A Blog by Jonathan Low


May 14, 2015

Scientists Believe They Can Predict Which Images Will Go Viral

We may not be happy about a proliferation of cat images, but at least we'll know why it's happening. JL

Jim Algar reports in Tech Times:

Just 1 out of 20 posted images gets even one share. Photos that were shared across different fan groups or friendship networks yielded a 67 percent success rate in predicting doubling. The most accurate predictor of increase? Speed.
Facebook researchers who've collaborated with university scientists to study the fate of photos on the social media site say such "cascades" of a photo are rare, and just 1 out of 20 posted images on Facebook gets even one share.
Only 1 out of 4,000 manages more than around 500 shares, they say.
Stanford computer scientist Jure Leskovec says that makes predicting the events difficult.
"It wasn't clear whether information cascades could be predicted because they happen so rarely," says Leskovic, who teaches in the university's Machine Learning Department.
In a study being presented to the International World Wide Web Conference in Seoul, South Korea, the Stanford researchers report how they successfully predicted 80 percent of the time when a particular photo cascade would achieve a doubling of shares.
In an analysis of 150,000 photos from Facebook, they searched for variables to help predict doubling instances, including the speed and rate at which particular photos were being shared.
They also looked at the configuration of the sharing process, noting that photos that were reposted in several networks were likely to generate stronger cascades.
The more a particular photo was shared, the more accurate their "cascade" prediction became, they reported, and when shares reached the hundreds the successful doubling prediction soared to 88 percent.
What was the most accurate predictor of a cascade increase? Speed, the researchers said; just examining how fast a cascade took off allowed doublings to be predicted in 78 percent of instances.
"Slow, persistent cascades don't really double in size," Leskovec says.
The configuration of a cascade -- how a particular photo was being shared -- was found to be the next best factor for predicting an increase, the researchers said.
Photos that were shared across different fan groups or friendship networks were evidence of a wide response in interest, yielded a 67 percent success rate in predicting doubling.
However, there was no sure and simple way to guarantee sharing, they said.
"It is very hard to quantify what going viral means," Leskovec says. "Anyone would say 'Gangnam Style' went viral, but that's a singular event," he said of the YouTube posting viewed around two billion times.
It's all a bit of a roll of the dice, he says.
"Even if you have the best cat picture ever, it could work for your network, but not for my boring academic friends," he explained. "You have to understand your network."


Post a Comment