{
  "id": 52036,
  "title": "Future Feature",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/52036",
  "author_name": "",
  "post_date": "2018-03-15T11:52:06.215588600Z",
  "votes": 6,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Are we allowed to use future features?\nMost of kernels are using those. If so, why?</p>",
  "messages": [
    {
      "id": "296535",
      "postDate": "03/15/2018 11:52:06",
      "content": "<p>Are we allowed to use future features?\nMost of kernels are using those. If so, why?</p>",
      "rawMarkdown": "Are we allowed to use future features?\nMost of kernels are using those. If so, why?",
      "votes": null
    },
    {
      "id": "296537",
      "postDate": "03/15/2018 11:53:05",
      "content": "<p>Could you elaborate what do you mean by future feature? I've seen several of the popular kernels and I haven't noticed anything of the sort...</p>",
      "rawMarkdown": "Could you elaborate what do you mean by future feature? I've seen several of the popular kernels and I haven't noticed anything of the sort...",
      "votes": null
    },
    {
      "id": "296541",
      "postDate": "03/15/2018 12:00:17",
      "content": "<p>For instance this <a href=\"https://www.kaggle.com/rteja1113/lightgbm-with-count-features\">kernel</a>, the feature are frequency of access. And training sets are the past before the access.</p>",
      "rawMarkdown": "For instance this [kernel][1], the feature are frequency of access. And training sets are the past before the access.\n\n\n  [1]: https://www.kaggle.com/rteja1113/lightgbm-with-count-features",
      "votes": null
    },
    {
      "id": "296542",
      "postDate": "03/15/2018 12:01:53",
      "content": "<p>Ok, from a business perspective: yes, it's useless since future info is utilized. In the purely ML/Kaggle context, labels are not used =&gt; I think risk of leakage is minimal.</p>",
      "rawMarkdown": "Ok, from a business perspective: yes, it's useless since future info is utilized. In the purely ML/Kaggle context, labels are not used =&gt; I think risk of leakage is minimal.",
      "votes": null
    },
    {
      "id": "296545",
      "postDate": "03/15/2018 12:08:29",
      "content": "<p>Hmm...🤔 highly doubt it, but thank you for letting me hear your thought!</p>",
      "rawMarkdown": "Hmm...🤔 highly doubt it, but thank you for letting me hear your thought!",
      "votes": null
    },
    {
      "id": "297108",
      "postDate": "03/16/2018 09:46:32",
      "content": "<p>Hi,\nFuture features are important in this competition and also  business context as well for this task.\nThat's why there is \"attributed_time\". As shown by many kernals there are 24 hour gap some times with click time and downloaded time. As an example a fraudulent IP may has first click and then number of clicks after one hour may 10000. So that information is important to predict the download ability of first click  </p>",
      "rawMarkdown": "Hi,\nFuture features are important in this competition and also  business context as well for this task.\nThat's why there is \"attributed_time\". As shown by many kernals there are 24 hour gap some times with click time and downloaded time. As an example a fraudulent IP may has first click and then number of clicks after one hour may 10000. So that information is important to predict the download ability of first click",
      "votes": null
    },
    {
      "id": "297113",
      "postDate": "03/16/2018 10:15:37",
      "content": "<p>Fair enough!! I could understand this problem!!! Thank you!!!</p>",
      "rawMarkdown": "Fair enough!! I could understand this problem!!! Thank you!!!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 296537,
      "author_name": "konradb",
      "author_url": "",
      "post_date": "03/15/2018 11:53:05",
      "content": "<p>Could you elaborate what do you mean by future feature? I've seen several of the popular kernels and I haven't noticed anything of the sort...</p>",
      "votes": null,
      "replies": [
        {
          "id": 296541,
          "author_name": "onodera",
          "author_url": "",
          "post_date": "03/15/2018 12:00:17",
          "content": "<p>For instance this <a href=\"https://www.kaggle.com/rteja1113/lightgbm-with-count-features\">kernel</a>, the feature are frequency of access. And training sets are the past before the access.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 296542,
          "author_name": "konradb",
          "author_url": "",
          "post_date": "03/15/2018 12:01:53",
          "content": "<p>Ok, from a business perspective: yes, it's useless since future info is utilized. In the purely ML/Kaggle context, labels are not used =&gt; I think risk of leakage is minimal.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 296545,
          "author_name": "onodera",
          "author_url": "",
          "post_date": "03/15/2018 12:08:29",
          "content": "<p>Hmm...🤔 highly doubt it, but thank you for letting me hear your thought!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 297108,
      "author_name": "isaranja",
      "author_url": "",
      "post_date": "03/16/2018 09:46:32",
      "content": "<p>Hi,\nFuture features are important in this competition and also  business context as well for this task.\nThat's why there is \"attributed_time\". As shown by many kernals there are 24 hour gap some times with click time and downloaded time. As an example a fraudulent IP may has first click and then number of clicks after one hour may 10000. So that information is important to predict the download ability of first click  </p>",
      "votes": null,
      "replies": [
        {
          "id": 297113,
          "author_name": "onodera",
          "author_url": "",
          "post_date": "03/16/2018 10:15:37",
          "content": "<p>Fair enough!! I could understand this problem!!! Thank you!!!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "296535": "Are we allowed to use future features?\nMost of kernels are using those. If so, why?",
    "296537": "Could you elaborate what do you mean by future feature? I've seen several of the popular kernels and I haven't noticed anything of the sort...",
    "296541": "For instance this [kernel][1], the feature are frequency of access. And training sets are the past before the access.\n\n\n  [1]: https://www.kaggle.com/rteja1113/lightgbm-with-count-features",
    "296542": "Ok, from a business perspective: yes, it's useless since future info is utilized. In the purely ML/Kaggle context, labels are not used =&gt; I think risk of leakage is minimal.",
    "296545": "Hmm...🤔 highly doubt it, but thank you for letting me hear your thought!",
    "297108": "Hi,\nFuture features are important in this competition and also  business context as well for this task.\nThat's why there is \"attributed_time\". As shown by many kernals there are 24 hour gap some times with click time and downloaded time. As an example a fraudulent IP may has first click and then number of clicks after one hour may 10000. So that information is important to predict the download ability of first click",
    "297113": "Fair enough!! I could understand this problem!!! Thank you!!!"
  },
  "source": "meta"
}