The problem is as follows:

Tags are used to classify images. An image with the tag "sky" has 90% chance of containing a sunset, with the tag "cloud" giving an image 50% chance of containing a sunset. Given an image with both tags, what is the likelihood that this image contains a sunset? The addition of extra tags should increase the likelihood, i.e. the probability will be at least equal to the highest tag (in this case 90%).

In general:

Given the probability of C occurring given A is x, and the probability of C occurring given B is y, what is the probability of C occurring given both A and B have occurred?

The general solution to this seems like something that should be somewhat obvious and easily searchable, but my mind is drawing a blank on what to search for here.

  • $\begingroup$ Logical observation: If all cloud images are sky images, and half of all cloud images are sunset images, then half of all cloud-sky images are sunset images. It seems clear that more information is needed. $\endgroup$ – Logophobic Jan 28 '16 at 11:54
  • $\begingroup$ I suppose having tag "cloud" and tag "sky" are not independent events, as clouds usually appear in the sky (except e.g. computing clouds). Also I doubt more tags should always increase the probability in this case: moonlit night images may well depict clouds in the sky without a sunset, so additional tag "night" shoud decrease the probability in question... $\endgroup$ – CiaPan Aug 22 '17 at 10:34

If there is independence then the probability of no sunset is: $$\frac{10}{100}\frac{50}{100}=\frac1{20}$$


A priori there are $8$ different cases: $s\in\{0,1\}$, $c\in\{0,1\}$, and $\odot\in\{0,1\}$ with unknown dependencies between $s$, $c$, and $\odot$. You have to find out whether the available data suffice to determine the eight probabilities $p_{s\,c\,\odot}$. If yes it is then easy to compute the conditional probability you are looking for.


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.