{
  "id": 12464,
  "title": "Do the top winners use the anonymity feathers?",
  "url": "/competitions/avazu-ctr-prediction/discussion/12464",
  "author_name": "",
  "post_date": "2015-02-10T06:03:36.933Z",
  "votes": null,
  "comment_count": 6,
  "views": 2909,
  "content": "<p>First,congrats to winners,and this is a fun competition.</p>\n<p>I use named feathers to create new feathers such as clicks per person per app_id/site_id and so on,the first/last time one person click the app_id/site_id/app_cate/site_cate ,totally create about 130 feathers,and it does not work ,only get 0.42 on the LB with the anonymity feathers.</p>\n<p>And I wonder how to use anonymity feather? Something wrong with my feather&nbsp;engineering?</p>\n<p>Would you share your feather engineering method?</p>",
  "messages": [
    {
      "id": "63880",
      "postDate": "02/10/2015 06:03:36",
      "content": "<p>First,congrats to winners,and this is a fun competition.</p>\n<p>I use named feathers to create new feathers such as clicks per person per app_id/site_id and so on,the first/last time one person click the app_id/site_id/app_cate/site_cate ,totally create about 130 feathers,and it does not work ,only get 0.42 on the LB with the anonymity feathers.</p>\n<p>And I wonder how to use anonymity feather? Something wrong with my feather&nbsp;engineering?</p>\n<p>Would you share your feather engineering method?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63894",
      "postDate": "02/10/2015 08:17:59",
      "content": "<p>Sorry for the dumb question, but what are &quot;feathers&quot; in machine learning context? Or do you simply mean &quot;features&quot; (as in predictors, inputs, regressors, covariates, independent variables)?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63896",
      "postDate": "02/10/2015 08:39:04",
      "content": "<p>I've used genetic algo to select interacted feature pairs, but this only get me to LB 0.391. Code for genetic feature&nbsp;selection shared <a href=\"https://github.com/xxinl/Kaggle.Avazu.CTR/blob/master/Cpp/feature_select_genetic.h\" target=\"_blank\">here</a>&nbsp;(c++).</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63898",
      "postDate": "02/10/2015 08:50:20",
      "content": "<p>I've used parallel version of SGD (runs multiple models in the same time) to discover interactions of interest. The interaction had to improve log loss on every day. So, for every interaction, 10 models are trained, every one being tested on different day (like leave one out CV). It was running for 20 hour. Here are the results.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63943",
      "postDate": "02/10/2015 17:23:46",
      "content": "<p>I wonder how results from Interactions.xlsx looks like, when plotted as heatmap, App_id seems to have some interesting capability for making valid interactions with other variables.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63948",
      "postDate": "02/10/2015 18:22:13",
      "content": "<p>[quote=xvnv;63943]</p>\n<p>I wonder how results from Interactions.xlsx looks like, when plotted as heatmap, App_id seems to have some interesting capability for making valid interactions with other variables.</p>\n<p>[/quote]</p>\n<p>Don't have it, sorry. You can sort it by name to see where app_id fits. All with score 10&nbsp; or 4 for days 27..30 should improve your score.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "73126",
      "postDate": "04/22/2015 08:56:56",
      "content": "<p>[quote=hey_world1293;63880]</p>\n<p>First,congrats to winners,and this is a fun competition.</p>\n<p>I use named feathers to create new feathers such as clicks per person per app_id/site_id and so on,the first/last time one person click the app_id/site_id/app_cate/site_cate ,totally create about 130 feathers,and it does not work ,only get 0.42 on the LB with the anonymity feathers.</p>\n<p>And I wonder how to use anonymity feather? Something wrong with my feather&nbsp;engineering?</p>\n<p>Would you share your feather engineering method?</p>\n<p>[/quote]</p>\n<p>the dataset doesn't have person feature, where the clicks per person come from?</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 63894,
      "author_name": "tobycheese",
      "author_url": "",
      "post_date": "02/10/2015 08:17:59",
      "content": "<p>Sorry for the dumb question, but what are &quot;feathers&quot; in machine learning context? Or do you simply mean &quot;features&quot; (as in predictors, inputs, regressors, covariates, independent variables)?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63896,
      "author_name": "xinliu",
      "author_url": "",
      "post_date": "02/10/2015 08:39:04",
      "content": "<p>I've used genetic algo to select interacted feature pairs, but this only get me to LB 0.391. Code for genetic feature&nbsp;selection shared <a href=\"https://github.com/xxinl/Kaggle.Avazu.CTR/blob/master/Cpp/feature_select_genetic.h\" target=\"_blank\">here</a>&nbsp;(c++).</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63898,
      "author_name": "gmilosev",
      "author_url": "",
      "post_date": "02/10/2015 08:50:20",
      "content": "<p>I've used parallel version of SGD (runs multiple models in the same time) to discover interactions of interest. The interaction had to improve log loss on every day. So, for every interaction, 10 models are trained, every one being tested on different day (like leave one out CV). It was running for 20 hour. Here are the results.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63943,
      "author_name": "oneoff",
      "author_url": "",
      "post_date": "02/10/2015 17:23:46",
      "content": "<p>I wonder how results from Interactions.xlsx looks like, when plotted as heatmap, App_id seems to have some interesting capability for making valid interactions with other variables.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63948,
      "author_name": "gmilosev",
      "author_url": "",
      "post_date": "02/10/2015 18:22:13",
      "content": "<p>[quote=xvnv;63943]</p>\n<p>I wonder how results from Interactions.xlsx looks like, when plotted as heatmap, App_id seems to have some interesting capability for making valid interactions with other variables.</p>\n<p>[/quote]</p>\n<p>Don't have it, sorry. You can sort it by name to see where app_id fits. All with score 10&nbsp; or 4 for days 27..30 should improve your score.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 73126,
      "author_name": "kaka10",
      "author_url": "",
      "post_date": "04/22/2015 08:56:56",
      "content": "<p>[quote=hey_world1293;63880]</p>\n<p>First,congrats to winners,and this is a fun competition.</p>\n<p>I use named feathers to create new feathers such as clicks per person per app_id/site_id and so on,the first/last time one person click the app_id/site_id/app_cate/site_cate ,totally create about 130 feathers,and it does not work ,only get 0.42 on the LB with the anonymity feathers.</p>\n<p>And I wonder how to use anonymity feather? Something wrong with my feather&nbsp;engineering?</p>\n<p>Would you share your feather engineering method?</p>\n<p>[/quote]</p>\n<p>the dataset doesn't have person feature, where the clicks per person come from?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "63880": "",
    "63894": "",
    "63896": "",
    "63898": "",
    "63943": "",
    "63948": "",
    "73126": ""
  },
  "source": "meta"
}