1

I've encountered an interview question:

Given several binary labels, each label represents a user will click a certain advertisement or not, we have a trained logistic model and its predicted probability of a user clicking the advertisement. How to evaluate this trained logistic model?

I answered by using different thresholds on the predicted probability, we can easily plot the ROC curve and then area under the curve should be a measure. The interviewer said that this is Okay, but can you give me other methods? I'm wondering how to answer this question. By the way, I failed at last. 

         click or not     predicted
user1        1               0.8
user2        1               0.6
user3        0               0.4
...         ...              ...  
usern        0               0.3
FihopZz
  • 1,923
  • 4
  • 11
  • 9
  • 1
    See: http://stats.stackexchange.com/a/171300/858 – Neil G Sep 07 '15 at 05:27
  • 1
    Maybe this helps: http://stats.stackexchange.com/questions/164204/how-to-compare-probability-predictive-ability-of-models-developed-from-logisti/164229#164229 –  Sep 07 '15 at 09:22

1 Answers1

0

You may want you consider the log loss function to evaluate the accuracy performance of the model. Check this link out that discusses the log loss function.

https://www.quora.com/What-is-an-intuitive-explanation-for-the-log-loss-function

stochasticcrap
  • 213
  • 1
  • 2
  • 5