Which would you rather have for lunch; something a good friend ordered for you, or the most popular lunch among people who also enjoyed such hits as Turkey Sandwich and Fish Taco?
A few years back, Netflix made headlines by offering a prize of one million US dollars to anyone who could build a better recommendation engine for their online movie rental site. Crazy? Not according to Netflix. Their assertion was that if they could get just ten percent better at recommending movies based on users’ ratings (1 to 5 stars) of previously-viewed films, their revenues would increase by much more than a paltry $1M.
What has always boggled my mind about this challenge is that it’s a classic case of struggling to find a technological solution to a distinctly human problem. Google labors to perfect computer algorithms that convert recorded speech to text. For all their research and computational might, they do a pretty poor job. Meanwhile, a small, smart company called CastingWords* uses Amazon’s “Mechanical Turk” service to assign transcription tasks to hundreds of eager, human laborers who work for pennies. The results are near-perfect.
Netflix is sitting on a nearly Wikipedia-sized repository of user-generated movie reviews. These reviews are cheaper than free to Netflix — since only active members can contribute them, people are actually paying for the privilege of reviewing films on Netflix’s site.
Netflix not only ignores these reviews in recommending movies, it also ignores your reaction to the reviews. This is the 100% human answer to their “technological” problem that has been staring them right in the face for years.
When I give Face/Off five stars, am I doing it because I love John Travolta Movies? Or Nick Cage flicks? Or John Woo films? Or hyper-violent '90s action? Or any film that features a speedboat chase? Netflix has no idea why I like or dislike a movie, so how can they predict what I might like or dislike next?
When I read the Netflix reviews of Face/Off, two things are abundantly clear: First, many people like that movie for reasons I don’t agree with. Second, people who don’t like Face/Off, with a few notable exceptions, are people whose cinematic opinions I can live without.
In Real Life I have friends who love Face/Off, and friends who hate it. And crazily enough, I respect all of their opinions. Any of these friends are welcome to recommend movies to me, and I will almost always take those recommendations.
That’s Netflix’s second mistake: Thinking that we always only want to watch movies that we’d rate highly. I don’t know about you, but I watch plenty of movies that I know won’t be five-star favorites. A friend’s strong endorsement is often the reason — even that guy who hates Face/Off.
But let’s get back to Netflix’s first mistake: Thinking that what everyone thinks matters. I’m sorry, but everyone is an idiot. Even with the challenge complete and the fancy new algorithm implemented, everyone seems to think that because I liked Mission Impossible III, I’ll jump for anything starring Tom Cruise. Or that liking the first four Steven Seagal films has anything to do with one’s opinions of his subsequent works.
Read some Netflix reviews. While a few are insightful, most are utter garbage.
If a person’s review seems garbage to me, then what good is their star rating to me? None whatsoever. So the majority of the data used by these million-dollar algorithms is worse than worthless. No wonder the results plateau, despite endless efforts.
Of course, your garbage might be my delicacy. On Netflix, you can rate reviews Helpful or Unhelpful. But Netflix obstinately sticks to it’s “everyone’s opinion” philosophy and uses these ratings to bubble "Helpful" reviews to the top of the list.
The list. No matter how you rate the reviews, you see the same list as everyone else.
You can probably imagine what I’m going to say next. Unless, of course, you work for Netflix.**
If I read a compelling, insightful review on Netflix, and I mark that review Helpful, then factor that person’s ratings higher in determining what films I might like. Similarly, if I mark a review as Unhelpful, then don’t let that fool’s opinions influence my recommendations.
Let me build a personal network of people whose opinions I respect, and let their recommendations populate my personalized Netflix experience. It will immediately become obvious that a few, cultivated opinions are worth much more that a watered down average of millions.
Oh, and speaking of millions, you can keep the check Netflix. Better recommendations will be reward enough. In case you hadn’t noticed, I’m hooked on you like strawberry crack.
