      "post_date": "02/19/2019 00:19:33",
      "content": "<p>Outlier detection doesn't seem consistent to me, but I will go with both anyway. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "473965": "Due to feature engineering and stacking leading to minimal gains, I believe that most people have used outlier detection to improve their score.  If this is true, then most people could have overfit their predictions trying to cater toward outliers.  What are you thoughts on this?  Do you believe that you overfit?\nYou can also select two predictions for the final test set, so I would select one prediction from a model with outlier detection and one from a model without it.",
    "474019": "You are absolutely right - everyone has just been doing massive outlier overrides based on LB feedback. But this is a recipe for disaster.\n\nAnd your submission selection strategy is valid - this is something we are strongly considering as well",
    "474107": "Outlier detection doesn't seem consistent to me, but I will go with both anyway."
  },
  "source": "meta"
}