It’s a simple principle which could help improve millions of lives.
Since time immemorial, color has been associated with emotion. Bright colors are associated with positive emotions, while neutral or darker ones are more likely to carry on negative feelings. This simple concept led Harvard’s Andrew Reece and Chris Danforth to believe that they could investigate if someone was suffering from depression only by looking at their Instagram photos.
They recruited some 500 workers from Amazon’s Mechanical Turk, 170 of which agreed to have their Instagram accounts analyzed. Of these, 70 users were clinically depressed.
Overall, researchers identified 40,000 photos, selecting the most recent 100 photos for each volunteer. For depressed individuals, they chose the 100 photographs posted before their diagnosis. They identified some relevant variables, such as hue, color saturation, contrast and so on, while also subjectively rating the “happiness” of a picture from 0 to 5. They also used a face recognition software to identify how many people there were in the photo.
The thing with Instagram is that it makes it very easy to apply filters in a way that’s representative for your mood, and you can greatly influence the aspect of a photo. Researchers quickly noticed that people suffering from depression preferred the ‘Inkwell’ filter, which converts color photographs to black-and-white images. They also favored pictures with more black or blue in them, and showed an overall preference towards darker colors.
Meanwhile, non-depressed volunteers preferred the ‘Valencia,’ a filter that lightens photographs. The detection algorithm isn’t perfect, but overall, it boasted a 70% accuracy, which could at the very least highlight people at risk for depression. The fact that this information can be so easily obtained from Instagram could make a big different for the hundreds of millions of users.
“These findings support the notion that major changes in individual psychology are transmitted in social-media use, and can be identified via computational methods,” say Reece and Danforth.