{
  "id": 118348,
  "title": "Leakage data makes me so upset.",
  "url": "/competitions/ashrae-energy-prediction/discussion/118348",
  "author_name": "",
  "post_date": "2019-11-20T22:34:49.034562100Z",
  "votes": null,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I have to say using leakage data is too tricky (but I used it too). I think I might be give it some time and wait the whole leakage data released.</p>\n\n<p>And I am really suspicious that the private LB can avoid these leakage data since it also contain <strong>2018 data</strong>.</p>",
  "messages": [
    {
      "id": "678020",
      "postDate": "11/20/2019 22:34:49",
      "content": "<p>I have to say using leakage data is too tricky (but I used it too). I think I might be give it some time and wait the whole leakage data released.</p>\n\n<p>And I am really suspicious that the private LB can avoid these leakage data since it also contain <strong>2018 data</strong>.</p>",
      "rawMarkdown": "I have to say using leakage data is too tricky (but I used it too). I think I might be give it some time and wait the whole leakage data released.\n\nAnd I am really suspicious that the private LB can avoid these leakage data since it also contain **2018 data**.",
      "votes": null
    },
    {
      "id": "678036",
      "postDate": "11/20/2019 23:13:39",
      "content": "<p>I don't think you need to worry about people merging ground truth labels onto test prediction submission. What you should worry about is people finding a way to make the leaked test labels useful for lowering private LB by training/validating using all available data. This part is extremely difficult and tricky. Look at people's score without leak <a href=\"https://www.kaggle.com/c/ashrae-energy-prediction/discussion/118214\">here</a>.</p>",
      "rawMarkdown": "I don't think you need to worry about people merging ground truth labels onto test prediction submission. What you should worry about is people finding a way to make the leaked test labels useful for lowering private LB by training/validating using all available data. This part is extremely difficult and tricky. Look at people's score without leak [here](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/118214).",
      "votes": null
    },
    {
      "id": "678064",
      "postDate": "11/21/2019 00:55:14",
      "content": "<p>Yes! what you said is correct! I think it might be better if we use leaked test data and train data as a new train data to train model. </p>",
      "rawMarkdown": "Yes! what you said is correct! I think it might be better if we use leaked test data and train data as a new train data to train model.",
      "votes": null
    },
    {
      "id": "678384",
      "postDate": "11/21/2019 10:00:34",
      "content": "<p>LB will be corrected after eliminating dummy predictions, but I do not think it will be much different from current position unless someone has a similar true LB score with Top scorers and do not replace dummies with leaked data.</p>",
      "rawMarkdown": "LB will be corrected after eliminating dummy predictions, but I do not think it will be much different from current position unless someone has a similar true LB score with Top scorers and do not replace dummies with leaked data.",
      "votes": null
    },
    {
      "id": "678489",
      "postDate": "11/21/2019 13:15:48",
      "content": "<p>I wonder how they can eliminate dummy predictions...\nMaybe the real prediction will have some little difference from true value so the perfectly same data will be recognized as dummy data?</p>",
      "rawMarkdown": "I wonder how they can eliminate dummy predictions...\nMaybe the real prediction will have some little difference from true value so the perfectly same data will be recognized as dummy data?",
      "votes": null
    },
    {
      "id": "678750",
      "postDate": "11/21/2019 20:10:19",
      "content": "<p>Official says they will exclude predictions publicly available. I do not know how they know if these are though.</p>",
      "rawMarkdown": "Official says they will exclude predictions publicly available. I do not know how they know if these are though.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 678036,
      "author_name": "teeyee314",
      "author_url": "",
      "post_date": "11/20/2019 23:13:39",
      "content": "<p>I don't think you need to worry about people merging ground truth labels onto test prediction submission. What you should worry about is people finding a way to make the leaked test labels useful for lowering private LB by training/validating using all available data. This part is extremely difficult and tricky. Look at people's score without leak <a href=\"https://www.kaggle.com/c/ashrae-energy-prediction/discussion/118214\">here</a>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 678064,
          "author_name": "yixinchen1",
          "author_url": "",
          "post_date": "11/21/2019 00:55:14",
          "content": "<p>Yes! what you said is correct! I think it might be better if we use leaked test data and train data as a new train data to train model. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 678384,
      "author_name": "hanjoonchoe",
      "author_url": "",
      "post_date": "11/21/2019 10:00:34",
      "content": "<p>LB will be corrected after eliminating dummy predictions, but I do not think it will be much different from current position unless someone has a similar true LB score with Top scorers and do not replace dummies with leaked data.</p>",
      "votes": null,
      "replies": [
        {
          "id": 678489,
          "author_name": "yixinchen1",
          "author_url": "",
          "post_date": "11/21/2019 13:15:48",
          "content": "<p>I wonder how they can eliminate dummy predictions...\nMaybe the real prediction will have some little difference from true value so the perfectly same data will be recognized as dummy data?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 678750,
          "author_name": "hanjoonchoe",
          "author_url": "",
          "post_date": "11/21/2019 20:10:19",
          "content": "<p>Official says they will exclude predictions publicly available. I do not know how they know if these are though.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "678020": "I have to say using leakage data is too tricky (but I used it too). I think I might be give it some time and wait the whole leakage data released.\n\nAnd I am really suspicious that the private LB can avoid these leakage data since it also contain **2018 data**.",
    "678036": "I don't think you need to worry about people merging ground truth labels onto test prediction submission. What you should worry about is people finding a way to make the leaked test labels useful for lowering private LB by training/validating using all available data. This part is extremely difficult and tricky. Look at people's score without leak [here](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/118214).",
    "678064": "Yes! what you said is correct! I think it might be better if we use leaked test data and train data as a new train data to train model.",
    "678384": "LB will be corrected after eliminating dummy predictions, but I do not think it will be much different from current position unless someone has a similar true LB score with Top scorers and do not replace dummies with leaked data.",
    "678489": "I wonder how they can eliminate dummy predictions...\nMaybe the real prediction will have some little difference from true value so the perfectly same data will be recognized as dummy data?",
    "678750": "Official says they will exclude predictions publicly available. I do not know how they know if these are though."
  },
  "source": "meta"
}