{
  "id": 28458,
  "title": "CNN + Linear SVM",
  "url": "/competitions/dogs-vs-cats-redux-kernels-edition/discussion/28458",
  "author_name": "mrgloom",
  "post_date": "2017-02-04T14:51:56.257000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>What is the best practice to obtain probability from SVM classifier?</p>\n\n<p>There is some related info:</p>\n\n<p><a href=\"http://stackoverflow.com/questions/26478000/converting-linearsvcs-decision-function-to-probabilities-scikit-learn-python\">http://stackoverflow.com/questions/26478000/converting-linearsvcs-decision-function-to-probabilities-scikit-learn-python</a></p>\n\n<p><a href=\"http://stackoverflow.com/questions/15111408/how-does-sklearn-svm-svcs-function-predict-proba-work-internally\">http://stackoverflow.com/questions/15111408/how-does-sklearn-svm-svcs-function-predict-proba-work-internally</a></p>\n\n<p>And I have tried <code>CalibratedClassifierCV</code> approach and also using <code>linear_model.LogisticRegression()</code> , accuracy of SVM classifier is about 96% but seems probabilities 'not tuned properly', so log loss highly penalize them.</p>\n\n<p>Any suggestions?</p>",
  "messages": [
    {
      "id": 159872,
      "postDate": "2017-02-04T14:51:56.257Z",
      "content": "<p>What is the best practice to obtain probability from SVM classifier?</p>\n\n<p>There is some related info:</p>\n\n<p><a href=\"http://stackoverflow.com/questions/26478000/converting-linearsvcs-decision-function-to-probabilities-scikit-learn-python\">http://stackoverflow.com/questions/26478000/converting-linearsvcs-decision-function-to-probabilities-scikit-learn-python</a></p>\n\n<p><a href=\"http://stackoverflow.com/questions/15111408/how-does-sklearn-svm-svcs-function-predict-proba-work-internally\">http://stackoverflow.com/questions/15111408/how-does-sklearn-svm-svcs-function-predict-proba-work-internally</a></p>\n\n<p>And I have tried <code>CalibratedClassifierCV</code> approach and also using <code>linear_model.LogisticRegression()</code> , accuracy of SVM classifier is about 96% but seems probabilities 'not tuned properly', so log loss highly penalize them.</p>\n\n<p>Any suggestions?</p>",
      "rawMarkdown": "What is the best practice to obtain probability from SVM classifier?\n\nThere is some related info:\n\nhttp://stackoverflow.com/questions/26478000/converting-linearsvcs-decision-function-to-probabilities-scikit-learn-python\n\nhttp://stackoverflow.com/questions/15111408/how-does-sklearn-svm-svcs-function-predict-proba-work-internally\n\nAnd I have tried `CalibratedClassifierCV` approach and also using `linear_model.LogisticRegression()` , accuracy of SVM classifier is about 96% but seems probabilities 'not tuned properly', so log loss highly penalize them.\n\nAny suggestions?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "159872": "What is the best practice to obtain probability from SVM classifier?\n\nThere is some related info:\n\nhttp://stackoverflow.com/questions/26478000/converting-linearsvcs-decision-function-to-probabilities-scikit-learn-python\n\nhttp://stackoverflow.com/questions/15111408/how-does-sklearn-svm-svcs-function-predict-proba-work-internally\n\nAnd I have tried `CalibratedClassifierCV` approach and also using `linear_model.LogisticRegression()` , accuracy of SVM classifier is about 96% but seems probabilities 'not tuned properly', so log loss highly penalize them.\n\nAny suggestions?"
  }
}