{
  "id": 5960,
  "title": " is there a single model which can reach to 0.5?",
  "url": "/competitions/expedia-personalized-sort/discussion/5960",
  "author_name": "",
  "post_date": "2013-10-04T12:43:43.500Z",
  "votes": null,
  "comment_count": 15,
  "views": 5063,
  "content": "<p>is there a single model which can reach to 0.5?</p>\n<p>thanks!</p>",
  "messages": [
    {
      "id": "32021",
      "postDate": "10/04/2013 12:43:43",
      "content": "<p>is there a single model which can reach to 0.5?</p>\n<p>thanks!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32029",
      "postDate": "10/04/2013 16:05:43",
      "content": "<p>Yes. :)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32031",
      "postDate": "10/04/2013 17:13:20",
      "content": "<p>Almost there!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32039",
      "postDate": "10/04/2013 23:49:59",
      "content": "<p>I use 2 models. Each one can beat 0.5.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32046",
      "postDate": "10/05/2013 07:37:50",
      "content": "<p>good job</p>\n<p>I use LambdaMART now , but hard to promote 0.46.</p>\n<p>are you generate other features which is not include in datasets?</p>\n<p>add cross-features?</p>\n<p>&nbsp;</p>\n<p>thanks!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32047",
      "postDate": "10/05/2013 07:46:19",
      "content": "<p>can give me some advice?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32048",
      "postDate": "10/05/2013 08:31:26",
      "content": "<p>Indeed, you need to do a bit of feature engineering first. R with a lot of RAM - more than 16 Gb - can do the job. Sorry I can't tell you more at this stage!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32054",
      "postDate": "10/05/2013 18:26:50",
      "content": "<p>With a laptop of 8GB RAM, you still can beat 0.5. Of course, it would require a bit more programming experience. &nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32057",
      "postDate": "10/06/2013 02:06:46",
      "content": "<p>I think my problem now : do not well in feature scale and model parameter</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32058",
      "postDate": "10/06/2013 02:09:42",
      "content": "<p>i am very curious about baseline randomforest&nbsp; method .</p>\n<p>we know there are some id feature: such as srch_id ,prop_id</p>\n<p>why are ids useful?</p>\n<p>I don't think that id values can be compared?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32060",
      "postDate": "10/06/2013 03:25:38",
      "content": "<p>Curious to ask how many examples people are running on? <br><br>Unfortunately the machine I run it on only has 4GB in it so I'm using only about 300k examples. Upgrading hardware in the next week or two...</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32061",
      "postDate": "10/06/2013 08:26:45",
      "content": "<p>[quote=Alec Radford;32060]</p>\n<p>Curious to ask how many examples people are running on? <br><br>Unfortunately the machine I run it on only has 4GB in it so I'm using only about 300k examples. Upgrading hardware in the next week or two...</p>\n<p>[/quote]</p>\n<p>At least for now, I'm using about 10% of the data set due to memory restrictions.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32330",
      "postDate": "10/15/2013 01:20:16",
      "content": "<p>you just use 10% train-datasets to reach rank-15. Your method is so beautiful</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32442",
      "postDate": "10/19/2013 16:10:08",
      "content": "<p>Anyone using an algorithm specifically designed for ranking? I know there are a lot available, but I don't know much about them. &nbsp;I've just used GBM and RF (best results have come from regression, not classification) with some feature engineering. &nbsp;I wonder if I could get more out of improving features or changing algorithms. &nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32447",
      "postDate": "10/20/2013 00:15:24",
      "content": "<p>As always, if we think RF or GBM are &quot;single&quot; models then the answer is an obvious YES.</p>\n<p>The data seem to be large enough so that even 10% can produce very decent models.</p>\n<p>&nbsp;</p>\n<p>I think a single (in the sense of a single &quot;GBM&quot;) model built with 10% data can get into top 10 on the leaderboard right now (0.525xx).</p>\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "32450",
      "postDate": "10/20/2013 01:54:19",
      "content": "<p>owen &#65292;so better result.</p>\n<p>I think&nbsp; you get good result in feature engineer&#12290;</p>\n<p>I use learning to rank method&#65292;but now below 0.5</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 32029,
      "author_name": "jeongyoonlee",
      "author_url": "",
      "post_date": "10/04/2013 16:05:43",
      "content": "<p>Yes. :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32031,
      "author_name": "alecradford",
      "author_url": "",
      "post_date": "10/04/2013 17:13:20",
      "content": "<p>Almost there!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32039,
      "author_name": "xavierconort",
      "author_url": "",
      "post_date": "10/04/2013 23:49:59",
      "content": "<p>I use 2 models. Each one can beat 0.5.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32046,
      "author_name": "dylanfan1",
      "author_url": "",
      "post_date": "10/05/2013 07:37:50",
      "content": "<p>good job</p>\n<p>I use LambdaMART now , but hard to promote 0.46.</p>\n<p>are you generate other features which is not include in datasets?</p>\n<p>add cross-features?</p>\n<p>&nbsp;</p>\n<p>thanks!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32047,
      "author_name": "dylanfan1",
      "author_url": "",
      "post_date": "10/05/2013 07:46:19",
      "content": "<p>can give me some advice?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32048,
      "author_name": "xavierconort",
      "author_url": "",
      "post_date": "10/05/2013 08:31:26",
      "content": "<p>Indeed, you need to do a bit of feature engineering first. R with a lot of RAM - more than 16 Gb - can do the job. Sorry I can't tell you more at this stage!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32054,
      "author_name": "songgc",
      "author_url": "",
      "post_date": "10/05/2013 18:26:50",
      "content": "<p>With a laptop of 8GB RAM, you still can beat 0.5. Of course, it would require a bit more programming experience. &nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32057,
      "author_name": "dylanfan1",
      "author_url": "",
      "post_date": "10/06/2013 02:06:46",
      "content": "<p>I think my problem now : do not well in feature scale and model parameter</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32058,
      "author_name": "dylanfan1",
      "author_url": "",
      "post_date": "10/06/2013 02:09:42",
      "content": "<p>i am very curious about baseline randomforest&nbsp; method .</p>\n<p>we know there are some id feature: such as srch_id ,prop_id</p>\n<p>why are ids useful?</p>\n<p>I don't think that id values can be compared?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32060,
      "author_name": "alecradford",
      "author_url": "",
      "post_date": "10/06/2013 03:25:38",
      "content": "<p>Curious to ask how many examples people are running on? <br><br>Unfortunately the machine I run it on only has 4GB in it so I'm using only about 300k examples. Upgrading hardware in the next week or two...</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32061,
      "author_name": "dynamic24",
      "author_url": "",
      "post_date": "10/06/2013 08:26:45",
      "content": "<p>[quote=Alec Radford;32060]</p>\n<p>Curious to ask how many examples people are running on? <br><br>Unfortunately the machine I run it on only has 4GB in it so I'm using only about 300k examples. Upgrading hardware in the next week or two...</p>\n<p>[/quote]</p>\n<p>At least for now, I'm using about 10% of the data set due to memory restrictions.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32330,
      "author_name": "dylanfan1",
      "author_url": "",
      "post_date": "10/15/2013 01:20:16",
      "content": "<p>you just use 10% train-datasets to reach rank-15. Your method is so beautiful</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32442,
      "author_name": "bensolucky",
      "author_url": "",
      "post_date": "10/19/2013 16:10:08",
      "content": "<p>Anyone using an algorithm specifically designed for ranking? I know there are a lot available, but I don't know much about them. &nbsp;I've just used GBM and RF (best results have come from regression, not classification) with some feature engineering. &nbsp;I wonder if I could get more out of improving features or changing algorithms. &nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32447,
      "author_name": "owenzhang1",
      "author_url": "",
      "post_date": "10/20/2013 00:15:24",
      "content": "<p>As always, if we think RF or GBM are &quot;single&quot; models then the answer is an obvious YES.</p>\n<p>The data seem to be large enough so that even 10% can produce very decent models.</p>\n<p>&nbsp;</p>\n<p>I think a single (in the sense of a single &quot;GBM&quot;) model built with 10% data can get into top 10 on the leaderboard right now (0.525xx).</p>\n<p>&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 32450,
      "author_name": "dylanfan1",
      "author_url": "",
      "post_date": "10/20/2013 01:54:19",
      "content": "<p>owen &#65292;so better result.</p>\n<p>I think&nbsp; you get good result in feature engineer&#12290;</p>\n<p>I use learning to rank method&#65292;but now below 0.5</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "32021": "",
    "32029": "",
    "32031": "",
    "32039": "",
    "32046": "",
    "32047": "",
    "32048": "",
    "32054": "",
    "32057": "",
    "32058": "",
    "32060": "",
    "32061": "",
    "32330": "",
    "32442": "",
    "32447": "",
    "32450": ""
  },
  "source": "meta"
}