{
  "id": 21583,
  "title": "Waiting for Idle_speculation solution",
  "url": "/competitions/expedia-hotel-recommendations/discussion/21583",
  "author_name": "",
  "post_date": "2016-06-11T00:01:32.037Z",
  "votes": 4,
  "comment_count": 13,
  "views": 2889,
  "content": "",
  "messages": [
    {
      "id": "123306",
      "postDate": "06/11/2016 00:01:32",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "123307",
      "postDate": "06/11/2016 00:01:50",
      "content": "<p>reserved :)</p>",
      "rawMarkdown": "reserved :)",
      "votes": null
    },
    {
      "id": "123310",
      "postDate": "06/11/2016 00:11:42",
      "content": "<p>Me too :)</p>",
      "rawMarkdown": "Me too :)",
      "votes": null
    },
    {
      "id": "123312",
      "postDate": "06/11/2016 00:40:18",
      "content": "<p>I'll write up something more detailed tomorrow.  Basically:</p>\n\n<ol>\n<li>map user cities and clusters to latitude and longitude using gradient descent</li>\n<li>build a factorization machine model for each cluster</li>\n<li>calculate historical click and book rates by a variety of factors</li>\n<li>build a modified &quot;rank:pairwise&quot; xgboost model on 1-3</li>\n</ol>\n\n<p>Now if you'll excuse me, I have some celebrating I need to take care of :)</p>",
      "rawMarkdown": "I'll write up something more detailed tomorrow.  Basically:\r\n\r\n 1. map user cities and clusters to latitude and longitude using gradient descent\r\n 2. build a factorization machine model for each cluster\r\n 3. calculate historical click and book rates by a variety of factors\r\n 4. build a modified \"rank:pairwise\" xgboost model on 1-3\r\n\r\nNow if you'll excuse me, I have some celebrating I need to take care of :)",
      "votes": null
    },
    {
      "id": "123314",
      "postDate": "06/11/2016 00:44:31",
      "content": "<p>That's real machine learning.. (2) Do you mean factorization on one hot encoded hotel_cluster column. And did you used LibFm</p>",
      "rawMarkdown": "That's real machine learning.. (2) Do you mean factorization on one hot encoded hotel_cluster column. And did you used LibFm",
      "votes": null
    },
    {
      "id": "123316",
      "postDate": "06/11/2016 01:07:57",
      "content": "<p>Congrats Idle. That's an outstanding and amazing solution. Also congrats for your CV discipline. </p>",
      "rawMarkdown": "Congrats Idle. That's an outstanding and amazing solution. Also congrats for your CV discipline.",
      "votes": null
    },
    {
      "id": "123318",
      "postDate": "06/11/2016 01:36:03",
      "content": "<p>Congrats to all winners and waiting to see Idle's amazing solution:)</p>",
      "rawMarkdown": "Congrats to all winners and waiting to see Idle's amazing solution:)",
      "votes": null
    },
    {
      "id": "123321",
      "postDate": "06/11/2016 01:51:47",
      "content": "<p>Genius</p>",
      "rawMarkdown": "Genius",
      "votes": null
    },
    {
      "id": "123329",
      "postDate": "06/11/2016 04:20:09",
      "content": "<p>Congrats on another landslide 1st place. Pleasure meeting you at Atlantic City. Looking forward to hearing about the details. You seem to have this xg &quot;pairwise&quot; thing down.</p>",
      "rawMarkdown": "Congrats on another landslide 1st place. Pleasure meeting you at Atlantic City. Looking forward to hearing about the details. You seem to have this xg \"pairwise\" thing down.",
      "votes": null
    },
    {
      "id": "123348",
      "postDate": "06/11/2016 08:07:26",
      "content": "<p>congrat to all the winners and the amazing Idle_speculation. looking farward to learning from you.</p>",
      "rawMarkdown": "congrat to all the winners and the amazing Idle_speculation. looking farward to learning from you.",
      "votes": null
    },
    {
      "id": "123350",
      "postDate": "06/11/2016 08:27:40",
      "content": "<p>Congratulations to all winners!  </p>\n\n<p>I would love to learn @idle_speculation techniques of organizing CV that allows avoiding LB submissions completely . It was 1 shot = 1 kill in this competition. Most impressive.</p>",
      "rawMarkdown": "Congratulations to all winners!  \r\n\r\nI would love to learn @idle_speculation techniques of organizing CV that allows avoiding LB submissions completely . It was 1 shot = 1 kill in this competition. Most impressive.",
      "votes": null
    },
    {
      "id": "123359",
      "postDate": "06/11/2016 09:45:43",
      "content": "<p>Congratulations to the top teams! It was close race for the 2nd.\nOutstanding performance idle_speculation!\nI am currently on vacation I will give details on Wedneday.</p>\n\n<p>Quick overview:</p>\n\n<ul>\n<li><p>Map city to globe and calculate lat-long</p></li>\n<li><p>create seasonality proxy for destinations</p></li>\n<li><p>Hotel cluster frequencies based on factors</p></li>\n<li><p>user preferences</p></li>\n</ul>\n\n<p>Split data based on leakage hotel matches 1:2\nTrained binary xgb models separately for each hotel clusters.\nI used 8-20% of the negative samples in each binary classifiers to speed up training.\nseparate feature selection and paropt helped.</p>",
      "rawMarkdown": "Congratulations to the top teams! It was close race for the 2nd.\r\nOutstanding performance idle_speculation!\r\nI am currently on vacation I will give details on Wedneday.\r\n\r\nQuick overview:\r\n\r\n* Map city to globe and calculate lat-long\r\n\r\n* create seasonality proxy for destinations\r\n\r\n* Hotel cluster frequencies based on factors\r\n\r\n* user preferences\r\n\r\nSplit data based on leakage hotel matches 1:2\r\nTrained binary xgb models separately for each hotel clusters.\r\nI used 8-20% of the negative samples in each binary classifiers to speed up training.\r\nseparate feature selection and paropt helped.",
      "votes": null
    },
    {
      "id": "123412",
      "postDate": "06/11/2016 18:39:00",
      "content": "<p>How did you compute latitude and longitude?</p>",
      "rawMarkdown": "How did you compute latitude and longitude?",
      "votes": null
    },
    {
      "id": "140327",
      "postDate": "10/19/2016 16:56:08",
      "content": "<p>@beluga could you share your code of mapping city with lat-lon?</p>",
      "rawMarkdown": "beluga could you share your code of mapping city with lat-lon?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 123307,
      "author_name": "davutpolat",
      "author_url": "",
      "post_date": "06/11/2016 00:01:50",
      "content": "<p>reserved :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123310,
      "author_name": "karimouda",
      "author_url": "",
      "post_date": "06/11/2016 00:11:42",
      "content": "<p>Me too :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123312,
      "author_name": "speculation",
      "author_url": "",
      "post_date": "06/11/2016 00:40:18",
      "content": "<p>I'll write up something more detailed tomorrow.  Basically:</p>\n\n<ol>\n<li>map user cities and clusters to latitude and longitude using gradient descent</li>\n<li>build a factorization machine model for each cluster</li>\n<li>calculate historical click and book rates by a variety of factors</li>\n<li>build a modified &quot;rank:pairwise&quot; xgboost model on 1-3</li>\n</ol>\n\n<p>Now if you'll excuse me, I have some celebrating I need to take care of :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123314,
      "author_name": "shahnawazakhtar",
      "author_url": "",
      "post_date": "06/11/2016 00:44:31",
      "content": "<p>That's real machine learning.. (2) Do you mean factorization on one hot encoded hotel_cluster column. And did you used LibFm</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123316,
      "author_name": "titericz",
      "author_url": "",
      "post_date": "06/11/2016 01:07:57",
      "content": "<p>Congrats Idle. That's an outstanding and amazing solution. Also congrats for your CV discipline. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123318,
      "author_name": "kuanchen",
      "author_url": "",
      "post_date": "06/11/2016 01:36:03",
      "content": "<p>Congrats to all winners and waiting to see Idle's amazing solution:)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123321,
      "author_name": "beedata",
      "author_url": "",
      "post_date": "06/11/2016 01:51:47",
      "content": "<p>Genius</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123329,
      "author_name": "mikeskim",
      "author_url": "",
      "post_date": "06/11/2016 04:20:09",
      "content": "<p>Congrats on another landslide 1st place. Pleasure meeting you at Atlantic City. Looking forward to hearing about the details. You seem to have this xg &quot;pairwise&quot; thing down.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123348,
      "author_name": "pichai",
      "author_url": "",
      "post_date": "06/11/2016 08:07:26",
      "content": "<p>congrat to all the winners and the amazing Idle_speculation. looking farward to learning from you.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123350,
      "author_name": "vykhand",
      "author_url": "",
      "post_date": "06/11/2016 08:27:40",
      "content": "<p>Congratulations to all winners!  </p>\n\n<p>I would love to learn @idle_speculation techniques of organizing CV that allows avoiding LB submissions completely . It was 1 shot = 1 kill in this competition. Most impressive.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123359,
      "author_name": "gaborfodor",
      "author_url": "",
      "post_date": "06/11/2016 09:45:43",
      "content": "<p>Congratulations to the top teams! It was close race for the 2nd.\nOutstanding performance idle_speculation!\nI am currently on vacation I will give details on Wedneday.</p>\n\n<p>Quick overview:</p>\n\n<ul>\n<li><p>Map city to globe and calculate lat-long</p></li>\n<li><p>create seasonality proxy for destinations</p></li>\n<li><p>Hotel cluster frequencies based on factors</p></li>\n<li><p>user preferences</p></li>\n</ul>\n\n<p>Split data based on leakage hotel matches 1:2\nTrained binary xgb models separately for each hotel clusters.\nI used 8-20% of the negative samples in each binary classifiers to speed up training.\nseparate feature selection and paropt helped.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 123412,
      "author_name": "thomasseleck",
      "author_url": "",
      "post_date": "06/11/2016 18:39:00",
      "content": "<p>How did you compute latitude and longitude?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 140327,
      "author_name": "danielchanthink",
      "author_url": "",
      "post_date": "10/19/2016 16:56:08",
      "content": "<p>@beluga could you share your code of mapping city with lat-lon?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "123306": "",
    "123307": "reserved :)",
    "123310": "Me too :)",
    "123312": "I'll write up something more detailed tomorrow.  Basically:\r\n\r\n 1. map user cities and clusters to latitude and longitude using gradient descent\r\n 2. build a factorization machine model for each cluster\r\n 3. calculate historical click and book rates by a variety of factors\r\n 4. build a modified \"rank:pairwise\" xgboost model on 1-3\r\n\r\nNow if you'll excuse me, I have some celebrating I need to take care of :)",
    "123314": "That's real machine learning.. (2) Do you mean factorization on one hot encoded hotel_cluster column. And did you used LibFm",
    "123316": "Congrats Idle. That's an outstanding and amazing solution. Also congrats for your CV discipline.",
    "123318": "Congrats to all winners and waiting to see Idle's amazing solution:)",
    "123321": "Genius",
    "123329": "Congrats on another landslide 1st place. Pleasure meeting you at Atlantic City. Looking forward to hearing about the details. You seem to have this xg \"pairwise\" thing down.",
    "123348": "congrat to all the winners and the amazing Idle_speculation. looking farward to learning from you.",
    "123350": "Congratulations to all winners!  \r\n\r\nI would love to learn @idle_speculation techniques of organizing CV that allows avoiding LB submissions completely . It was 1 shot = 1 kill in this competition. Most impressive.",
    "123359": "Congratulations to the top teams! It was close race for the 2nd.\r\nOutstanding performance idle_speculation!\r\nI am currently on vacation I will give details on Wedneday.\r\n\r\nQuick overview:\r\n\r\n* Map city to globe and calculate lat-long\r\n\r\n* create seasonality proxy for destinations\r\n\r\n* Hotel cluster frequencies based on factors\r\n\r\n* user preferences\r\n\r\nSplit data based on leakage hotel matches 1:2\r\nTrained binary xgb models separately for each hotel clusters.\r\nI used 8-20% of the negative samples in each binary classifiers to speed up training.\r\nseparate feature selection and paropt helped.",
    "123412": "How did you compute latitude and longitude?",
    "140327": "beluga could you share your code of mapping city with lat-lon?"
  },
  "source": "meta"
}