{
  "id": 121261,
  "title": "Is it actually worth using leaked data?",
  "url": "/competitions/ashrae-energy-prediction/discussion/121261",
  "author_name": "Thiago Preischadt",
  "post_date": "2019-12-12T04:59:14.231000",
  "votes": 6,
  "comment_count": 20,
  "views": 0,
  "content": "<p>Sorry maybe this has been asked but it's not quite clear to me.</p>\n\n<p>I've seen it's been said that leaks won't be in the private set, does that mean that it won't at all affect the private score? Or is it possible that leaks can be used to train models better and private LB can be similar to public LB? I really don't care much about the leaked data since I just wanna learn from the competition.</p>\n\n<p>TL;DR\nWill it affect me on the private set if I just ignore leaks or is that only a setback in public LB?</p>",
  "messages": [
    {
      "id": 693155,
      "postDate": "2019-12-12T04:59:14.230Z",
      "content": "<p>Sorry maybe this has been asked but it's not quite clear to me.</p>\n\n<p>I've seen it's been said that leaks won't be in the private set, does that mean that it won't at all affect the private score? Or is it possible that leaks can be used to train models better and private LB can be similar to public LB? I really don't care much about the leaked data since I just wanna learn from the competition.</p>\n\n<p>TL;DR\nWill it affect me on the private set if I just ignore leaks or is that only a setback in public LB?</p>",
      "rawMarkdown": "Sorry maybe this has been asked but it's not quite clear to me.\n\nI've seen it's been said that leaks won't be in the private set, does that mean that it won't at all affect the private score? Or is it possible that leaks can be used to train models better and private LB can be similar to public LB? I really don't care much about the leaked data since I just wanna learn from the competition.\n\nTL;DR\nWill it affect me on the private set if I just ignore leaks or is that only a setback in public LB?",
      "votes": 6
    },
    {
      "id": 698382,
      "postDate": "2019-12-19T06:49:09.107Z",
      "content": "<p>One of my model without including leak data in training has lb 1.062, when I trained using leak data, lb improved to 1.056. It can be possible because I trained using leak data and then make prediction for the same. </p>\n\n<p>One way i could have tested it by setting meter readings in all the leak data in test set To zero and then compare the results of model trained with or without leak data, unfortunately don’t have submissions now.</p>",
      "rawMarkdown": "One of my model without including leak data in training has lb 1.062, when I trained using leak data, lb improved to 1.056. It can be possible because I trained using leak data and then make prediction for the same. \n\nOne way i could have tested it by setting meter readings in all the leak data in test set To zero and then compare the results of model trained with or without leak data, unfortunately don’t have submissions now.",
      "votes": 2,
      "replies": [
        {
          "id": 698386,
          "postDate": "2019-12-19T06:59:22.463Z",
          "content": "<p>The real question is does it improve other sites without leak?</p>",
          "rawMarkdown": "The real question is does it improve other sites without leak?"
        },
        {
          "id": 698402,
          "postDate": "2019-12-19T07:20:22.840Z",
          "content": "<p>We’ll know that tonight ;-)</p>\n\n<p>I honestly hope for you and Fred that you will stay Gold after the «&nbsp;Shake up&nbsp;».</p>",
          "rawMarkdown": "We’ll know that tonight ;-)\n\nI honestly hope for you and Fred that you will stay Gold after the «&nbsp;Shake up&nbsp;».\n"
        },
        {
          "id": 698560,
          "postDate": "2019-12-19T11:48:52.490Z",
          "content": "<p>I tested it, it improves.</p>",
          "rawMarkdown": "I tested it, it improves.",
          "votes": 1
        }
      ]
    },
    {
      "id": 693193,
      "postDate": "2019-12-12T06:06:09.667Z",
      "content": "<p>It is not directly useful but you can use leaked data to increase the training model size. </p>",
      "rawMarkdown": "It is not directly useful but you can use leaked data to increase the training model size. ",
      "votes": 2,
      "replies": [
        {
          "id": 693221,
          "postDate": "2019-12-12T06:43:02.933Z",
          "content": "<p>I see, but that may cause overfitting I suppose?</p>",
          "rawMarkdown": "I see, but that may cause overfitting I suppose?",
          "votes": 5
        },
        {
          "id": 693475,
          "postDate": "2019-12-12T12:18:33.183Z",
          "content": "<p>it didnt make my LB score better </p>",
          "rawMarkdown": "it didnt make my LB score better ",
          "votes": 4
        }
      ]
    },
    {
      "id": 696928,
      "postDate": "2019-12-17T08:48:30.490Z",
      "content": "<p>有泄露的数据吗 </p>",
      "rawMarkdown": "有泄露的数据吗 ",
      "votes": -10
    },
    {
      "id": 698400,
      "postDate": "2019-12-19T07:18:48.480Z",
      "content": "<p>Deleted...</p>",
      "rawMarkdown": "Deleted..."
    },
    {
      "id": 693651,
      "postDate": "2019-12-12T16:14:41.947Z",
      "content": "<p>That eases my mind</p>",
      "rawMarkdown": "That eases my mind"
    },
    {
      "id": 693553,
      "postDate": "2019-12-12T13:59:12.993Z",
      "content": "<p>the same question, is there a risk of cancelling my score if I use these leaky data?</p>",
      "rawMarkdown": "the same question, is there a risk of cancelling my score if I use these leaky data?",
      "replies": [
        {
          "id": 693639,
          "postDate": "2019-12-12T15:51:25.723Z",
          "content": "<p>I don't think so honestly</p>",
          "rawMarkdown": "I don't think so honestly",
          "votes": 1
        },
        {
          "id": 693652,
          "postDate": "2019-12-12T16:19:52.723Z",
          "content": "<p>No, they just won't score the rows with leaked data.</p>",
          "rawMarkdown": "No, they just won't score the rows with leaked data."
        },
        {
          "id": 693713,
          "postDate": "2019-12-12T17:56:50.253Z",
          "content": "<p>And how «&nbsp;they&nbsp;» should know your submission comes from leaked data ?</p>",
          "rawMarkdown": "And how «&nbsp;they&nbsp;» should know your submission comes from leaked data ?"
        },
        {
          "id": 693772,
          "postDate": "2019-12-12T19:23:43.887Z",
          "content": "<p>I mean, they will remove the rows whose meter reading is leaked publically before calculating the private leaderboard score.</p>",
          "rawMarkdown": "I mean, they will remove the rows whose meter reading is leaked publically before calculating the private leaderboard score."
        },
        {
          "id": 693777,
          "postDate": "2019-12-12T19:30:24.517Z",
          "content": "<p>This was not my comment : how will «&nbsp;they&nbsp;» find your submission comes from leaked data ? Imagine your guesses were right, imagine you’ve been creative, imagine you found the right algorithm... Who can prove that you cheated ? </p>\n\n<p>And moreover, if you used leaked data, Kaggle is responsible for that, and this is not the first time they do such a mistake, not the the first time and not the last. It happened so many times in the past.</p>",
          "rawMarkdown": "This was not my comment : how will «&nbsp;they&nbsp;» find your submission comes from leaked data ? Imagine your guesses were right, imagine you’ve been creative, imagine you found the right algorithm... Who can prove that you cheated ? \n\nAnd moreover, if you used leaked data, Kaggle is responsible for that, and this is not the first time they do such a mistake, not the the first time and not the last. It happened so many times in the past.",
          "votes": 1
        },
        {
          "id": 694054,
          "postDate": "2019-12-13T05:44:53.717Z",
          "content": "<p>You can use leaked data for training the model + submitting it. No problem. Kaggle will just remove the rows whose meter reading was leaked publically (like site-0,site-1,site-2,site-3,site-4,site15 I think). </p>\n\n<p>If you find a leaked data on the internet, but do not publish it on Kaggle. No one will know, and you will probably get a prize. </p>",
          "rawMarkdown": "You can use leaked data for training the model + submitting it. No problem. Kaggle will just remove the rows whose meter reading was leaked publically (like site-0,site-1,site-2,site-3,site-4,site15 I think). \n\nIf you find a leaked data on the internet, but do not publish it on Kaggle. No one will know, and you will probably get a prize. "
        },
        {
          "id": 697214,
          "postDate": "2019-12-17T16:15:51.287Z",
          "content": "<p>And you will eventually get disqualified ;)</p>\n\n<blockquote>\n  <p>If you find a leaked data on the internet, but do not publish it on Kaggle. No one will know, and you will probably get a prize.</p>\n</blockquote>\n\n<p>That's just stupid, sorry.</p>",
          "rawMarkdown": "And you will eventually get disqualified ;)\n\n&gt; If you find a leaked data on the internet, but do not publish it on Kaggle. No one will know, and you will probably get a prize.\n\nThat's just stupid, sorry."
        },
        {
          "id": 697245,
          "postDate": "2019-12-17T16:45:37.587Z",
          "content": "<p>It is not a kernel only competition, so if someone finds one, and do not publish it, there is a very good chance, he/she can get away with it. Right?</p>",
          "rawMarkdown": "It is not a kernel only competition, so if someone finds one, and do not publish it, there is a very good chance, he/she can get away with it. Right?"
        }
      ]
    },
    {
      "id": 705003,
      "postDate": "2019-12-28T09:33:32.207Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 698382,
      "author_name": "HarshitMehta",
      "author_url": "",
      "post_date": "2019-12-19T06:49:09.107000",
      "content": "<p>One of my model without including leak data in training has lb 1.062, when I trained using leak data, lb improved to 1.056. It can be possible because I trained using leak data and then make prediction for the same. </p>\n\n<p>One way i could have tested it by setting meter readings in all the leak data in test set To zero and then compare the results of model trained with or without leak data, unfortunately don’t have submissions now.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 698386,
          "author_name": "Gunes Evitan",
          "author_url": "",
          "post_date": "2019-12-19T06:59:22.463000",
          "content": "<p>The real question is does it improve other sites without leak?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 698402,
          "author_name": "mezoganet",
          "author_url": "",
          "post_date": "2019-12-19T07:20:22.840000",
          "content": "<p>We’ll know that tonight ;-)</p>\n\n<p>I honestly hope for you and Fred that you will stay Gold after the «&nbsp;Shake up&nbsp;».</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 698560,
          "author_name": "Kamal Chhirang",
          "author_url": "",
          "post_date": "2019-12-19T11:48:52.490000",
          "content": "<p>I tested it, it improves.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 693193,
      "author_name": "Kamal Chhirang",
      "author_url": "",
      "post_date": "2019-12-12T06:06:09.667000",
      "content": "<p>It is not directly useful but you can use leaked data to increase the training model size. </p>",
      "votes": 2,
      "replies": [
        {
          "id": 693221,
          "author_name": "Thiago Preischadt",
          "author_url": "",
          "post_date": "2019-12-12T06:43:02.933000",
          "content": "<p>I see, but that may cause overfitting I suppose?</p>",
          "votes": 5,
          "replies": []
        },
        {
          "id": 693475,
          "author_name": "Mohamad Falah",
          "author_url": "",
          "post_date": "2019-12-12T12:18:33.183000",
          "content": "<p>it didnt make my LB score better </p>",
          "votes": 4,
          "replies": []
        }
      ]
    },
    {
      "id": 696928,
      "author_name": "zhengchenglong",
      "author_url": "",
      "post_date": "2019-12-17T08:48:30.490000",
      "content": "<p>有泄露的数据吗 </p>",
      "votes": -10,
      "replies": []
    },
    {
      "id": 698400,
      "author_name": "mezoganet",
      "author_url": "",
      "post_date": "2019-12-19T07:18:48.480000",
      "content": "<p>Deleted...</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 693651,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-12T16:14:41.947000",
      "content": "<p>That eases my mind</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 693553,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-12T13:59:12.993000",
      "content": "<p>the same question, is there a risk of cancelling my score if I use these leaky data?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 693639,
          "author_name": "Thiago Preischadt",
          "author_url": "",
          "post_date": "2019-12-12T15:51:25.723000",
          "content": "<p>I don't think so honestly</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 693652,
          "author_name": "Kamal Chhirang",
          "author_url": "",
          "post_date": "2019-12-12T16:19:52.723000",
          "content": "<p>No, they just won't score the rows with leaked data.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 693713,
          "author_name": "mezoganet",
          "author_url": "",
          "post_date": "2019-12-12T17:56:50.253000",
          "content": "<p>And how «&nbsp;they&nbsp;» should know your submission comes from leaked data ?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 693772,
          "author_name": "Kamal Chhirang",
          "author_url": "",
          "post_date": "2019-12-12T19:23:43.887000",
          "content": "<p>I mean, they will remove the rows whose meter reading is leaked publically before calculating the private leaderboard score.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 693777,
          "author_name": "mezoganet",
          "author_url": "",
          "post_date": "2019-12-12T19:30:24.517000",
          "content": "<p>This was not my comment : how will «&nbsp;they&nbsp;» find your submission comes from leaked data ? Imagine your guesses were right, imagine you’ve been creative, imagine you found the right algorithm... Who can prove that you cheated ? </p>\n\n<p>And moreover, if you used leaked data, Kaggle is responsible for that, and this is not the first time they do such a mistake, not the the first time and not the last. It happened so many times in the past.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 694054,
          "author_name": "Kamal Chhirang",
          "author_url": "",
          "post_date": "2019-12-13T05:44:53.717000",
          "content": "<p>You can use leaked data for training the model + submitting it. No problem. Kaggle will just remove the rows whose meter reading was leaked publically (like site-0,site-1,site-2,site-3,site-4,site15 I think). </p>\n\n<p>If you find a leaked data on the internet, but do not publish it on Kaggle. No one will know, and you will probably get a prize. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 697214,
          "author_name": "danzel",
          "author_url": "",
          "post_date": "2019-12-17T16:15:51.287000",
          "content": "<p>And you will eventually get disqualified ;)</p>\n\n<blockquote>\n  <p>If you find a leaked data on the internet, but do not publish it on Kaggle. No one will know, and you will probably get a prize.</p>\n</blockquote>\n\n<p>That's just stupid, sorry.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 697245,
          "author_name": "Kamal Chhirang",
          "author_url": "",
          "post_date": "2019-12-17T16:45:37.587000",
          "content": "<p>It is not a kernel only competition, so if someone finds one, and do not publish it, there is a very good chance, he/she can get away with it. Right?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 705003,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-12-28T09:33:32.207000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "693155": "Sorry maybe this has been asked but it's not quite clear to me.\n\nI've seen it's been said that leaks won't be in the private set, does that mean that it won't at all affect the private score? Or is it possible that leaks can be used to train models better and private LB can be similar to public LB? I really don't care much about the leaked data since I just wanna learn from the competition.\n\nTL;DR\nWill it affect me on the private set if I just ignore leaks or is that only a setback in public LB?",
    "698382": "One of my model without including leak data in training has lb 1.062, when I trained using leak data, lb improved to 1.056. It can be possible because I trained using leak data and then make prediction for the same. \n\nOne way i could have tested it by setting meter readings in all the leak data in test set To zero and then compare the results of model trained with or without leak data, unfortunately don’t have submissions now.",
    "693193": "It is not directly useful but you can use leaked data to increase the training model size. ",
    "696928": "有泄露的数据吗 ",
    "698400": "Deleted...",
    "693651": "That eases my mind",
    "693553": "the same question, is there a risk of cancelling my score if I use these leaky data?",
    "705003": ""
  }
}