{
  "id": 55535,
  "title": "Deal Probabilities",
  "url": "/competitions/avito-demand-prediction/discussion/55535",
  "author_name": "",
  "post_date": "2018-04-28T05:00:47.372456400Z",
  "votes": 9,
  "comment_count": 2,
  "views": 0,
  "content": "<p>So, how are the deal probabilities calculated for the items in the training set ? Is it based on Avito's classification model ?\nAsking because, what if instead of considering this as a regression problem, I take it as a classification problem and try to predict softmax probabilities of the items</p>",
  "messages": [
    {
      "id": "320274",
      "postDate": "04/28/2018 05:00:47",
      "content": "<p>So, how are the deal probabilities calculated for the items in the training set ? Is it based on Avito's classification model ?\nAsking because, what if instead of considering this as a regression problem, I take it as a classification problem and try to predict softmax probabilities of the items</p>",
      "rawMarkdown": "So, how are the deal probabilities calculated for the items in the training set ? Is it based on Avito's classification model ?\nAsking because, what if instead of considering this as a regression problem, I take it as a classification problem and try to predict softmax probabilities of the items",
      "votes": null
    },
    {
      "id": "322195",
      "postDate": "05/02/2018 14:30:52",
      "content": "<p>I'd guess it is based on actual data where deal probability  = actual deal / #views</p>",
      "rawMarkdown": "I'd guess it is based on actual data where deal probability  = actual deal / #views",
      "votes": null
    },
    {
      "id": "322387",
      "postDate": "05/02/2018 20:59:37",
      "content": "<p>Treating this as a classification problem is a bad idea. Consider the following cases:</p>\n\n<p>Regression:\nactual = 0.25, predicted = 0.25   --&gt; mean squared error = (0.25 - 0.25)**2 = 0</p>\n\n<p>Classification:\nactual = 0.25, predicted = 0.25  --&gt; binary cross-entropy error = -(0.25*log(0.25) + (1 - 0.25)*log(1 - 0.25))</p>\n\n<p>For all deal probability other than 0 or 1 classification model will see non-zero error even when prefect predictions are made.</p>",
      "rawMarkdown": "Treating this as a classification problem is a bad idea. Consider the following cases:\n\nRegression:\nactual = 0.25, predicted = 0.25   --&gt; mean squared error = (0.25 - 0.25)**2 = 0\n\nClassification:\nactual = 0.25, predicted = 0.25  --&gt; binary cross-entropy error = -(0.25*log(0.25) + (1 - 0.25)*log(1 - 0.25))\n\nFor all deal probability other than 0 or 1 classification model will see non-zero error even when prefect predictions are made.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 322195,
      "author_name": "sixianghu",
      "author_url": "",
      "post_date": "05/02/2018 14:30:52",
      "content": "<p>I'd guess it is based on actual data where deal probability  = actual deal / #views</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 322387,
      "author_name": "sourabhjha",
      "author_url": "",
      "post_date": "05/02/2018 20:59:37",
      "content": "<p>Treating this as a classification problem is a bad idea. Consider the following cases:</p>\n\n<p>Regression:\nactual = 0.25, predicted = 0.25   --&gt; mean squared error = (0.25 - 0.25)**2 = 0</p>\n\n<p>Classification:\nactual = 0.25, predicted = 0.25  --&gt; binary cross-entropy error = -(0.25*log(0.25) + (1 - 0.25)*log(1 - 0.25))</p>\n\n<p>For all deal probability other than 0 or 1 classification model will see non-zero error even when prefect predictions are made.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "320274": "So, how are the deal probabilities calculated for the items in the training set ? Is it based on Avito's classification model ?\nAsking because, what if instead of considering this as a regression problem, I take it as a classification problem and try to predict softmax probabilities of the items",
    "322195": "I'd guess it is based on actual data where deal probability  = actual deal / #views",
    "322387": "Treating this as a classification problem is a bad idea. Consider the following cases:\n\nRegression:\nactual = 0.25, predicted = 0.25   --&gt; mean squared error = (0.25 - 0.25)**2 = 0\n\nClassification:\nactual = 0.25, predicted = 0.25  --&gt; binary cross-entropy error = -(0.25*log(0.25) + (1 - 0.25)*log(1 - 0.25))\n\nFor all deal probability other than 0 or 1 classification model will see non-zero error even when prefect predictions are made."
  },
  "source": "meta"
}