Facebook tunes ranking of stories in news feeds
MENLO PARK – Facebook’s quest to be a personalized newspaper for the Internet age continued on Tuesday with tweaks aimed at making sure members spy hot stories from their friends.
Changes to the leading social network’s formula for figuring out which posts will be of interest included “bumping” up potentially intriguing stories that went unnoticed during prior visits to Facebook.
“It is hard for users to get back to old things; you have to scroll through things you have already seen,” Facebook news feed team engineering Lars Backstrom said while discussing the latest changes.
“We wanted to make it so people weren’t missing important stories that didn’t make top slots but were just below the fold.”
Signals weighed in the machine learning algorithm were modified to bump-up a story considered more interesting than fresher material that formerly got priority simply for being newer.
“We tweaked the model,” Backstrom said, noting that about 30,000 signals are balanced in the algorithm.
“Instead of just taking the new stories, we would take all stories that were new to you, that you haven’t seen, even if it isn’t the freshest.”
A test of the change showed that the number of stories people read in news feeds rose to 70 percent from 57 percent with “bumping,” according to Facebook.
“Story Bumping is going to be a really nice tool for people if they…are sitting with a Facebook account and have run out of things to look at,” said Facebook vice president of product Chris Cox. “It will bump up new stuff.”
News feeds were also modified to take into account the “last actor” a member interacted with and then give that friend’s posts temporary priority since they seem to be up to something interesting.
“We wanted to capture your current state of mind as you were using Facebook,” Backstrom said.
“A lot of signals are long term, such as the relationship with each friend; we wanted a real time factor.”
Facebook’s ranking software assigns numerical scores to the roughly 1,500 stories typically eligible for delivery to a member’s news feed and displays the top 300.
Powerful factors for ranking are relationships, along with how often a member comments, shares, “likes,” or otherwise acts on posts of friends. Hiding posts sinks content from that person in news feed rankings.
“Our goal is to create the best personalized newspaper for each of our readers,” Backstrom said.
Facebook engineers are experimenting with ways for News Feeds to better handle chronological posts, such as a friend firing off play-by-play updates from a sporting event.
Backstrom’s team meets each Tuesday to brainstorm ways to improve the Facebook news feed, with worthwhile ideas tested internally among workers or with a tiny fraction of the social network’s more than one billion members.
“It starts with intuition and then that gets written into code as a feature,” said Cox. “Then we look at interactions.”
Ads displayed as promoted posts in news feeds are handled separately from content generated by people’s friends or family members at Facebook, according to the ranking team.
“We figure out the most relevant news feed with the organic content, and then, as a newspaper or television program might do, we create advertising slots,” said Facebook product manager Will Cathcart.
Backstrom compared the job of ranking news feed posts to the challenge faced by Internet search engines Google or Bing when it comes to quickly determining optimal results for queries.
“Facebook is one of the only places where you have a problem on the same scale as what Google or Bing is doing but you have to use different techniques because of the personal aspects of it,” Backstrom said.