/Predicting Popularity

Predicting Popularity

The following two tabs change content below.

Joseph Iacoviello

Latest posts by Joseph Iacoviello (see all)

Artificial intelligence in computer science is often viewed as emulating or predicting the actions of a single intelligent individual, such as a human or an animal. Not many people picture artificial intelligence as emulating and predicting the actions of a large group of individuals, an entire population. These forms of artificial intelligence often are used to predict the popularity of various memes.

Aditya Khosla, a PhD candidate at MIT’s Computer Science and Artificial intelligence Lab, is currently working on an artificial intelligence designed to predict the popularity of your photos before you even upload them. He has been crunching through 2.3 million Flickr photos to discover what makes a photo popular on the internet. The algorithm takes into account several different factors about the photo. The first factors it considers are the social factors of the photo, such as how many followers you have, the length of the title, and the number of tags you included with the photo. The algorithm then looks at the properties of the actual photo. It measures features such as color, gradient, and objects found inside the photo. Aditya put his tool online for demonstration along with his paper at http://popularity.csail.mit.edu/. To get a rating for your photo, you upload a photo or point the tool to a site, and it gives you a score from 1 to 10. This score is a rough logarithm of how many views that image will get per day, an image with a score 6 will receive 2^6 = 64 views. Because this is a rough estimate, this score is more useful as a relative score.

Using this data, Aditya has also developed a program to tweak photos of faces to make them more memorable. The program takes a photo and refines each one slightly, making thousands of new photos, gradually making each one more memorable as the program continues. Not even the researchers working on the program know exactly what changes the program makes. The study’s senior author, Aude Oliva, presented this program at the International Conference on Computer Vision in December of 2013. “It looks like the faces that are more memorable are a little slimmer, but that is just my interpretation,” Dr. Oliva said. “And those that are more forgettable look a little rounder, but we really do not know.”

This isn’t the first time people have used social data to predict popularity. A group of researchers from Tottori University in Japan developed a mathematical formula designed to predict the box office success of a movie given various social factors such as advertising, word-of-mouth, and social networks. Professor Akira Ishii, the lead researcher, hopes to sell the algorithm to marketers and other people who could make use of such data, commenting that: “Currently the calculations are such that they can’t really be used other than by the students working in our lab. In the future we’d like to incorporate the equations into software that is easy to use by marketers or non-technical people that aren’t so familiar with mathematics. Our dream is that our software could be used by advertisers or for conferences in a broad range of industries.”

Prediction and emulation of large populations, though not often thought of as AI, is a valuable part of artificial intelligence. These are the algorithms that make all the Big Data collection that companies are doing worth the effort. With large amounts of data and predictive algorithms, companies can predict how a specific population will react to certain events, and use that data to sell a product, or profit in some other way.