Auto-thumbnail generator model based on interval measures for tube sites I think video streaming services e.g. tube like sites could become the next playground for data scientists, particularly I was quite interested in play rate enhacing and want to do some research about this topic. It came to me some ideas on how user-systen information could be analyzed in the way of building something like an auto-thumbnail generator model that selects the highest density frame of the clip, this would involve working with datasets where each measure looks more like an interval. Let's say a clip starts on second 0 and user1 decides to skip directly to second 120 after watching only the first 15 seconds and then almost inmediatly jumps backs to the beginning - let's say second 10. 
The data would look something like this:
0:15
120:121
10:X
I wonder if there exists any kind of framework or technique to work with this sort of intervalar measures, hopefully my explanation makes sense.
 A: It seems to me that the natural thing to do would be to count the number of times each frame had been played. So the interval 0:15, in your notation, would correspond to +1 play for each of the (360-450, depending on framerate) frames in that interval. So this would be (highly correlated) count data. The frames between 10 and 15s would end up with count 2, so those would be the most popular.
I think an interesting approach, which could perhaps be combined with what you suggest, is recent work done by Guillermo Sapiro, Rene Vidal, and Ehsan Elhamifar (abstract) about finding "representative" points in a high-dimensional space. I disagree with their terminology--I would call them "characteristic" or better yet "generative" points--but I think it would be a good approach to finding good thumbnails. I can't find the full article online, but their approach finds a few points which in a sense generate the rest of the data--then you could use your method to choose the most interesting of these candidate points, based on which is played most often (or which best represents frames that were played most often).
You might also need to make some adjustment for the beginning, since many people might start watching and then stop, which would over-weight the first seconds. But that might not be bad after all.
Anyway, lots to think about in this area.
