{
  "id": 26160,
  "title": "Beat the BTB's with vowpal wabbit! ~0.641",
  "url": "/competitions/outbrain-click-prediction/discussion/26160",
  "author_name": "",
  "post_date": "2016-12-05T04:24:13.267Z",
  "votes": 10,
  "comment_count": 4,
  "views": 752,
  "content": "<p>Hi all, </p>\n\n<p>The following code was inspired in the FTRL script, the idea is to use the preprocessing to use vowpal wabbit that allows a few more options than the python script. The \"leak variables\" were left out, but can be included by modifying the scripts.</p>\n\n<p>Please feel free to add improvements/suggestions to the code or report bugs if you find them. Probably github is the best for this. </p>\n\n<p>Repo with the code:\n<a href=\"https://github.com/martinbel/kaggle-outbrain-click-prediction\">https://github.com/martinbel/kaggle-outbrain-click-prediction</a></p>\n\n<p>General <a href=\"https://martinbel.github.io/outbrain-click-prediction-with-vowpal-wabbit.html\">instructions</a> to reproduce the results.</p>",
  "messages": [
    {
      "id": "148460",
      "postDate": "12/05/2016 04:24:13",
      "content": "<p>Hi all, </p>\n\n<p>The following code was inspired in the FTRL script, the idea is to use the preprocessing to use vowpal wabbit that allows a few more options than the python script. The \"leak variables\" were left out, but can be included by modifying the scripts.</p>\n\n<p>Please feel free to add improvements/suggestions to the code or report bugs if you find them. Probably github is the best for this. </p>\n\n<p>Repo with the code:\n<a href=\"https://github.com/martinbel/kaggle-outbrain-click-prediction\">https://github.com/martinbel/kaggle-outbrain-click-prediction</a></p>\n\n<p>General <a href=\"https://martinbel.github.io/outbrain-click-prediction-with-vowpal-wabbit.html\">instructions</a> to reproduce the results.</p>",
      "rawMarkdown": "Hi all, \r\n\r\nThe following code was inspired in the FTRL script, the idea is to use the preprocessing to use vowpal wabbit that allows a few more options than the python script. The \"leak variables\" were left out, but can be included by modifying the scripts.\r\n\r\nPlease feel free to add improvements/suggestions to the code or report bugs if you find them. Probably github is the best for this. \r\n\r\nRepo with the code:\r\nhttps://github.com/martinbel/kaggle-outbrain-click-prediction\r\n\r\nGeneral [instructions][1] to reproduce the results.\r\n\r\n\r\n  [1]: https://martinbel.github.io/outbrain-click-prediction-with-vowpal-wabbit.html",
      "votes": null
    },
    {
      "id": "148561",
      "postDate": "12/05/2016 17:08:12",
      "content": "<p>Hi Marbel,</p>\n\n<p>Thanks for posting. I was looking into using interactions yesterday and this is a perfect and clearly explained introduction for VW.</p>",
      "rawMarkdown": "Hi Marbel,\r\n\r\nThanks for posting. I was looking into using interactions yesterday and this is a perfect and clearly explained introduction for VW.",
      "votes": null
    },
    {
      "id": "148573",
      "postDate": "12/05/2016 17:59:46",
      "content": "<p>@onepau Thanks! It's based mostly on <a href=\"https://github.com/MLWave/kaggle-criteo\">this</a> and the <a href=\"https://www.kaggle.com/sudalairajkumar/outbrain-click-prediction/ftrl-starter-with-leakage-vars\">ftrl scrips</a> used for preprocessing.</p>",
      "rawMarkdown": "onepau Thanks! It's based mostly on [this](https://github.com/MLWave/kaggle-criteo) and the [ftrl scrips](https://www.kaggle.com/sudalairajkumar/outbrain-click-prediction/ftrl-starter-with-leakage-vars) used for preprocessing.",
      "votes": null
    },
    {
      "id": "148587",
      "postDate": "12/05/2016 19:19:08",
      "content": "<p>@marbel Thanks for sharing. Did you generate feature based on (display_id, ad_id) pair?  </p>",
      "rawMarkdown": "marbel Thanks for sharing. Did you generate feature based on (display_id, ad_id) pair?",
      "votes": null
    },
    {
      "id": "148606",
      "postDate": "12/05/2016 20:35:52",
      "content": "<p>@FengLi You can't use the display_id directly, as they don't overlap in the train/test set. But you can make features that describe each display_id. I've used ad_id directly and made features from it. \nIf you uncomment the line:  <code>if t &gt;= 1:break print t, row</code>\nyou will find it easier to understand what the csv2vw.py script is doing. It's all there. </p>",
      "rawMarkdown": "FengLi You can't use the display_id directly, as they don't overlap in the train/test set. But you can make features that describe each display_id. I've used ad_id directly and made features from it. \r\nIf you uncomment the line:  `if t >= 1:break print t, row`\r\nyou will find it easier to understand what the csv2vw.py script is doing. It's all there.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 148561,
      "author_name": "onepau",
      "author_url": "",
      "post_date": "12/05/2016 17:08:12",
      "content": "<p>Hi Marbel,</p>\n\n<p>Thanks for posting. I was looking into using interactions yesterday and this is a perfect and clearly explained introduction for VW.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 148573,
      "author_name": "marbel",
      "author_url": "",
      "post_date": "12/05/2016 17:59:46",
      "content": "<p>@onepau Thanks! It's based mostly on <a href=\"https://github.com/MLWave/kaggle-criteo\">this</a> and the <a href=\"https://www.kaggle.com/sudalairajkumar/outbrain-click-prediction/ftrl-starter-with-leakage-vars\">ftrl scrips</a> used for preprocessing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 148587,
      "author_name": "beedata",
      "author_url": "",
      "post_date": "12/05/2016 19:19:08",
      "content": "<p>@marbel Thanks for sharing. Did you generate feature based on (display_id, ad_id) pair?  </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 148606,
      "author_name": "marbel",
      "author_url": "",
      "post_date": "12/05/2016 20:35:52",
      "content": "<p>@FengLi You can't use the display_id directly, as they don't overlap in the train/test set. But you can make features that describe each display_id. I've used ad_id directly and made features from it. \nIf you uncomment the line:  <code>if t &gt;= 1:break print t, row</code>\nyou will find it easier to understand what the csv2vw.py script is doing. It's all there. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "148460": "Hi all, \r\n\r\nThe following code was inspired in the FTRL script, the idea is to use the preprocessing to use vowpal wabbit that allows a few more options than the python script. The \"leak variables\" were left out, but can be included by modifying the scripts.\r\n\r\nPlease feel free to add improvements/suggestions to the code or report bugs if you find them. Probably github is the best for this. \r\n\r\nRepo with the code:\r\nhttps://github.com/martinbel/kaggle-outbrain-click-prediction\r\n\r\nGeneral [instructions][1] to reproduce the results.\r\n\r\n\r\n  [1]: https://martinbel.github.io/outbrain-click-prediction-with-vowpal-wabbit.html",
    "148561": "Hi Marbel,\r\n\r\nThanks for posting. I was looking into using interactions yesterday and this is a perfect and clearly explained introduction for VW.",
    "148573": "onepau Thanks! It's based mostly on [this](https://github.com/MLWave/kaggle-criteo) and the [ftrl scrips](https://www.kaggle.com/sudalairajkumar/outbrain-click-prediction/ftrl-starter-with-leakage-vars) used for preprocessing.",
    "148587": "marbel Thanks for sharing. Did you generate feature based on (display_id, ad_id) pair?",
    "148606": "FengLi You can't use the display_id directly, as they don't overlap in the train/test set. But you can make features that describe each display_id. I've used ad_id directly and made features from it. \r\nIf you uncomment the line:  `if t >= 1:break print t, row`\r\nyou will find it easier to understand what the csv2vw.py script is doing. It's all there."
  },
  "source": "meta"
}