{
  "id": 68319,
  "title": "Potential Solutions to Leak",
  "url": "/competitions/ga-customer-revenue-prediction/discussion/68319",
  "author_name": "",
  "post_date": "2018-10-11T12:21:33.194778800Z",
  "votes": 6,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Let's share our ideas for how Kaggle can address the leaked data.</p>\n\n<p>My idea is twofold. First, change to a kernel's only competition and release all the data (up to the current date) as the training set. Then have the test set be the data that is developed after the end of the competition (similar to Two Sigma) competition. The biggest downside is that we won't know the competition results until a while after the competition ends (need time for the data to develop), but I would like to hear other ideas!</p>",
  "messages": [
    {
      "id": "402253",
      "postDate": "10/11/2018 12:21:33",
      "content": "<p>Let's share our ideas for how Kaggle can address the leaked data.</p>\n\n<p>My idea is twofold. First, change to a kernel's only competition and release all the data (up to the current date) as the training set. Then have the test set be the data that is developed after the end of the competition (similar to Two Sigma) competition. The biggest downside is that we won't know the competition results until a while after the competition ends (need time for the data to develop), but I would like to hear other ideas!</p>",
      "rawMarkdown": "Let's share our ideas for how Kaggle can address the leaked data.\n\nMy idea is twofold. First, change to a kernel's only competition and release all the data (up to the current date) as the training set. Then have the test set be the data that is developed after the end of the competition (similar to Two Sigma) competition. The biggest downside is that we won't know the competition results until a while after the competition ends (need time for the data to develop), but I would like to hear other ideas!",
      "votes": null
    },
    {
      "id": "402267",
      "postDate": "10/11/2018 12:40:09",
      "content": "<p>That is a better way of handling time-series competitions in general. Another option could be to use data from a different store for the same dates where the store's information is not publicly available. Either way there could be quite a lot of rework involved, but in both cases some existing work could be reusable.</p>",
      "rawMarkdown": "That is a better way of handling time-series competitions in general. Another option could be to use data from a different store for the same dates where the store's information is not publicly available. Either way there could be quite a lot of rework involved, but in both cases some existing work could be reusable.",
      "votes": null
    },
    {
      "id": "402416",
      "postDate": "10/11/2018 17:20:32",
      "content": "<p>Using data from a different store would be a really good solution. I think another option would be cancelling the competition, but that's a lot of time wasted for kagglers. I didn't check the leak, but the way people are describing it we would spend more time looking how to connect the two sources than doing ml or anything usefull...</p>",
      "rawMarkdown": "Using data from a different store would be a really good solution. I think another option would be cancelling the competition, but that's a lot of time wasted for kagglers. I didn't check the leak, but the way people are describing it we would spend more time looking how to connect the two sources than doing ml or anything usefull...",
      "votes": null
    },
    {
      "id": "402452",
      "postDate": "10/11/2018 18:32:45",
      "content": "<p>I am new to this but I am sure there must have been leaks before and if yes then what did kaggle do deal with it. Does anyone know and why is kaggle not doing anything now?</p>",
      "rawMarkdown": "I am new to this but I am sure there must have been leaks before and if yes then what did kaggle do deal with it. Does anyone know and why is kaggle not doing anything now?",
      "votes": null
    },
    {
      "id": "402523",
      "postDate": "10/11/2018 20:33:15",
      "content": "<p>That could be a good way to continue this competition, throwing away not all the work, but some part of it (may be larger part of course). As already mentioned in this thread, test test after the end of competition would make it more \"time-series\" competition than it is now and was before the last leak.</p>",
      "rawMarkdown": "That could be a good way to continue this competition, throwing away not all the work, but some part of it (may be larger part of course). As already mentioned in this thread, test test after the end of competition would make it more \"time-series\" competition than it is now and was before the last leak.",
      "votes": null
    },
    {
      "id": "402658",
      "postDate": "10/12/2018 04:24:07",
      "content": "<p>For different store, the feature importance revenue distribution will be different, so it's unfair for those who already spent time on feature selection or feature engineering, and it is same as restarting the competition to me...</p>",
      "rawMarkdown": "For different store, the feature importance revenue distribution will be different, so it's unfair for those who already spent time on feature selection or feature engineering, and it is same as restarting the competition to me...",
      "votes": null
    },
    {
      "id": "402677",
      "postDate": "10/12/2018 05:39:31",
      "content": "<p>How Abt making this competition a classification problem ? I.e whether a particular session will be a convert or not.. and deleting few columns like \"bounce\" which is an obvoius indicator that session isn't converted to a purchase</p>",
      "rawMarkdown": "How Abt making this competition a classification problem ? I.e whether a particular session will be a convert or not.. and deleting few columns like \"bounce\" which is an obvoius indicator that session isn't converted to a purchase",
      "votes": null
    },
    {
      "id": "402707",
      "postDate": "10/12/2018 07:09:48",
      "content": "<p>Not good enough. pageviews=1 is basically the same, so you would also have to delete that one. And this is the number 1 predictor.</p>",
      "rawMarkdown": "Not good enough. pageviews=1 is basically the same, so you would also have to delete that one. And this is the number 1 predictor.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 402267,
      "author_name": "datahobbit",
      "author_url": "",
      "post_date": "10/11/2018 12:40:09",
      "content": "<p>That is a better way of handling time-series competitions in general. Another option could be to use data from a different store for the same dates where the store's information is not publicly available. Either way there could be quite a lot of rework involved, but in both cases some existing work could be reusable.</p>",
      "votes": null,
      "replies": [
        {
          "id": 402416,
          "author_name": "jsaguiar",
          "author_url": "",
          "post_date": "10/11/2018 17:20:32",
          "content": "<p>Using data from a different store would be a really good solution. I think another option would be cancelling the competition, but that's a lot of time wasted for kagglers. I didn't check the leak, but the way people are describing it we would spend more time looking how to connect the two sources than doing ml or anything usefull...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 402452,
          "author_name": "moizzz",
          "author_url": "",
          "post_date": "10/11/2018 18:32:45",
          "content": "<p>I am new to this but I am sure there must have been leaks before and if yes then what did kaggle do deal with it. Does anyone know and why is kaggle not doing anything now?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 402658,
          "author_name": "joe1995",
          "author_url": "",
          "post_date": "10/12/2018 04:24:07",
          "content": "<p>For different store, the feature importance revenue distribution will be different, so it's unfair for those who already spent time on feature selection or feature engineering, and it is same as restarting the competition to me...</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 402523,
      "author_name": "chesnokov",
      "author_url": "",
      "post_date": "10/11/2018 20:33:15",
      "content": "<p>That could be a good way to continue this competition, throwing away not all the work, but some part of it (may be larger part of course). As already mentioned in this thread, test test after the end of competition would make it more \"time-series\" competition than it is now and was before the last leak.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 402677,
      "author_name": "rajuspartan",
      "author_url": "",
      "post_date": "10/12/2018 05:39:31",
      "content": "<p>How Abt making this competition a classification problem ? I.e whether a particular session will be a convert or not.. and deleting few columns like \"bounce\" which is an obvoius indicator that session isn't converted to a purchase</p>",
      "votes": null,
      "replies": [
        {
          "id": 402707,
          "author_name": "erikbruin",
          "author_url": "",
          "post_date": "10/12/2018 07:09:48",
          "content": "<p>Not good enough. pageviews=1 is basically the same, so you would also have to delete that one. And this is the number 1 predictor.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "402253": "Let's share our ideas for how Kaggle can address the leaked data.\n\nMy idea is twofold. First, change to a kernel's only competition and release all the data (up to the current date) as the training set. Then have the test set be the data that is developed after the end of the competition (similar to Two Sigma) competition. The biggest downside is that we won't know the competition results until a while after the competition ends (need time for the data to develop), but I would like to hear other ideas!",
    "402267": "That is a better way of handling time-series competitions in general. Another option could be to use data from a different store for the same dates where the store's information is not publicly available. Either way there could be quite a lot of rework involved, but in both cases some existing work could be reusable.",
    "402416": "Using data from a different store would be a really good solution. I think another option would be cancelling the competition, but that's a lot of time wasted for kagglers. I didn't check the leak, but the way people are describing it we would spend more time looking how to connect the two sources than doing ml or anything usefull...",
    "402452": "I am new to this but I am sure there must have been leaks before and if yes then what did kaggle do deal with it. Does anyone know and why is kaggle not doing anything now?",
    "402523": "That could be a good way to continue this competition, throwing away not all the work, but some part of it (may be larger part of course). As already mentioned in this thread, test test after the end of competition would make it more \"time-series\" competition than it is now and was before the last leak.",
    "402658": "For different store, the feature importance revenue distribution will be different, so it's unfair for those who already spent time on feature selection or feature engineering, and it is same as restarting the competition to me...",
    "402677": "How Abt making this competition a classification problem ? I.e whether a particular session will be a convert or not.. and deleting few columns like \"bounce\" which is an obvoius indicator that session isn't converted to a purchase",
    "402707": "Not good enough. pageviews=1 is basically the same, so you would also have to delete that one. And this is the number 1 predictor."
  },
  "source": "meta"
}