{
  "id": 12455,
  "title": "Awfully quiet",
  "url": "/competitions/avazu-ctr-prediction/discussion/12455",
  "author_name": "",
  "post_date": "2015-02-09T20:10:58.117Z",
  "votes": 2,
  "comment_count": 9,
  "views": 1948,
  "content": "<p>So let me break the silence :). I'm running SGD AdaGrad (C# implementation no hashing - dicts instead). 2 epohs, 18 interactions, no L2 and L1 in form of feature occurrence. I'm not using single model, but like 1500 of them built around device_ip / app_id and site_id. They are all trained on the fly, but only some of them end up making final predictions.</p>\n<p>I hopped for better LB score from them, but I guess: sometimes less is more.</p>\n<p>I've found most problems with site_id 17d1b03f which does not exists in train set but is 25% of test set. Didn't really found out how to effectively handle it.</p>\n<p>Anyway, it was a fun, looking forward to read about other solutions.</p>\n<p>Good luck to you all.</p>\n<p>br,</p>\n<p>Goran M.</p>",
  "messages": [
    {
      "id": "63824",
      "postDate": "02/09/2015 20:10:58",
      "content": "<p>So let me break the silence :). I'm running SGD AdaGrad (C# implementation no hashing - dicts instead). 2 epohs, 18 interactions, no L2 and L1 in form of feature occurrence. I'm not using single model, but like 1500 of them built around device_ip / app_id and site_id. They are all trained on the fly, but only some of them end up making final predictions.</p>\n<p>I hopped for better LB score from them, but I guess: sometimes less is more.</p>\n<p>I've found most problems with site_id 17d1b03f which does not exists in train set but is 25% of test set. Didn't really found out how to effectively handle it.</p>\n<p>Anyway, it was a fun, looking forward to read about other solutions.</p>\n<p>Good luck to you all.</p>\n<p>br,</p>\n<p>Goran M.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63827",
      "postDate": "02/09/2015 20:21:50",
      "content": "<p>Mayby you can put somewhere in your algorithm average of clicks column when it ask about&nbsp; 17d1b03f .</p>\n\n<p>I'm using regression on manipulated dataset, it achives OOS = 0.3820499, but can't scale to full size solution due to memory problems with R (merging and manipulating two large ffdf objects is horrible).</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63828",
      "postDate": "02/09/2015 20:37:18",
      "content": "<p>And maybe it's not a bad idea - tuning it manually, maybe next time. I'm out of submissions anyway :).</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63829",
      "postDate": "02/09/2015 20:48:38",
      "content": "<p>Inputation of random class would be somehow equivalent to using average value or response variable.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63830",
      "postDate": "02/09/2015 20:53:53",
      "content": "<p>There's an insight into the data that I'm not seeing. How do I know this? I've approached this problem many different ways, and they all plateau at around my current LB score. The fact the Top 5 is a step change away from the pack strongly suggests the rest of us are missing something.</p>\n<p>I'm looking forward to learning what that was, so I can kick myself. &nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63831",
      "postDate": "02/09/2015 21:02:20",
      "content": "<p>Just my guess : I think there is no insight only spliting dataset in right way and then combining the outputs, spliting for example on hour etc. </p>\n\n<p>Does 0.01 of difference in LogLosses -&gt; 1% difference in classification error (1 when ==, 0 otherwise) ?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63832",
      "postDate": "02/09/2015 21:13:39",
      "content": "<p>Guess we're all waiting for the resolution and descriptions of the winners solutions.</p>\n<p>I myself can hardly wait to read about what the winners did in terms of feature engineering!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63835",
      "postDate": "02/09/2015 21:21:15",
      "content": "<p>Don't say too much yet.. I'm looking for an 0.00004 improvement.&nbsp;<br>The person above me&nbsp;might not&nbsp;like it if I got a hunch from the forum..</p>\n<p>A&nbsp;C# adagrad implementation in 2.5 hours would be a bit of a&nbsp;harsh schedule so no harm done yet :)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63837",
      "postDate": "02/09/2015 21:37:44",
      "content": "<p>[quote=Julian de Wit;63835]</p>\n<p>Don't say too much yet.. I'm looking for an 0.00004 improvement.&nbsp;<br>The person above me&nbsp;might not&nbsp;like it if I got a hunch from the forum..</p>\n<p>A&nbsp;C# adagrad implementation in 2.5 hours would be a bit of a&nbsp;harsh schedule so no harm done yet :)</p>\n<p>[/quote]</p>\n<p>Sure thing. Good luck!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63838",
      "postDate": "02/09/2015 21:53:26",
      "content": "<p>I have a selfish request. One of my favorite parts of the competition is coming up, when the winners and other top competitors start generously posting their winning solutions.&nbsp;Usually these&nbsp;are&nbsp;pretty specific descriptions of algorithms and models.</p>\n<p>For my own edification, I'd love&nbsp;to hear more about how people approached the problem from a more general meta sense. How you&nbsp;thought about the problem, your&nbsp;cross-validation approach, and how you&nbsp;decided on which models or features to use, or other conceptual insights.&nbsp;Thanks in advance for your generosity in sharing your knowledge!!!</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 63827,
      "author_name": "oneoff",
      "author_url": "",
      "post_date": "02/09/2015 20:21:50",
      "content": "<p>Mayby you can put somewhere in your algorithm average of clicks column when it ask about&nbsp; 17d1b03f .</p>\n\n<p>I'm using regression on manipulated dataset, it achives OOS = 0.3820499, but can't scale to full size solution due to memory problems with R (merging and manipulating two large ffdf objects is horrible).</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63828,
      "author_name": "gmilosev",
      "author_url": "",
      "post_date": "02/09/2015 20:37:18",
      "content": "<p>And maybe it's not a bad idea - tuning it manually, maybe next time. I'm out of submissions anyway :).</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63829,
      "author_name": "oneoff",
      "author_url": "",
      "post_date": "02/09/2015 20:48:38",
      "content": "<p>Inputation of random class would be somehow equivalent to using average value or response variable.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63830,
      "author_name": "inversion",
      "author_url": "",
      "post_date": "02/09/2015 20:53:53",
      "content": "<p>There's an insight into the data that I'm not seeing. How do I know this? I've approached this problem many different ways, and they all plateau at around my current LB score. The fact the Top 5 is a step change away from the pack strongly suggests the rest of us are missing something.</p>\n<p>I'm looking forward to learning what that was, so I can kick myself. &nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63831,
      "author_name": "oneoff",
      "author_url": "",
      "post_date": "02/09/2015 21:02:20",
      "content": "<p>Just my guess : I think there is no insight only spliting dataset in right way and then combining the outputs, spliting for example on hour etc. </p>\n\n<p>Does 0.01 of difference in LogLosses -&gt; 1% difference in classification error (1 when ==, 0 otherwise) ?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63832,
      "author_name": "tobycheese",
      "author_url": "",
      "post_date": "02/09/2015 21:13:39",
      "content": "<p>Guess we're all waiting for the resolution and descriptions of the winners solutions.</p>\n<p>I myself can hardly wait to read about what the winners did in terms of feature engineering!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63835,
      "author_name": "juliandewit",
      "author_url": "",
      "post_date": "02/09/2015 21:21:15",
      "content": "<p>Don't say too much yet.. I'm looking for an 0.00004 improvement.&nbsp;<br>The person above me&nbsp;might not&nbsp;like it if I got a hunch from the forum..</p>\n<p>A&nbsp;C# adagrad implementation in 2.5 hours would be a bit of a&nbsp;harsh schedule so no harm done yet :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63837,
      "author_name": "inversion",
      "author_url": "",
      "post_date": "02/09/2015 21:37:44",
      "content": "<p>[quote=Julian de Wit;63835]</p>\n<p>Don't say too much yet.. I'm looking for an 0.00004 improvement.&nbsp;<br>The person above me&nbsp;might not&nbsp;like it if I got a hunch from the forum..</p>\n<p>A&nbsp;C# adagrad implementation in 2.5 hours would be a bit of a&nbsp;harsh schedule so no harm done yet :)</p>\n<p>[/quote]</p>\n<p>Sure thing. Good luck!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63838,
      "author_name": "danielyoo",
      "author_url": "",
      "post_date": "02/09/2015 21:53:26",
      "content": "<p>I have a selfish request. One of my favorite parts of the competition is coming up, when the winners and other top competitors start generously posting their winning solutions.&nbsp;Usually these&nbsp;are&nbsp;pretty specific descriptions of algorithms and models.</p>\n<p>For my own edification, I'd love&nbsp;to hear more about how people approached the problem from a more general meta sense. How you&nbsp;thought about the problem, your&nbsp;cross-validation approach, and how you&nbsp;decided on which models or features to use, or other conceptual insights.&nbsp;Thanks in advance for your generosity in sharing your knowledge!!!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "63824": "",
    "63827": "",
    "63828": "",
    "63829": "",
    "63830": "",
    "63831": "",
    "63832": "",
    "63835": "",
    "63837": "",
    "63838": ""
  },
  "source": "meta"
}