I read some paper (* about predicting user retention in StumbledUpon) and saw the authors provide a list of features with accuracy of each feature with the following explanation:

enter image description here As decision trees use information gain to find the best feature, we instead use decision trees with single features and report their accuracies on the decision tree classification for retained and non-retained URLs. Table 2 shows the ranking based on chi-square measure and the decision tree accuracy on individual features.

I am confused about how they can do it. I can just imagine using chi-squared to make a rank and to know the most important features (features evaluation), but in this case I don't know how to calculate the accuracy of each feature. Thank you.

  • $\begingroup$ Your question is very unclear. Can you please edit your answer to make it explicit exactly what you're asking please? $\endgroup$
    – Glen_b
    Jun 2, 2014 at 8:53
  • $\begingroup$ @Glen_b thank you for comment, I already add a figure in my question, the problem is how they can calculate the accuracy for each feature? $\endgroup$
    – user46543
    Jun 2, 2014 at 9:12
  • 1
    $\begingroup$ A paper you don't cite gave results you don't understand. Sorry, but this is very unclear. Even if you gave the reference, we would still have to wade through it to understand the question. $\endgroup$
    – Nick Cox
    Jun 2, 2014 at 9:17
  • 3
    $\begingroup$ I'd write to the authors. There are email addresses. You're new here, but please try a little role reversal: People who answer questions do so for various reasons, including the question being interesting and their wanting to help others. But your question is still: Please read through a densely written 11 page paper and tell me what the authors are doing. Would you do that for someone else? Questions work well here when they can be understood by reading what the OP writes, in at most a few minutes. Your question doesn't qualify in my view. $\endgroup$
    – Nick Cox
    Jun 2, 2014 at 9:45
  • 2
    $\begingroup$ I'd also advise asking one of the authors via email (and then answering your own question here). There seems to be more than one way to measure predictive accuracy in classification. One way is likely standard in this particular application area (so someone here may well be able to give a very educated guess), but the way to be sure is to ask them what they did. It's what I'd do if I had this question - just ask the authors what the definition of accuracy was in Table 2. If you do that, take some time over composing your question - make it as clear and concise as you can (they're likely busy). $\endgroup$
    – Glen_b
    Jun 2, 2014 at 9:57


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.