{
  "id": 11264,
  "title": "Separating app and web ads?",
  "url": "/competitions/avazu-ctr-prediction/discussion/11264",
  "author_name": "",
  "post_date": "2014-12-16T14:09:08.977Z",
  "votes": 1,
  "comment_count": 10,
  "views": 3374,
  "content": "<p>Since the app* features are null for web ads and site* features are null for app ads, I tried separating them by training two models and got moderate improvements on both one day and two days validation.However, the LB score just went the opposite way. Anyone had luck doing this?</p>",
  "messages": [
    {
      "id": "60351",
      "postDate": "12/16/2014 14:09:08",
      "content": "<p>Since the app* features are null for web ads and site* features are null for app ads, I tried separating them by training two models and got moderate improvements on both one day and two days validation.However, the LB score just went the opposite way. Anyone had luck doing this?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "60354",
      "postDate": "12/16/2014 14:22:32",
      "content": "<p>yep, it improves 0.0003 of our ensemble</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "60356",
      "postDate": "12/16/2014 14:34:21",
      "content": "<p>I did this also but didn't see improvement on the leaderboard.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "60357",
      "postDate": "12/16/2014 14:46:38",
      "content": "<p>Thanks for the reply :). According to my previous results, the change on leaderboard has been between changes on my two validation sets(the 10th day and the last two days). It improves 0.0005 on both validation sets, but gets 0.0005 loss up on the leaderboard which I haven't figured out.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "60365",
      "postDate": "12/16/2014 16:26:13",
      "content": "<p>I've checked the app* and site* features and I don't see any nulls, how did you determine they could be separated?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "60368",
      "postDate": "12/16/2014 16:42:53",
      "content": "<p>[quote=Jerome Conan;60365]</p>\n<p>I've checked the app* and site* features and I don't see any nulls, how did you determine they could be separated?</p>\n<p>[/quote]</p>\n<p>They are hashed as &quot;85f751fd&quot; for site_id and &quot;ecad2386&quot; for app_id.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "61461",
      "postDate": "12/18/2014 09:49:39",
      "content": "<p>&nbsp;how can we say for sure that &quot;85f751fd&quot; for site_id and &quot;ecad2386&quot; for app_id are null values.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "61468",
      "postDate": "12/18/2014 11:11:45",
      "content": "<p>[quote=avinashj;61461]</p>\n<p>&nbsp;how can we say for sure that &quot;85f751fd&quot; for site_id and &quot;ecad2386&quot; for app_id are null values.&nbsp;</p>\n<p>[/quote]</p>\n<p>Each line contains exactly one of them. :)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63672",
      "postDate": "02/06/2015 01:48:27",
      "content": "<p>Isn't it&nbsp;1fbe01fe for site_id? I just counted the frequency of&nbsp;the levels of each variable. The null one is bound to be the most common one in both cases.</p>\n<p>Probably other variables can be used to divide the dataset.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63690",
      "postDate": "02/06/2015 11:46:46",
      "content": "<p>[quote=marbel;63672]</p>\n<p>Isn't it&nbsp;1fbe01fe for site_id? I just counted the frequency of&nbsp;the levels of each variable. The null one is bound to be the most common one in both cases.</p>\n<p>Probably other variables can be used to divide the dataset.&nbsp;</p>\n<p>[/quote]</p>\n<p>The null is not&nbsp;necessarily the most common one. Suppose there are 10000 app ads and 20000 site ads, the null for site appears 10000 times, while the site_id for a site&nbsp;with a significant amount of traffic may appear 15000 times.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63840",
      "postDate": "02/09/2015 22:26:19",
      "content": "<p>This helped a bit,&nbsp;0.0002021compared to the submission I previously had.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 60354,
      "author_name": "jiweiliu",
      "author_url": "",
      "post_date": "12/16/2014 14:22:32",
      "content": "<p>yep, it improves 0.0003 of our ensemble</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 60356,
      "author_name": "jwjohnson314",
      "author_url": "",
      "post_date": "12/16/2014 14:34:21",
      "content": "<p>I did this also but didn't see improvement on the leaderboard.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 60357,
      "author_name": "infinite",
      "author_url": "",
      "post_date": "12/16/2014 14:46:38",
      "content": "<p>Thanks for the reply :). According to my previous results, the change on leaderboard has been between changes on my two validation sets(the 10th day and the last two days). It improves 0.0005 on both validation sets, but gets 0.0005 loss up on the leaderboard which I haven't figured out.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 60365,
      "author_name": "jeromic",
      "author_url": "",
      "post_date": "12/16/2014 16:26:13",
      "content": "<p>I've checked the app* and site* features and I don't see any nulls, how did you determine they could be separated?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 60368,
      "author_name": "infinite",
      "author_url": "",
      "post_date": "12/16/2014 16:42:53",
      "content": "<p>[quote=Jerome Conan;60365]</p>\n<p>I've checked the app* and site* features and I don't see any nulls, how did you determine they could be separated?</p>\n<p>[/quote]</p>\n<p>They are hashed as &quot;85f751fd&quot; for site_id and &quot;ecad2386&quot; for app_id.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 61461,
      "author_name": "avinashj",
      "author_url": "",
      "post_date": "12/18/2014 09:49:39",
      "content": "<p>&nbsp;how can we say for sure that &quot;85f751fd&quot; for site_id and &quot;ecad2386&quot; for app_id are null values.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 61468,
      "author_name": "infinite",
      "author_url": "",
      "post_date": "12/18/2014 11:11:45",
      "content": "<p>[quote=avinashj;61461]</p>\n<p>&nbsp;how can we say for sure that &quot;85f751fd&quot; for site_id and &quot;ecad2386&quot; for app_id are null values.&nbsp;</p>\n<p>[/quote]</p>\n<p>Each line contains exactly one of them. :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63672,
      "author_name": "marbel",
      "author_url": "",
      "post_date": "02/06/2015 01:48:27",
      "content": "<p>Isn't it&nbsp;1fbe01fe for site_id? I just counted the frequency of&nbsp;the levels of each variable. The null one is bound to be the most common one in both cases.</p>\n<p>Probably other variables can be used to divide the dataset.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63690,
      "author_name": "infinite",
      "author_url": "",
      "post_date": "02/06/2015 11:46:46",
      "content": "<p>[quote=marbel;63672]</p>\n<p>Isn't it&nbsp;1fbe01fe for site_id? I just counted the frequency of&nbsp;the levels of each variable. The null one is bound to be the most common one in both cases.</p>\n<p>Probably other variables can be used to divide the dataset.&nbsp;</p>\n<p>[/quote]</p>\n<p>The null is not&nbsp;necessarily the most common one. Suppose there are 10000 app ads and 20000 site ads, the null for site appears 10000 times, while the site_id for a site&nbsp;with a significant amount of traffic may appear 15000 times.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63840,
      "author_name": "marbel",
      "author_url": "",
      "post_date": "02/09/2015 22:26:19",
      "content": "<p>This helped a bit,&nbsp;0.0002021compared to the submission I previously had.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "60351": "",
    "60354": "",
    "60356": "",
    "60357": "",
    "60365": "",
    "60368": "",
    "61461": "",
    "61468": "",
    "63672": "",
    "63690": "",
    "63840": ""
  },
  "source": "meta"
}