{
  "id": 14494,
  "title": "Welcome to Avito Context Click Prediction!",
  "url": "/competitions/avito-context-ad-clicks/discussion/14494",
  "author_name": "",
  "post_date": "2015-06-02T21:12:50.843Z",
  "votes": 15,
  "comment_count": 15,
  "views": 4953,
  "content": "<p>Dear participants,</p>\n\n<p>Welcome to Avito Context Click Prediction competition where you are asked to predict probability that a user of an online classified will click on a specific ad. In this competition you are given a lot of behavioral information about each user that&nbsp;can be used to infer his/her interests. We hope that this information combined with other statistics that we have provided can help with better ad targeting.</p>\n<p>You have a unique opportunity to practice your feature preparation and modeling skills with both structured and unstructured data to solve one of the worlds most interesting and challenging problems. In many existing CTR prediction competitions you are either limited with the type of available data or rather you are already given some numerical features that are shrouded in mystery. By taking part in this competition you can learn by yourself full data mining process in programmatic advertising: from understanding raw semi-structured data to creating robust CTR predictive model and testing its accuracy.</p>\n<p>Good Luck!</p>\n\n<p>Ivan</p>",
  "messages": [
    {
      "id": "80703",
      "postDate": "06/02/2015 21:12:50",
      "content": "<p>Dear participants,</p>\n\n<p>Welcome to Avito Context Click Prediction competition where you are asked to predict probability that a user of an online classified will click on a specific ad. In this competition you are given a lot of behavioral information about each user that&nbsp;can be used to infer his/her interests. We hope that this information combined with other statistics that we have provided can help with better ad targeting.</p>\n<p>You have a unique opportunity to practice your feature preparation and modeling skills with both structured and unstructured data to solve one of the worlds most interesting and challenging problems. In many existing CTR prediction competitions you are either limited with the type of available data or rather you are already given some numerical features that are shrouded in mystery. By taking part in this competition you can learn by yourself full data mining process in programmatic advertising: from understanding raw semi-structured data to creating robust CTR predictive model and testing its accuracy.</p>\n<p>Good Luck!</p>\n\n<p>Ivan</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81070",
      "postDate": "06/05/2015 23:44:32",
      "content": "<p>Hi Ivan,</p>\n<p>Hope you are well.</p>\n<p>I am an academic in business analytics (Deakin University, Australia) and found this dataset to be quite related to my research area. I was wondering if I can use this dataset for research purposes.</p>\n<p>Regards,</p>\n<p>Ali Tamaddoni&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81178",
      "postDate": "06/07/2015 20:12:51",
      "content": "<p style=\"text-align: left\">Hi&nbsp;Ali,</p>\n<p style=\"text-align: left\">Feel free to use this dataset but for academic purposes only.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81217",
      "postDate": "06/08/2015 08:41:41",
      "content": "<p>Hi Ivan,</p>\n<p>Thank you very much. Much appreciated.</p>\n<p>Regards,</p>\n<p>Ali Tamaddoni</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81611",
      "postDate": "06/11/2015 15:14:53",
      "content": "<p>Hello,</p>\n<p>I just have a question regarding the attribute 'level' in both the 'location' and 'category' tables.</p>\n<p>What does it signify?</p>\n\n<p>Thank you</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81643",
      "postDate": "06/11/2015 23:05:03",
      "content": "<p>Hello,</p>\n<p><strong>Search</strong></p>\n<p>Level in&nbsp;Location (once you join SearchInfo and Location on SearchLocationID = LocationID) represents geo level&nbsp;where a search (and impression) took place. Values could be 3 (city), 2 (region), or 1 (whole country).</p>\n<p>Level in&nbsp;Category (once you join SearchInfo and Category&nbsp;on SearchCategoryID&nbsp;= CategoryID)&nbsp;represents category&nbsp;level where a search (and impression) took place. Values&nbsp;could be&nbsp;3 (subcategory), 2 (category), or&nbsp;1&nbsp;(no specific category or subcategory were selected by visitor).</p>\n<p><strong>Ads</strong></p>\n<p>Level in Location (once you join AdsInfo and Location using&nbsp;LocationID) represents geo level belongs to&nbsp;a non-contextual ad. It has to&nbsp;be 3 (city). Note that for context ad LocationID&nbsp;is NULL.</p>\n<p>Level in Category (once you join AdsInfo&nbsp;and Category using&nbsp;CategoryID) represents category level belongs to any&nbsp;ad. Values&nbsp;could be either 3 (subcategory, for all non-contextual and some contextual ads), or 2 (category, for some contextual ads).</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82476",
      "postDate": "06/20/2015 17:37:38",
      "content": "<p>A question: what is histCTR? &nbsp;Where, exactly, does it come from?</p>\n<p>Thank you!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82528",
      "postDate": "06/22/2015 07:49:15",
      "content": "<p>HistCTR&nbsp;is&nbsp;history-based estimation of click-through rate for contextual ads, calculated when the ad is showed.</p>\n<p>It derives&nbsp;largely from the context ad itself and similar context ads.</p>\n<p>You are free to ignore that field in your models.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82723",
      "postDate": "06/25/2015 12:34:47",
      "content": "<p>Hello,</p>\n<p>I have a question regarding the output format that should be submitted; should it be strictly 0 or 1, or it can be any value in the continuous interval [0,1], for example, 0.2, 0.3433 or 0.64??</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82732",
      "postDate": "06/25/2015 14:29:46",
      "content": "<p>[quote=GNU_dreamer;82723]</p>\n<p>Hello,</p>\n<p>I have a question regarding the output format that should be submitted; should it be strictly 0 or 1, or it can be any value in the continuous interval [0,1], for example, 0.2, 0.3433 or 0.64??</p>\n<p>[/quote]</p>\n<p>[0,1]. Any numerical value that satisfies: 0.0 &lt;= value &lt;= 1.0</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82740",
      "postDate": "06/25/2015 15:59:26",
      "content": "<p>&quot;IsClick - 1 if there was a click on this ad. Otherwise 0. For non-contextual ads this field equals NULL. The goal of this competition is to make a click prediction model for contextual ads.&quot;</p>\n\n<p>Event &quot;IsClick&quot; has been &nbsp;= 1,&nbsp;Event &quot;IsClick&quot; was not = 0.</p>\n<p>What is mean&nbsp;&quot;IsClick&quot; =&nbsp;0.5 ?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82743",
      "postDate": "06/25/2015 16:31:28",
      "content": "<p>[quote=Michael Swan;82740]</p>\n<p>&quot;IsClick - 1 if there was a click on this ad. Otherwise 0. For non-contextual ads this field equals NULL. The goal of this competition is to make a click prediction model for contextual ads.&quot;</p>\n<p>Event &quot;IsClick&quot; has been &nbsp;= 1,&nbsp;Event &quot;IsClick&quot; was not = 0.</p>\n<p>What is mean&nbsp;&quot;IsClick&quot; =&nbsp;0.5 ?</p>\n<p>[/quote]</p>\n<p>Means that the user thought of clicking and give up halfway.</p>\n<p>Seriously, the real event can be only {0,1}, clicked or non clicked. But the&nbsp;predictions are the probability of clicking, So it can be anywhere between 0 and 1. &nbsp;&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "82772",
      "postDate": "06/26/2015 06:34:27",
      "content": "<p>[quote=GNU_dreamer;82723]</p>\n<p>Hello,</p>\n<p>I have a question regarding the output format that should be submitted; should it be strictly 0 or 1, or it can be any value in the continuous interval [0,1], for example, 0.2, 0.3433 or 0.64??</p>\n<p>[/quote]</p>\n<p>It should be strictly between 0 and 1. Note that log-loss evaluation function is not defined at 0 and 1.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "84302",
      "postDate": "07/13/2015 19:12:43",
      "content": "<p>The submission file format is mentioned different in &quot;Evaluation&quot; Section and Sample Submission file! :-)</p>\n\n<p>We could easily figure that out though!</p>\n\n<p>Not a very popular competition I see!</p>",
      "rawMarkdown": "The submission file format is mentioned different in \"Evaluation\" Section and Sample Submission file! :-)\r\n\r\nWe could easily figure that out though!\r\n\r\nNot a very popular competition I see!",
      "votes": null
    },
    {
      "id": "85605",
      "postDate": "07/15/2015 21:41:36",
      "content": "<p>Hi,</p>\n\n<p>Can you confirm that the number of ads per search in the *SearchStream.tsv files is random and does not reflect the true number of ads shown to the user?  If this is the case, I suspect we are losing the ability to factor in an individual user's propensity to click. </p>\n\n<p>I am also wondering, why the IsClick column was set to NULL for non-contextual ads?  Again it appears this inhibits our ability to incorporate information from these non-contextual ads into our model.  From a user's perspective, do non-contextual ads appear different from contextual ads?</p>\n\n<p>Thank you in advance!</p>",
      "rawMarkdown": "Hi,\r\n \r\nCan you confirm that the number of ads per search in the *SearchStream.tsv files is random and does not reflect the true number of ads shown to the user?  If this is the case, I suspect we are losing the ability to factor in an individual user's propensity to click. \r\n \r\nI am also wondering, why the IsClick column was set to NULL for non-contextual ads?  Again it appears this inhibits our ability to incorporate information from these non-contextual ads into our model.  From a user's perspective, do non-contextual ads appear different from contextual ads?\r\n \r\nThank you in advance!",
      "votes": null
    },
    {
      "id": "1724590",
      "postDate": "03/16/2022 11:13:07",
      "content": "<p>Where can I find an academic reference to data?</p>",
      "rawMarkdown": "Where can I find an academic reference to data?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1724590,
      "author_name": "egormaslikov",
      "author_url": "",
      "post_date": "03/16/2022 11:13:07",
      "content": "<p>Where can I find an academic reference to data?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81070,
      "author_name": "alit336958",
      "author_url": "",
      "post_date": "06/05/2015 23:44:32",
      "content": "<p>Hi Ivan,</p>\n<p>Hope you are well.</p>\n<p>I am an academic in business analytics (Deakin University, Australia) and found this dataset to be quite related to my research area. I was wondering if I can use this dataset for research purposes.</p>\n<p>Regards,</p>\n<p>Ali Tamaddoni&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81178,
      "author_name": "ivanguz",
      "author_url": "",
      "post_date": "06/07/2015 20:12:51",
      "content": "<p style=\"text-align: left\">Hi&nbsp;Ali,</p>\n<p style=\"text-align: left\">Feel free to use this dataset but for academic purposes only.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81217,
      "author_name": "alit336958",
      "author_url": "",
      "post_date": "06/08/2015 08:41:41",
      "content": "<p>Hi Ivan,</p>\n<p>Thank you very much. Much appreciated.</p>\n<p>Regards,</p>\n<p>Ali Tamaddoni</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81611,
      "author_name": "gnudreamer",
      "author_url": "",
      "post_date": "06/11/2015 15:14:53",
      "content": "<p>Hello,</p>\n<p>I just have a question regarding the attribute 'level' in both the 'location' and 'category' tables.</p>\n<p>What does it signify?</p>\n\n<p>Thank you</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81643,
      "author_name": "nightbat",
      "author_url": "",
      "post_date": "06/11/2015 23:05:03",
      "content": "<p>Hello,</p>\n<p><strong>Search</strong></p>\n<p>Level in&nbsp;Location (once you join SearchInfo and Location on SearchLocationID = LocationID) represents geo level&nbsp;where a search (and impression) took place. Values could be 3 (city), 2 (region), or 1 (whole country).</p>\n<p>Level in&nbsp;Category (once you join SearchInfo and Category&nbsp;on SearchCategoryID&nbsp;= CategoryID)&nbsp;represents category&nbsp;level where a search (and impression) took place. Values&nbsp;could be&nbsp;3 (subcategory), 2 (category), or&nbsp;1&nbsp;(no specific category or subcategory were selected by visitor).</p>\n<p><strong>Ads</strong></p>\n<p>Level in Location (once you join AdsInfo and Location using&nbsp;LocationID) represents geo level belongs to&nbsp;a non-contextual ad. It has to&nbsp;be 3 (city). Note that for context ad LocationID&nbsp;is NULL.</p>\n<p>Level in Category (once you join AdsInfo&nbsp;and Category using&nbsp;CategoryID) represents category level belongs to any&nbsp;ad. Values&nbsp;could be either 3 (subcategory, for all non-contextual and some contextual ads), or 2 (category, for some contextual ads).</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82476,
      "author_name": "dotdashdashdash",
      "author_url": "",
      "post_date": "06/20/2015 17:37:38",
      "content": "<p>A question: what is histCTR? &nbsp;Where, exactly, does it come from?</p>\n<p>Thank you!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82528,
      "author_name": "nightbat",
      "author_url": "",
      "post_date": "06/22/2015 07:49:15",
      "content": "<p>HistCTR&nbsp;is&nbsp;history-based estimation of click-through rate for contextual ads, calculated when the ad is showed.</p>\n<p>It derives&nbsp;largely from the context ad itself and similar context ads.</p>\n<p>You are free to ignore that field in your models.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82723,
      "author_name": "gnudreamer",
      "author_url": "",
      "post_date": "06/25/2015 12:34:47",
      "content": "<p>Hello,</p>\n<p>I have a question regarding the output format that should be submitted; should it be strictly 0 or 1, or it can be any value in the continuous interval [0,1], for example, 0.2, 0.3433 or 0.64??</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82732,
      "author_name": "leustagos",
      "author_url": "",
      "post_date": "06/25/2015 14:29:46",
      "content": "<p>[quote=GNU_dreamer;82723]</p>\n<p>Hello,</p>\n<p>I have a question regarding the output format that should be submitted; should it be strictly 0 or 1, or it can be any value in the continuous interval [0,1], for example, 0.2, 0.3433 or 0.64??</p>\n<p>[/quote]</p>\n<p>[0,1]. Any numerical value that satisfies: 0.0 &lt;= value &lt;= 1.0</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82740,
      "author_name": "michaelswan",
      "author_url": "",
      "post_date": "06/25/2015 15:59:26",
      "content": "<p>&quot;IsClick - 1 if there was a click on this ad. Otherwise 0. For non-contextual ads this field equals NULL. The goal of this competition is to make a click prediction model for contextual ads.&quot;</p>\n\n<p>Event &quot;IsClick&quot; has been &nbsp;= 1,&nbsp;Event &quot;IsClick&quot; was not = 0.</p>\n<p>What is mean&nbsp;&quot;IsClick&quot; =&nbsp;0.5 ?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82743,
      "author_name": "leustagos",
      "author_url": "",
      "post_date": "06/25/2015 16:31:28",
      "content": "<p>[quote=Michael Swan;82740]</p>\n<p>&quot;IsClick - 1 if there was a click on this ad. Otherwise 0. For non-contextual ads this field equals NULL. The goal of this competition is to make a click prediction model for contextual ads.&quot;</p>\n<p>Event &quot;IsClick&quot; has been &nbsp;= 1,&nbsp;Event &quot;IsClick&quot; was not = 0.</p>\n<p>What is mean&nbsp;&quot;IsClick&quot; =&nbsp;0.5 ?</p>\n<p>[/quote]</p>\n<p>Means that the user thought of clicking and give up halfway.</p>\n<p>Seriously, the real event can be only {0,1}, clicked or non clicked. But the&nbsp;predictions are the probability of clicking, So it can be anywhere between 0 and 1. &nbsp;&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 82772,
      "author_name": "mas313",
      "author_url": "",
      "post_date": "06/26/2015 06:34:27",
      "content": "<p>[quote=GNU_dreamer;82723]</p>\n<p>Hello,</p>\n<p>I have a question regarding the output format that should be submitted; should it be strictly 0 or 1, or it can be any value in the continuous interval [0,1], for example, 0.2, 0.3433 or 0.64??</p>\n<p>[/quote]</p>\n<p>It should be strictly between 0 and 1. Note that log-loss evaluation function is not defined at 0 and 1.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 84302,
      "author_name": "phanisrikanth",
      "author_url": "",
      "post_date": "07/13/2015 19:12:43",
      "content": "<p>The submission file format is mentioned different in &quot;Evaluation&quot; Section and Sample Submission file! :-)</p>\n\n<p>We could easily figure that out though!</p>\n\n<p>Not a very popular competition I see!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 85605,
      "author_name": "rhinodaveb",
      "author_url": "",
      "post_date": "07/15/2015 21:41:36",
      "content": "<p>Hi,</p>\n\n<p>Can you confirm that the number of ads per search in the *SearchStream.tsv files is random and does not reflect the true number of ads shown to the user?  If this is the case, I suspect we are losing the ability to factor in an individual user's propensity to click. </p>\n\n<p>I am also wondering, why the IsClick column was set to NULL for non-contextual ads?  Again it appears this inhibits our ability to incorporate information from these non-contextual ads into our model.  From a user's perspective, do non-contextual ads appear different from contextual ads?</p>\n\n<p>Thank you in advance!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "80703": "",
    "81070": "",
    "81178": "",
    "81217": "",
    "81611": "",
    "81643": "",
    "82476": "",
    "82528": "",
    "82723": "",
    "82732": "",
    "82740": "",
    "82743": "",
    "82772": "",
    "84302": "The submission file format is mentioned different in \"Evaluation\" Section and Sample Submission file! :-)\r\n\r\nWe could easily figure that out though!\r\n\r\nNot a very popular competition I see!",
    "85605": "Hi,\r\n \r\nCan you confirm that the number of ads per search in the *SearchStream.tsv files is random and does not reflect the true number of ads shown to the user?  If this is the case, I suspect we are losing the ability to factor in an individual user's propensity to click. \r\n \r\nI am also wondering, why the IsClick column was set to NULL for non-contextual ads?  Again it appears this inhibits our ability to incorporate information from these non-contextual ads into our model.  From a user's perspective, do non-contextual ads appear different from contextual ads?\r\n \r\nThank you in advance!",
    "1724590": "Where can I find an academic reference to data?"
  },
  "source": "meta"
}