Researchers at the University of California San Diego are working on an algorithm to determine whether you’re a punk or a goth by the style of your social media posts.
The team analyzed pictures to try to classify people into one of eight subcultures based on their appearance. This included hipsters and goths as well as surfers, bikers and bikers.
The algorithm can be trained to look for tattoos, trendy hairstyles, and jewelry, for instance, based on the pictures you post on social media.
The websites can offer a more personalized experience. A surfer might receive recommendations on holidays, while a punk could be updated about gigs by their favorite band. What better way to stay on top of the latest organic and fair trade coffee products than to receive updates as soon as they are available?
What it does
Researchers use a classification algorithm known as multi-label. They are used widely in vision analysis for concluding clues found in images. The algorithm takes a series of photos with labels such as “cat,” ‘car,” and ’emo’ and finds the features that are most likely to predict the title for a new picture. The algorithm relies on the assumption that photos with similar feature values will likely have similar labels.
If it sees a photo and notices a pair of horn-rimmed spectacles, a waxed mustache, and a lumberjack t-shirt, and it’s told it’s looking at a “hipster,” it can then move to another picture and recognizes a quinoa fan by their appearance.
Researchers say that the algorithm is, on average, 48% accurate, while chance only gets answers right 9% of the time. You will get the right answer on average 11 times if you try to guess the contents of a photo without seeing it. This machine is better but still not as good as a person using their full street knowledge.
The algorithm breaks down each image into feature values using a “parts-and-attributes” approach. In this instance, attributes like tattoos, colors, haircuts, and jewelry were scanned on features such as the face, neck, arms, and torso of each subject.
The algorithm uses the labeled pictures to create a classifier. This type of problem is perfectly suited for Google machinery, as it could be used to identify features that are indicative of certain social groups without the need to specify the types of features, such as the face, the head, the top of the head (where the hat would go), the neck, the torso, and the arms.
What is it for?
Sites can provide a more personalized experience if they can recognize your personality based on how you appear.
This approach has some issues. In the current state of technology, Facebooking goths would have a good chance of getting ads for fixed-wheel bicycle repairs in their news feed. The researchers say that while 48% accuracy is better than random chance, they want their algorithm to perform like a human and will continue to work to improve it.
The deeper question is whether or not you can make any assumptions about a person’s interests based solely on their appearance. A goth’s preference for black clothing doesn’t mean that their hobbies are similar to those of a surfer.
It is important to consider whether we would like our Internet experience to be customized in this manner. It can be annoying to see ads and search results tailored based on our gender. Facebook thinks that just because someone is female, they will be interested in celebrity diet news.