{
  "id": 16084,
  "title": "2st PLACE - WINNER SOLUTION - Gzs_iceberg",
  "url": "/competitions/avito-context-ad-clicks/discussion/16084",
  "author_name": "Gzs_iceberg",
  "post_date": "2015-08-23T03:49:18.473000",
  "votes": 7,
  "comment_count": 7,
  "views": 3172,
  "content": "<p>Hi, all</p>\n\n<p>First, thanks to the Kaggle team, kagglers and Avito for such great competition.</p>\n\n<p>All codes and documents, please see <a href=\"https://github.com/Gzsiceberg/kaggle-avito\">here</a></p>\n\n<p>Your comments are very welcome.</p>\n\n<p>Best regards</p>\n\n<p>Gzs_iceberg</p>",
  "messages": [
    {
      "id": 90142,
      "postDate": "2015-08-23T03:49:18.473Z",
      "content": "<p>Hi, all</p>\n\n<p>First, thanks to the Kaggle team, kagglers and Avito for such great competition.</p>\n\n<p>All codes and documents, please see <a href=\"https://github.com/Gzsiceberg/kaggle-avito\">here</a></p>\n\n<p>Your comments are very welcome.</p>\n\n<p>Best regards</p>\n\n<p>Gzs_iceberg</p>",
      "rawMarkdown": "Hi, all\r\n\r\nFirst, thanks to the Kaggle team, kagglers and Avito for such great competition.\r\n\r\nAll codes and documents, please see [here][1]\r\n\r\nYour comments are very welcome.\r\n\r\nBest regards\r\n\r\nGzs_iceberg\r\n\r\n\r\n  [1]: https://github.com/Gzsiceberg/kaggle-avito",
      "votes": 7
    },
    {
      "id": 90353,
      "postDate": "2015-08-25T17:39:39.733Z",
      "content": "<p>Thanks Gzs_iceberg.</p>\n\n<p>My solution and documentation can be found here:</p>\n\n<p><a href=\"https://github.com/owenzhang/kaggle-avito\">https://github.com/owenzhang/kaggle-avito</a></p>\n\n<p>Cheers,\nOwen</p>",
      "rawMarkdown": "Thanks Gzs_iceberg.\r\n\r\nMy solution and documentation can be found here:\r\n\r\n https://github.com/owenzhang/kaggle-avito\r\n\r\nCheers,\r\nOwen",
      "votes": 6
    },
    {
      "id": 90633,
      "postDate": "2015-08-28T11:36:13.743Z",
      "content": "<p>@MarDo, I confess -- I have a 256GB local machine. \nWhen I build/tune models, I used 10% sample data.</p>",
      "rawMarkdown": "@MarDo, I confess -- I have a 256GB local machine. \r\nWhen I build/tune models, I used 10% sample data.",
      "votes": 1
    },
    {
      "id": 185152,
      "postDate": "2017-05-24T08:40:20.897Z",
      "content": "<p>Thanks for the solutions!</p>\n\n<p>@Gzs_iceberg:\nI have a problem about the negative down sampling.\nbecause the evaluation is logloss, not the F1 or AUC. I think the negative down sampling can make the ratio of positive and negative become very different between train and test，this can make the logloss in test become larger.</p>",
      "rawMarkdown": "Thanks for the solutions!\n\n@Gzs_iceberg:\nI have a problem about the negative down sampling.\nbecause the evaluation is logloss, not the F1 or AUC. I think the negative down sampling can make the ratio of positive and negative become very different between train and test，this can make the logloss in test become larger."
    },
    {
      "id": 90559,
      "postDate": "2015-08-27T19:57:32.883Z",
      "content": "<p>Thanks for the solutions!</p>\n\n<p>@Owen:  You mention that 256GB RAM are required for the full solution. May I ask how is your working mode? Do you develop locally on the desktop using subsets of the data and then scale to a cloud instance? If yes, what would be your recommendation regarding a cloud instance provider?</p>",
      "rawMarkdown": "Thanks for the solutions!\r\n\r\n@Owen:  You mention that 256GB RAM are required for the full solution. May I ask how is your working mode? Do you develop locally on the desktop using subsets of the data and then scale to a cloud instance? If yes, what would be your recommendation regarding a cloud instance provider?"
    },
    {
      "id": 90147,
      "postDate": "2015-08-23T05:13:51.453Z",
      "content": "<p>I have fixed it. </p>\n\n<p>Sorry about that.</p>",
      "rawMarkdown": "I have fixed it. \r\n\r\nSorry about that."
    },
    {
      "id": 90146,
      "postDate": "2015-08-23T05:10:08.197Z",
      "content": "<p>Hello!</p>\n\n<p>Thanks for the nice solution!</p>\n\n<p>When compiling your code the error encounters: files &quot;hash_filter.hpp&quot; and &quot;random.hpp&quot; were missed (called from /IceLR/src/fm_model/ffm.hpp).</p>\n\n<p>Could you, please, suggest how to solve that issue. </p>",
      "rawMarkdown": "Hello!\r\n\r\nThanks for the nice solution!\r\n\r\nWhen compiling your code the error encounters: files \"hash_filter.hpp\" and \"random.hpp\" were missed (called from /IceLR/src/fm_model/ffm.hpp).\r\n\r\nCould you, please, suggest how to solve that issue. "
    },
    {
      "id": 90695,
      "postDate": "2015-08-28T20:02:05.567Z",
      "rawMarkdown": "",
      "votes": 2,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 90353,
      "author_name": "Owen",
      "author_url": "",
      "post_date": "2015-08-25T17:39:39.733000",
      "content": "<p>Thanks Gzs_iceberg.</p>\n\n<p>My solution and documentation can be found here:</p>\n\n<p><a href=\"https://github.com/owenzhang/kaggle-avito\">https://github.com/owenzhang/kaggle-avito</a></p>\n\n<p>Cheers,\nOwen</p>",
      "votes": 6,
      "replies": []
    },
    {
      "id": 90633,
      "author_name": "Owen",
      "author_url": "",
      "post_date": "2015-08-28T11:36:13.743000",
      "content": "<p>@MarDo, I confess -- I have a 256GB local machine. \nWhen I build/tune models, I used 10% sample data.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 185152,
      "author_name": "小月月",
      "author_url": "",
      "post_date": "2017-05-24T08:40:20.897000",
      "content": "<p>Thanks for the solutions!</p>\n\n<p>@Gzs_iceberg:\nI have a problem about the negative down sampling.\nbecause the evaluation is logloss, not the F1 or AUC. I think the negative down sampling can make the ratio of positive and negative become very different between train and test，this can make the logloss in test become larger.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 90559,
      "author_name": "MarDo",
      "author_url": "",
      "post_date": "2015-08-27T19:57:32.883000",
      "content": "<p>Thanks for the solutions!</p>\n\n<p>@Owen:  You mention that 256GB RAM are required for the full solution. May I ask how is your working mode? Do you develop locally on the desktop using subsets of the data and then scale to a cloud instance? If yes, what would be your recommendation regarding a cloud instance provider?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 90147,
      "author_name": "Gzs_iceberg",
      "author_url": "",
      "post_date": "2015-08-23T05:13:51.453000",
      "content": "<p>I have fixed it. </p>\n\n<p>Sorry about that.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 90146,
      "author_name": "night_bat",
      "author_url": "",
      "post_date": "2015-08-23T05:10:08.197000",
      "content": "<p>Hello!</p>\n\n<p>Thanks for the nice solution!</p>\n\n<p>When compiling your code the error encounters: files &quot;hash_filter.hpp&quot; and &quot;random.hpp&quot; were missed (called from /IceLR/src/fm_model/ffm.hpp).</p>\n\n<p>Could you, please, suggest how to solve that issue. </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 90695,
      "author_name": "",
      "author_url": "",
      "post_date": "2015-08-28T20:02:05.567000",
      "content": "",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "90142": "Hi, all\r\n\r\nFirst, thanks to the Kaggle team, kagglers and Avito for such great competition.\r\n\r\nAll codes and documents, please see [here][1]\r\n\r\nYour comments are very welcome.\r\n\r\nBest regards\r\n\r\nGzs_iceberg\r\n\r\n\r\n  [1]: https://github.com/Gzsiceberg/kaggle-avito",
    "90353": "Thanks Gzs_iceberg.\r\n\r\nMy solution and documentation can be found here:\r\n\r\n https://github.com/owenzhang/kaggle-avito\r\n\r\nCheers,\r\nOwen",
    "90633": "@MarDo, I confess -- I have a 256GB local machine. \r\nWhen I build/tune models, I used 10% sample data.",
    "185152": "Thanks for the solutions!\n\n@Gzs_iceberg:\nI have a problem about the negative down sampling.\nbecause the evaluation is logloss, not the F1 or AUC. I think the negative down sampling can make the ratio of positive and negative become very different between train and test，this can make the logloss in test become larger.",
    "90559": "Thanks for the solutions!\r\n\r\n@Owen:  You mention that 256GB RAM are required for the full solution. May I ask how is your working mode? Do you develop locally on the desktop using subsets of the data and then scale to a cloud instance? If yes, what would be your recommendation regarding a cloud instance provider?",
    "90147": "I have fixed it. \r\n\r\nSorry about that.",
    "90146": "Hello!\r\n\r\nThanks for the nice solution!\r\n\r\nWhen compiling your code the error encounters: files \"hash_filter.hpp\" and \"random.hpp\" were missed (called from /IceLR/src/fm_model/ffm.hpp).\r\n\r\nCould you, please, suggest how to solve that issue. ",
    "90695": ""
  }
}