What is the interpretation of output 0.5 of a typical classifier?

I made a prediction and the probability of that data point being from the True class is 0.5.

  • $\begingroup$ Can you give a little more context? I’m not quite sure what you’re asking. Did you make a prediction and the probability of that data point being from the True class is 0.5? $\endgroup$ – gingermander Dec 17 '17 at 14:19
  • $\begingroup$ @gingermander Yes, Exactly and How will you explain that ? $\endgroup$ – Abhishek Sharma Dec 17 '17 at 14:30

It sounds like your model is saying there is a 50% chance that this data point could be in either class. Since this is a special case, prediction algorithms need to pick which class to choose when there is an equal probability for each class. In this case, it looks like your model defaults to the True class.

There are some models in sklearn like random forest and SVM that have predict_proba functions. You might find more information if you look there or check if the model you are using has documentation on how its probabilities are calculated.

| improve this answer | |

Depending on your algorithm, it may have different interpretations.

Suppose you are using SVM with kernels, it means your input data is exactly between two data points with different classes or if you are using linear kernels in SVM it means that your data is on the separator line.

If you are using Neural Networks, this is the Probability of each class. In such cases there maybe different explanations for this phenomenon.

  1. Suppose that you are trying to classify images of green and purple cats! and suppose that you have two same cats with both colors, then your classifier may output this result
  2. Another interpretation may also exist. suppose that you have unbalanced data-set for your two classes. suppose that there are same feature vectors of both classes. in such cases, the distribution of your samples overlap. A class which has more data samples would be the winner and will occupy the space much more than the class with less samples. In such cases, if you get 50 percent as the output, You can not tell that both classes have same expectations.

    If I want to clarify this more, I would do that with this example. Suppose that you have a car classifier for distinguishing between white and blue cars. during training you had 100 images of blue car and 20 images of white car. During recall phase, if for an arbitrary image you have 50 percent for each class, you can never say that the image has same probability to be each one at all


| improve this answer | |
  • $\begingroup$ I loved the answer you gave but it is too much of text, so tried segregating as that of the suggested edit. Let me know if you are not happy with that. $\endgroup$ – Toros91 Jan 11 '18 at 8:38
  • $\begingroup$ I've updated the result $\endgroup$ – Toros91 Feb 14 '18 at 2:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.