{
  "id": 4517,
  "title": "Frequently Asked Questions",
  "url": "/competitions/cause-effect-pairs/discussion/4517",
  "author_name": "",
  "post_date": "2013-05-06T22:26:29.210Z",
  "votes": null,
  "comment_count": 5,
  "views": 1923,
  "content": "<p>I created a FAQ page to answer some of the most frequently asked questions in some more details:</p>\r\n<p>http://www.causality.inf.ethz.ch/cause-effect.php?page=FAQ</p>\r\n<p>You will also have a chance to ask questions to the experts if you attend our <a title=\"Reading Group\" href=\"https://sites.google.com/a/chalearn.org/readinggroup/\">\r\nweekly teleconference seminar</a> meeting on thursdays at 9am PT. The first one will go over a new\r\n<a title=\"cause-effect pair slides\" href=\"http://clopinet.com/isabelle/Projects/causality/CauseEffectPairs_may1_2013.pptx\">\r\nextended version og my slide deck</a>.</p>",
  "messages": [
    {
      "id": "23959",
      "postDate": "05/06/2013 22:26:29",
      "content": "<p>I created a FAQ page to answer some of the most frequently asked questions in some more details:</p>\r\n<p>http://www.causality.inf.ethz.ch/cause-effect.php?page=FAQ</p>\r\n<p>You will also have a chance to ask questions to the experts if you attend our <a title=\"Reading Group\" href=\"https://sites.google.com/a/chalearn.org/readinggroup/\">\r\nweekly teleconference seminar</a> meeting on thursdays at 9am PT. The first one will go over a new\r\n<a title=\"cause-effect pair slides\" href=\"http://clopinet.com/isabelle/Projects/causality/CauseEffectPairs_may1_2013.pptx\">\r\nextended version og my slide deck</a>.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "25239",
      "postDate": "06/07/2013 02:31:15",
      "content": "<p>This is my first attempt at competition so I have some rather obvious questions to the seasoned participants but not for a first timer.</p>\r\n<p>1) Can I assume the type of data (&nbsp;numerical, binary, category ) have common range? What I mean for example a category for one training vector may have 2 4 8 ... while in another vector it would be 6 7 12 can I assume for example the category range is 1-12\r\n OR is each category of each training set unique.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "25241",
      "postDate": "06/07/2013 09:45:18",
      "content": "<p>Cruncher - this hasn't been a stereotypical competition. It's probably not the best one to have as your first! Go and have a look at the Amazon competition.&nbsp; The data will be changing for this competition so I wouldn't waste your time at the moment. (Look\r\n at the leaderboard -noone's entering anymore) Have a look at one of the knowledge competitions.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "25255",
      "postDate": "06/07/2013 18:25:03",
      "content": "<p>[quote=cruncher;25239]</p>\r\n<p>This is my first attempt at competition so I have some rather obvious questions to the seasoned participants but not for a first timer.</p>\r\n<p>1) Can I assume the type of data (&nbsp;numerical, binary, category ) have common range? What I mean for example a category for one training vector may have 2 4 8 ... while in another vector it would be 6 7 12 can I assume for example the category range is 1-12\r\n OR is each category of each training set unique.</p>\r\n<p>[/quote]</p>\r\n<p>We recently changed the data normalization.</p>\r\n<p>In all cases, binary variables are 0/1 encoded. Categorical variables are encoded by a number between 1 an num_categories.&nbsp;</p>\r\n<p>In the original data, the numerical variables are integers between 0 and max_val, where max_val does not exceed 9999</p>\r\n<p>In the supplementary data (SUP1 and SUP2) the numerical variables are integers that resulted from standardizing variables then multiplying by 10000 then rounding to the nearest integer.</p>\r\n<p>The competition is very active with a growing number of participants who keep making submissions to the leaderboard (114 this past week). If you make submissions on the validation set, be aware that the validation set is normalized like the initially released\r\n training data whereas the final evaluation test data will be normalized like the SUP2 data. This is because the 4 causal categories were not balanced with respect to variable value mean and max and number of unique values, hence it was possible to rely on\r\n these features, unrelated to the problem at hand, to get good classification results. The new normalization takes care of this bias. A validation set that is a subset of the final data will be released when we finish preparing the final data.</p>\r\n<p>What you can do in the mean time is: use cross-validation to self evaluate and make submissions using the original validation data, but using for training data properly normalized (like the supplementary data).</p>\r\n<p></p>\r\n<p></p>\r\n<p></p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "25261",
      "postDate": "06/07/2013 21:26:15",
      "content": "<p>Ok please let me repeat back my understanding of your message.</p>\r\n<p>use sup2 to train since the final money test set will be of type.</p>\r\n<p>the validation set will follow sup2 normalization method and will come later.</p>\r\n<p>leader boaris means nothing until new validation set released</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "25271",
      "postDate": "06/08/2013 02:03:32",
      "content": "<p>You should not necessarily limit yourself to training with sup2 data, you can train on as much data as you want from any source, including data you generate yourself and real data publicly available such as:&nbsp;http://webdav.tuebingen.mpg.de/cause-effect/.&nbsp;</p>\r\n<p>You may want to self-evaluate yourself on sup2 data because the final evaluation data&nbsp;will be normalized similarly and/or use cross-validation on sup2 data.</p>\r\n<p>The lederboard performances are probably not a good indication of what the final standings will be.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 25239,
      "author_name": "cruncher4",
      "author_url": "",
      "post_date": "06/07/2013 02:31:15",
      "content": "<p>This is my first attempt at competition so I have some rather obvious questions to the seasoned participants but not for a first timer.</p>\r\n<p>1) Can I assume the type of data (&nbsp;numerical, binary, category ) have common range? What I mean for example a category for one training vector may have 2 4 8 ... while in another vector it would be 6 7 12 can I assume for example the category range is 1-12\r\n OR is each category of each training set unique.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 25241,
      "author_name": "domcastro",
      "author_url": "",
      "post_date": "06/07/2013 09:45:18",
      "content": "<p>Cruncher - this hasn't been a stereotypical competition. It's probably not the best one to have as your first! Go and have a look at the Amazon competition.&nbsp; The data will be changing for this competition so I wouldn't waste your time at the moment. (Look\r\n at the leaderboard -noone's entering anymore) Have a look at one of the knowledge competitions.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 25255,
      "author_name": "iguyon",
      "author_url": "",
      "post_date": "06/07/2013 18:25:03",
      "content": "<p>[quote=cruncher;25239]</p>\r\n<p>This is my first attempt at competition so I have some rather obvious questions to the seasoned participants but not for a first timer.</p>\r\n<p>1) Can I assume the type of data (&nbsp;numerical, binary, category ) have common range? What I mean for example a category for one training vector may have 2 4 8 ... while in another vector it would be 6 7 12 can I assume for example the category range is 1-12\r\n OR is each category of each training set unique.</p>\r\n<p>[/quote]</p>\r\n<p>We recently changed the data normalization.</p>\r\n<p>In all cases, binary variables are 0/1 encoded. Categorical variables are encoded by a number between 1 an num_categories.&nbsp;</p>\r\n<p>In the original data, the numerical variables are integers between 0 and max_val, where max_val does not exceed 9999</p>\r\n<p>In the supplementary data (SUP1 and SUP2) the numerical variables are integers that resulted from standardizing variables then multiplying by 10000 then rounding to the nearest integer.</p>\r\n<p>The competition is very active with a growing number of participants who keep making submissions to the leaderboard (114 this past week). If you make submissions on the validation set, be aware that the validation set is normalized like the initially released\r\n training data whereas the final evaluation test data will be normalized like the SUP2 data. This is because the 4 causal categories were not balanced with respect to variable value mean and max and number of unique values, hence it was possible to rely on\r\n these features, unrelated to the problem at hand, to get good classification results. The new normalization takes care of this bias. A validation set that is a subset of the final data will be released when we finish preparing the final data.</p>\r\n<p>What you can do in the mean time is: use cross-validation to self evaluate and make submissions using the original validation data, but using for training data properly normalized (like the supplementary data).</p>\r\n<p></p>\r\n<p></p>\r\n<p></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 25261,
      "author_name": "cruncher4",
      "author_url": "",
      "post_date": "06/07/2013 21:26:15",
      "content": "<p>Ok please let me repeat back my understanding of your message.</p>\r\n<p>use sup2 to train since the final money test set will be of type.</p>\r\n<p>the validation set will follow sup2 normalization method and will come later.</p>\r\n<p>leader boaris means nothing until new validation set released</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 25271,
      "author_name": "iguyon",
      "author_url": "",
      "post_date": "06/08/2013 02:03:32",
      "content": "<p>You should not necessarily limit yourself to training with sup2 data, you can train on as much data as you want from any source, including data you generate yourself and real data publicly available such as:&nbsp;http://webdav.tuebingen.mpg.de/cause-effect/.&nbsp;</p>\r\n<p>You may want to self-evaluate yourself on sup2 data because the final evaluation data&nbsp;will be normalized similarly and/or use cross-validation on sup2 data.</p>\r\n<p>The lederboard performances are probably not a good indication of what the final standings will be.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "23959": "",
    "25239": "",
    "25241": "",
    "25255": "",
    "25261": "",
    "25271": ""
  },
  "source": "meta"
}