{
  "id": 15541,
  "title": "What the ~0.041- group knows that the ~0.43- doesn't?",
  "url": "/competitions/avito-context-ad-clicks/discussion/15541",
  "author_name": "",
  "post_date": "2015-07-25T23:37:54.397Z",
  "votes": 3,
  "comment_count": 12,
  "views": 2260,
  "content": "<p>Looking at the LB I think it's obvious that the ~0.041- group knows something about the data that the rest doesn't.</p>\n\n<p>Can't wait to see what it is though I hope I'll find out myself before the competition end. </p>",
  "messages": [
    {
      "id": "86983",
      "postDate": "07/25/2015 23:37:54",
      "content": "<p>Looking at the LB I think it's obvious that the ~0.041- group knows something about the data that the rest doesn't.</p>\n\n<p>Can't wait to see what it is though I hope I'll find out myself before the competition end. </p>",
      "rawMarkdown": "Looking at the LB I think it's obvious that the ~0.041- group knows something about the data that the rest doesn't.\r\n\r\nCan't wait to see what it is though I hope I'll find out myself before the competition end.",
      "votes": null
    },
    {
      "id": "87120",
      "postDate": "07/27/2015 02:25:41",
      "content": "<p>We know black magic and divination!</p>",
      "rawMarkdown": "We know black magic and divination!",
      "votes": null
    },
    {
      "id": "87138",
      "postDate": "07/27/2015 08:28:46",
      "content": "<p>[quote=Leustagos;87120]</p>\n\n<p>We know black magic and divination!</p>\n\n<p>[/quote]</p>\n\n<p>I know it!</p>\n\n<p>Though, for me it didn't help.</p>\n\n<p>I feel it something else.</p>",
      "rawMarkdown": "[quote=Leustagos;87120]\r\n\r\nWe know black magic and divination!\r\n\r\n[/quote]\r\n\r\nI know it!\r\n\r\nThough, for me it didn't help.\r\n\r\nI feel it something else.",
      "votes": null
    },
    {
      "id": "87157",
      "postDate": "07/27/2015 12:47:08",
      "content": "<p>[quote=clustifier;87138]</p>\n\n<p>[quote=Leustagos;87120]</p>\n\n<p>We know black magic and divination!</p>\n\n<p>[/quote]</p>\n\n<p>I know it!</p>\n\n<p>Though, for me it didn't help.</p>\n\n<p>I feel it something else.</p>\n\n<p>[/quote]</p>\n\n<p>Not really. There are good features, but nothing out of the ordinary. We probably just set up a good validation scheme to definie good features quickly by sampling or using linear ftrl. </p>",
      "rawMarkdown": "[quote=clustifier;87138]\r\n\r\n[quote=Leustagos;87120]\r\n\r\nWe know black magic and divination!\r\n\r\n[/quote]\r\n\r\nI know it!\r\n\r\nThough, for me it didn't help.\r\n\r\nI feel it something else.\r\n\r\n[/quote]\r\n\r\nNot really. There are good features, but nothing out of the ordinary. We probably just set up a good validation scheme to definie good features quickly by sampling or using linear ftrl.",
      "votes": null
    },
    {
      "id": "87165",
      "postDate": "07/27/2015 14:45:56",
      "content": "<p>Hey, Leustagos,\nit is against the rules, Black Magic (<a href=\"https://www.kaggle.com/users/23831/black-magic\">https://www.kaggle.com/users/23831/black-magic</a>) is not your team member :) </p>",
      "rawMarkdown": "Hey, Leustagos,\r\nit is against the rules, Black Magic (https://www.kaggle.com/users/23831/black-magic) is not your team member :)",
      "votes": null
    },
    {
      "id": "87194",
      "postDate": "07/27/2015 18:28:56",
      "content": "<p>That explains it.  I only know block multiplication and division.</p>",
      "rawMarkdown": "That explains it.  I only know block multiplication and division.",
      "votes": null
    },
    {
      "id": "87222",
      "postDate": "07/27/2015 22:06:13",
      "content": "<p>[quote=Sergey Yurgenson;87165]</p>\n\n<p>Hey, Leustagos,\nit is against the rules, Black Magic (<a href=\"https://www.kaggle.com/users/23831/black-magic\">https://www.kaggle.com/users/23831/black-magic</a>) is not your team member :) </p>\n\n<p>[/quote]</p>\n\n<p>@SY, thanks for the heads up. We will stick to sorcery and wizardry from now on... :P</p>",
      "rawMarkdown": "[quote=Sergey Yurgenson;87165]\r\n\r\nHey, Leustagos,\r\nit is against the rules, Black Magic (https://www.kaggle.com/users/23831/black-magic) is not your team member :) \r\n\r\n[/quote]\r\n\r\n @SY, thanks for the heads up. We will stick to sorcery and wizardry from now on... :P",
      "votes": null
    },
    {
      "id": "87315",
      "postDate": "07/28/2015 15:54:52",
      "content": "<p>Can't wait to find out which features I didn't use or didn't use correctly.</p>",
      "rawMarkdown": "Can't wait to find out which features I didn't use or didn't use correctly.",
      "votes": null
    },
    {
      "id": "87318",
      "postDate": "07/28/2015 16:33:48",
      "content": "<p>And at the same time, people like me in the ~0.45 group are wondering what features the 0.43 group are using. I guess it lies in PhoneRequest or VisitStream. For me, there is no personalization at all, just one-model-for-all.</p>\n\n<p>Can't wait to find out which features I didn't use or didn't use correctly.</p>",
      "rawMarkdown": "And at the same time, people like me in the ~0.45 group are wondering what features the 0.43 group are using. I guess it lies in PhoneRequest or VisitStream. For me, there is no personalization at all, just one-model-for-all.\r\n\r\nCan't wait to find out which features I didn't use or didn't use correctly.",
      "votes": null
    },
    {
      "id": "87321",
      "postDate": "07/28/2015 17:03:31",
      "content": "<p>It seems like FTRL was used for feature selection, while something more heavy-weight, like LR, for actual fit.</p>",
      "rawMarkdown": "It seems like FTRL was used for feature selection, while something more heavy-weight, like LR, for actual fit.",
      "votes": null
    },
    {
      "id": "87323",
      "postDate": "07/28/2015 17:34:33",
      "content": "<p>@clustifier has a point :)</p>\n\n<p><img src=\"http://oi59.tinypic.com/m7vqzd.jpg\" alt=\"Current LB scores\" title></p>",
      "rawMarkdown": "clustifier has a point :)\r\n\r\n![Current LB scores][1]\r\n\r\n\r\n  [1]: http://oi59.tinypic.com/m7vqzd.jpg",
      "votes": null
    },
    {
      "id": "87327",
      "postDate": "07/28/2015 18:07:48",
      "content": "<p>The ~0.042 can be achieved without using VisitsStream and PhoneRequest. </p>",
      "rawMarkdown": "The ~0.042 can be achieved without using VisitsStream and PhoneRequest.",
      "votes": null
    },
    {
      "id": "87331",
      "postDate": "07/28/2015 18:55:10",
      "content": "<p>I'm pretty sure my model is so complicated. I created for about 40 features.</p>\n\n<p>The problem is that i can't eliminate bad features because it takes so long to train it. Even with a small subset.</p>",
      "rawMarkdown": "I'm pretty sure my model is so complicated. I created for about 40 features.\r\n\r\nThe problem is that i can't eliminate bad features because it takes so long to train it. Even with a small subset.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 87120,
      "author_name": "leustagos",
      "author_url": "",
      "post_date": "07/27/2015 02:25:41",
      "content": "<p>We know black magic and divination!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87138,
      "author_name": "clustifier",
      "author_url": "",
      "post_date": "07/27/2015 08:28:46",
      "content": "<p>[quote=Leustagos;87120]</p>\n\n<p>We know black magic and divination!</p>\n\n<p>[/quote]</p>\n\n<p>I know it!</p>\n\n<p>Though, for me it didn't help.</p>\n\n<p>I feel it something else.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87157,
      "author_name": "leustagos",
      "author_url": "",
      "post_date": "07/27/2015 12:47:08",
      "content": "<p>[quote=clustifier;87138]</p>\n\n<p>[quote=Leustagos;87120]</p>\n\n<p>We know black magic and divination!</p>\n\n<p>[/quote]</p>\n\n<p>I know it!</p>\n\n<p>Though, for me it didn't help.</p>\n\n<p>I feel it something else.</p>\n\n<p>[/quote]</p>\n\n<p>Not really. There are good features, but nothing out of the ordinary. We probably just set up a good validation scheme to definie good features quickly by sampling or using linear ftrl. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87165,
      "author_name": "ccccat",
      "author_url": "",
      "post_date": "07/27/2015 14:45:56",
      "content": "<p>Hey, Leustagos,\nit is against the rules, Black Magic (<a href=\"https://www.kaggle.com/users/23831/black-magic\">https://www.kaggle.com/users/23831/black-magic</a>) is not your team member :) </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87194,
      "author_name": "willieliao",
      "author_url": "",
      "post_date": "07/27/2015 18:28:56",
      "content": "<p>That explains it.  I only know block multiplication and division.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87222,
      "author_name": "leustagos",
      "author_url": "",
      "post_date": "07/27/2015 22:06:13",
      "content": "<p>[quote=Sergey Yurgenson;87165]</p>\n\n<p>Hey, Leustagos,\nit is against the rules, Black Magic (<a href=\"https://www.kaggle.com/users/23831/black-magic\">https://www.kaggle.com/users/23831/black-magic</a>) is not your team member :) </p>\n\n<p>[/quote]</p>\n\n<p>@SY, thanks for the heads up. We will stick to sorcery and wizardry from now on... :P</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87315,
      "author_name": "clustifier",
      "author_url": "",
      "post_date": "07/28/2015 15:54:52",
      "content": "<p>Can't wait to find out which features I didn't use or didn't use correctly.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87318,
      "author_name": "hitchpy",
      "author_url": "",
      "post_date": "07/28/2015 16:33:48",
      "content": "<p>And at the same time, people like me in the ~0.45 group are wondering what features the 0.43 group are using. I guess it lies in PhoneRequest or VisitStream. For me, there is no personalization at all, just one-model-for-all.</p>\n\n<p>Can't wait to find out which features I didn't use or didn't use correctly.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87321,
      "author_name": "bekbolatov",
      "author_url": "",
      "post_date": "07/28/2015 17:03:31",
      "content": "<p>It seems like FTRL was used for feature selection, while something more heavy-weight, like LR, for actual fit.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87323,
      "author_name": "bekbolatov",
      "author_url": "",
      "post_date": "07/28/2015 17:34:33",
      "content": "<p>@clustifier has a point :)</p>\n\n<p><img src=\"http://oi59.tinypic.com/m7vqzd.jpg\" alt=\"Current LB scores\" title></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87327,
      "author_name": "skirpichenko",
      "author_url": "",
      "post_date": "07/28/2015 18:07:48",
      "content": "<p>The ~0.042 can be achieved without using VisitsStream and PhoneRequest. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 87331,
      "author_name": "rushter",
      "author_url": "",
      "post_date": "07/28/2015 18:55:10",
      "content": "<p>I'm pretty sure my model is so complicated. I created for about 40 features.</p>\n\n<p>The problem is that i can't eliminate bad features because it takes so long to train it. Even with a small subset.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "86983": "Looking at the LB I think it's obvious that the ~0.041- group knows something about the data that the rest doesn't.\r\n\r\nCan't wait to see what it is though I hope I'll find out myself before the competition end.",
    "87120": "We know black magic and divination!",
    "87138": "[quote=Leustagos;87120]\r\n\r\nWe know black magic and divination!\r\n\r\n[/quote]\r\n\r\nI know it!\r\n\r\nThough, for me it didn't help.\r\n\r\nI feel it something else.",
    "87157": "[quote=clustifier;87138]\r\n\r\n[quote=Leustagos;87120]\r\n\r\nWe know black magic and divination!\r\n\r\n[/quote]\r\n\r\nI know it!\r\n\r\nThough, for me it didn't help.\r\n\r\nI feel it something else.\r\n\r\n[/quote]\r\n\r\nNot really. There are good features, but nothing out of the ordinary. We probably just set up a good validation scheme to definie good features quickly by sampling or using linear ftrl.",
    "87165": "Hey, Leustagos,\r\nit is against the rules, Black Magic (https://www.kaggle.com/users/23831/black-magic) is not your team member :)",
    "87194": "That explains it.  I only know block multiplication and division.",
    "87222": "[quote=Sergey Yurgenson;87165]\r\n\r\nHey, Leustagos,\r\nit is against the rules, Black Magic (https://www.kaggle.com/users/23831/black-magic) is not your team member :) \r\n\r\n[/quote]\r\n\r\n @SY, thanks for the heads up. We will stick to sorcery and wizardry from now on... :P",
    "87315": "Can't wait to find out which features I didn't use or didn't use correctly.",
    "87318": "And at the same time, people like me in the ~0.45 group are wondering what features the 0.43 group are using. I guess it lies in PhoneRequest or VisitStream. For me, there is no personalization at all, just one-model-for-all.\r\n\r\nCan't wait to find out which features I didn't use or didn't use correctly.",
    "87321": "It seems like FTRL was used for feature selection, while something more heavy-weight, like LR, for actual fit.",
    "87323": "clustifier has a point :)\r\n\r\n![Current LB scores][1]\r\n\r\n\r\n  [1]: http://oi59.tinypic.com/m7vqzd.jpg",
    "87327": "The ~0.042 can be achieved without using VisitsStream and PhoneRequest.",
    "87331": "I'm pretty sure my model is so complicated. I created for about 40 features.\r\n\r\nThe problem is that i can't eliminate bad features because it takes so long to train it. Even with a small subset."
  },
  "source": "meta"
}