{
  "id": 12647,
  "title": "What changes would you make in a production environment?",
  "url": "/competitions/avazu-ctr-prediction/discussion/12647",
  "author_name": "",
  "post_date": "2015-03-01T17:07:10.643Z",
  "votes": null,
  "comment_count": 1,
  "views": 866,
  "content": "<p>A main purpose&nbsp;of these CTR models is to rank a set of ads at impression time, and determine which will generate the most revenue.&nbsp;In the industry, ad revenue is often times measured by Revenue Per 1000 Impressions, or RPM.</p>\n<p style=\"padding-left: 30px\"><code>RPM = CTR * Revenue&nbsp;Per Click * 1000;</code></p>\n<p>With this idea, I&#8217;m curious what changes you would make in a production environment to maximize RPM.</p>\n<p>Are there other features that you would collect? There is a plethora of additional data available at impression time - even so far as what the weather is like in a user&#8217;s area.</p>\n<p>How well suited are the models proposed here for predicting clicks in a many-advertisement environment? Often times, 5 or more ads are displayed adjacently, for example on a Google Search page. What would you change to account for this?</p>",
  "messages": [
    {
      "id": "65184",
      "postDate": "03/01/2015 17:07:10",
      "content": "<p>A main purpose&nbsp;of these CTR models is to rank a set of ads at impression time, and determine which will generate the most revenue.&nbsp;In the industry, ad revenue is often times measured by Revenue Per 1000 Impressions, or RPM.</p>\n<p style=\"padding-left: 30px\"><code>RPM = CTR * Revenue&nbsp;Per Click * 1000;</code></p>\n<p>With this idea, I&#8217;m curious what changes you would make in a production environment to maximize RPM.</p>\n<p>Are there other features that you would collect? There is a plethora of additional data available at impression time - even so far as what the weather is like in a user&#8217;s area.</p>\n<p>How well suited are the models proposed here for predicting clicks in a many-advertisement environment? Often times, 5 or more ads are displayed adjacently, for example on a Google Search page. What would you change to account for this?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65190",
      "postDate": "03/01/2015 18:46:12",
      "content": "<p>In that case I wouldn't have been minimalizing logloss, because if you answer 0.0 you get enormous penalty (but in industry you don't lose that much).<br><br>I do expect that calculating FM and ensembling models in real world application would be exasperatedly long, so I'd stick to ftrl without ensambling (especially for testing, when you have to anwer in miliseconds).</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 65190,
      "author_name": "masterblaster",
      "author_url": "",
      "post_date": "03/01/2015 18:46:12",
      "content": "<p>In that case I wouldn't have been minimalizing logloss, because if you answer 0.0 you get enormous penalty (but in industry you don't lose that much).<br><br>I do expect that calculating FM and ensembling models in real world application would be exasperatedly long, so I'd stick to ftrl without ensambling (especially for testing, when you have to anwer in miliseconds).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "65184": "",
    "65190": ""
  },
  "source": "meta"
}