On 26 Apr 2014, at 12:52, P.J. Alling <[email protected]> wrote:
> I'd love to do a statistical study, using real statistics on this. However > that just seems like too much work. On the other hand, based on the page, > pictures of people are more popular than scenery in instagram, Duh. Pictures > of Women, (and maybe men in miniskirts, he doesn't specify), double Duh. MIT > should just stick to what they do best, creating world beating Blackjack > counting teams. > > On 4/26/2014 11:58 AM, Christine Aguila wrote: >> Hi Everyone: >> >> I saw this short piece just now on the Verge. Here’s paragraph 2: >> >> "Khosla says his algorithm allows him to predict how many views your photo >> will get before you even upload it. The algorithm considers social factors >> such as how many followers a user has, the number of tags on the photo, and >> the length of the title. It also measures content factors such as texture, >> color, gradient, and objects present in the photo. (Miniskirts, bright >> colors, people instead of scenery = good. Plungers = bad. Pink and yellow >> miniskirts, even better. Green plungers, horrible.)” >> >> http://www.theverge.com/2014/4/24/5647270/mit-algorithm-predicts-how-popular-your-instagram-photo-will-be >> >> Cheers, Christine > > Without going to the page, I wonder if there is not a fundamental flaw. I can believe that it is possible to predict that women in min-skirts will draw more views than scenery on a site designed for sharing pictures of people. However, so what? Attracting views is one thing, but the popularity of an image is something totally else. My definition of popularity would include not only number of views but also number of downloads, number of prints made, weighted by the size of the prints made, degree of variability of over time (where a popular image would continue to get views after the initial spike), etc. So, in short, I would argue that number of views is mostly irrelevant if you are concerned with “popularity.” stan -- PDML Pentax-Discuss Mail List [email protected] http://pdml.net/mailman/listinfo/pdml_pdml.net to UNSUBSCRIBE from the PDML, please visit the link directly above and follow the directions.
