{
  "id": 5272,
  "title": "IGCI works bad on the Training and SUP datasets",
  "url": "/competitions/cause-effect-pairs/discussion/5272",
  "author_name": "",
  "post_date": "2013-08-02T12:05:21.977Z",
  "votes": null,
  "comment_count": 7,
  "views": 2032,
  "content": "<p>Hi, all.</p>\n<p>Recently I tried to use the IGCI algorithm as the features in training the classifier. Below is the result on the training dataset (CEfinal_train).&nbsp;</p>\n<p><img src=\"untitled.png\" alt></p>\n<p>where class 1 denote A-&gt;B, class 3 denote B-&gt;A, and class 2 otherwise. The x-axis denote the output value of</p>\n<p><code>igci(data.a,data.b,2,1)</code></p>\n<p>and y-axis denote</p>\n<p><code>igci(data.a,data.b,2,2)</code></p>\n<p>However, the features are not distinguishable.&nbsp;</p>\n<p>Any suggestions ?</p>",
  "messages": [
    {
      "id": "28047",
      "postDate": "08/02/2013 12:05:21",
      "content": "<p>Hi, all.</p>\n<p>Recently I tried to use the IGCI algorithm as the features in training the classifier. Below is the result on the training dataset (CEfinal_train).&nbsp;</p>\n<p><img src=\"untitled.png\" alt></p>\n<p>where class 1 denote A-&gt;B, class 3 denote B-&gt;A, and class 2 otherwise. The x-axis denote the output value of</p>\n<p><code>igci(data.a,data.b,2,1)</code></p>\n<p>and y-axis denote</p>\n<p><code>igci(data.a,data.b,2,2)</code></p>\n<p>However, the features are not distinguishable.&nbsp;</p>\n<p>Any suggestions ?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28050",
      "postDate": "08/02/2013 13:55:32",
      "content": "<p>It used to work on the previous data set, before renormalization. The same happened for other methods as well. I'm not sure what happened, but it definitely takes away from the credibility of this effort.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28054",
      "postDate": "08/02/2013 14:20:59",
      "content": "<p>IGCI is an algorithm designed for noise free data and the data of the challenge is not noise free. The fact that is worked for the old unnormalized data may be coincidental and is not supported by the theory. The normalized data evens out the marginal distributions of the 4 causal classes, a change that affects IGCI.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28062",
      "postDate": "08/02/2013 15:04:39",
      "content": "<p>Yes, I just went back and everything seems to have changed. I will see if I can work this out :) However there seems to be rarely any code exchange in the forum community.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28063",
      "postDate": "08/02/2013 15:11:32",
      "content": "<p>@isabella&nbsp;Thanks for your reply, and any suggestions in extracting these features out ? I see the code/paper on&nbsp;http://webdav.tuebingen.mpg.de/causality/ , However it seems harder for me to figure all these out.&nbsp;</p>\n<p>BTW, in the BMC paper ( <strong>New methods for separating causes from effects in genomics data</strong> BMC genomics: 2012), it says that using IGCI on noisy data seems to achieve better AUC than other Causal-Effect functions. Maybe I should tweak IGCI for this challenge or devise some new method.</p>\n<p>And also, it (<strong>BMC2012</strong>) proposed a ensemble method.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28066",
      "postDate": "08/02/2013 15:19:13",
      "content": "<p>I know IGCI seems to work well on some noisy data for theoretically unexplained reasons. Slides explaining IGCI are found at:&nbsp;https://sites.google.com/a/chalearn.org/readinggroup/</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28071",
      "postDate": "08/02/2013 16:22:46",
      "content": "<p>@Isabelle Thank you for the clarification. Is it possible to give out more information about the new normalization procedure, granted it doesn't give out too much information. Thanks.</p>\n<p>&nbsp;</p>\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28072",
      "postDate": "08/02/2013 16:32:41",
      "content": "<p>See the post:</p>\n<p>https://www.kaggle.com/c/cause-effect-pairs/forums/t/5055/bias-in-original-data</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 28050,
      "author_name": "ebilal",
      "author_url": "",
      "post_date": "08/02/2013 13:55:32",
      "content": "<p>It used to work on the previous data set, before renormalization. The same happened for other methods as well. I'm not sure what happened, but it definitely takes away from the credibility of this effort.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28054,
      "author_name": "iguyon",
      "author_url": "",
      "post_date": "08/02/2013 14:20:59",
      "content": "<p>IGCI is an algorithm designed for noise free data and the data of the challenge is not noise free. The fact that is worked for the old unnormalized data may be coincidental and is not supported by the theory. The normalized data evens out the marginal distributions of the 4 causal classes, a change that affects IGCI.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28062,
      "author_name": "liubenyuan",
      "author_url": "",
      "post_date": "08/02/2013 15:04:39",
      "content": "<p>Yes, I just went back and everything seems to have changed. I will see if I can work this out :) However there seems to be rarely any code exchange in the forum community.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28063,
      "author_name": "liubenyuan",
      "author_url": "",
      "post_date": "08/02/2013 15:11:32",
      "content": "<p>@isabella&nbsp;Thanks for your reply, and any suggestions in extracting these features out ? I see the code/paper on&nbsp;http://webdav.tuebingen.mpg.de/causality/ , However it seems harder for me to figure all these out.&nbsp;</p>\n<p>BTW, in the BMC paper ( <strong>New methods for separating causes from effects in genomics data</strong> BMC genomics: 2012), it says that using IGCI on noisy data seems to achieve better AUC than other Causal-Effect functions. Maybe I should tweak IGCI for this challenge or devise some new method.</p>\n<p>And also, it (<strong>BMC2012</strong>) proposed a ensemble method.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28066,
      "author_name": "iguyon",
      "author_url": "",
      "post_date": "08/02/2013 15:19:13",
      "content": "<p>I know IGCI seems to work well on some noisy data for theoretically unexplained reasons. Slides explaining IGCI are found at:&nbsp;https://sites.google.com/a/chalearn.org/readinggroup/</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28071,
      "author_name": "ebilal",
      "author_url": "",
      "post_date": "08/02/2013 16:22:46",
      "content": "<p>@Isabelle Thank you for the clarification. Is it possible to give out more information about the new normalization procedure, granted it doesn't give out too much information. Thanks.</p>\n<p>&nbsp;</p>\n<p>&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 28072,
      "author_name": "iguyon",
      "author_url": "",
      "post_date": "08/02/2013 16:32:41",
      "content": "<p>See the post:</p>\n<p>https://www.kaggle.com/c/cause-effect-pairs/forums/t/5055/bias-in-original-data</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "28047": "",
    "28050": "",
    "28054": "",
    "28062": "",
    "28063": "",
    "28066": "",
    "28071": "",
    "28072": ""
  },
  "source": "meta"
}