Rise Of Recommendation Systems: How Machines Figure Out The Things We Want
NOEL KING, HOST:
You know the recommendations that companies give you sometimes, so the movie that Netflix says you're going to want to watch next or the top news article on your Twitter feed, all of these are driven by what are known as recommendation algorithms. Mary Childs from our Planet Money podcast reports that these algorithms can be traced back to one person who was trying to clean out his inbox.
MARY CHILDS, BYLINE: In 1990, Doug Terry was an inventor at the place for computer inventions back when email was new. It didn't really have rules or etiquette yet, so inboxes were full of chain letters and reply-all threads. Terry wanted to invent a new way to filter it, to differentiate between his boss requesting a meeting and the 43rd reply on a thread about baseball.
DOUG TERRY: Computers were very good at doing the mundane tasks but were never going to be good at doing the judgmental things, the subjective reasoning that humans could do.
CHILDS: So Terry and his team create buttons for incoming email, like it and hate it. This email is good. I like it. This email is bad. Don't show me this kind of email. And their inboxes get cleaner. They invite some co-workers to also start hitting like and hate on their emails, which leads to the next big insight, the one that will change the internet. Terry doesn't need to rate every single email because his colleagues are doing it, too, with very similar emails.
TERRY: That's the community aspect. Then other people could say, oh, show me articles sent to this newsgroup on baseball that Doug liked or don't send me anything that Doug didn't like.
CHILDS: He calls this collaborative filtering, and it works. It reduces the time he spends on his inbox. And because this is early computer days at the computer research lab, his idea becomes part of the canon. Over the next 20 years, the idea of collaborative filtering evolves into this huge industry powering all kinds of recommendations, so much so that our relationship to them changes.
JINGJING ZHANG: As long as we utilize recommendations for our decision-making, we are very vulnerable to the side effects brought by those systems.
CHILDS: Jingjing Zhang of Indiana University is on a team that researches these systems. A few years ago, they created a series of experiments to see if recommendations affect preferences. They gave students a song to listen to with a rating that the students thought was tailored to them. They asked the students how much they would be willing to pay for a song. It turned out the students were willing to pay way more for things that the computer said they would like, even if those ratings were manipulated. So basically, when a machine tells us we're going to like something, we trust the machine more than ourselves.
ZHANG: These will make me purchase things that I don't really need or I don't really like.
CHILDS: Today, instead of friends and colleagues filtering each other's emails, it's assumptions from machines owned by companies that want to profit off our behavior.
ZHANG: Over time, this will make the system less effective, less accurate and provide less diverse recommendations. Eventually, this impact will make the system provide similar items to everybody, regardless of personal taste.
CHILDS: The first versions of recommendation systems were meant to save us time and help us find what we want - to read our minds. But there's a huge difference between that and an algorithm designed to lead your mind, to take you to a new destination that you maybe wouldn't have chosen. Mary Childs, NPR News.
(SOUNDBITE OF MOSS OF AURA'S "DIM")
NPR transcripts are created on a rush deadline by Verb8tm, Inc., an NPR contractor, and produced using a proprietary transcription process developed with NPR. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.