A fine post by Tren Griffin this morning about Netflix’ Reed Hastings. A very important quote:
“We are investing heavily in [machine learning] because we want it to be: you turn on Netflix and there’s a row, there’s like four choices, and you just want to watch them all. To get to that consistent view is where we are targeting.”
If this is true, and I believe that Netflix (and Amazon) are using machine learning, why are the recommendations always terrible?
I’ve seen a stat that shows people spend more than 20 minutes daily looking for things to watch. Anecdotally no one has ever told me they liked the Netflix (or Amazon) searching or browsing ability. And of course if you add in a mix of people e.g. Mom, Dad, Kid 1, Kid 2, then the matrix of options and bad choices appears infinite.
A few thoughts on why these brilliant minds and powerful machines are failing:
- The tools being used seem to believe that because you liked or bought one thing, you probably want another. How often do you buy something from Amazon, e.g. a broom, only for the rest of the Internet to scream “you probably need a second, maybe third, maybe fourth broom!!!” And you know what? I never do. If I bought something, it was fairly unique. And I won’t need it again for another many years. The same goes for movies – if I watched a RomCom, odds are good I will watch something different next time.
- Sample size is small, cost of making a choice is high. I might watch one or two shows a week on Netflix. I simply don’t have 4+ hours available. So each movie I select is a big deal – it can’t suck. Contrast this with YouTube which has a ton of content and the “buy in” is very small. If I select a bad 8-minute clip it won’t matter much. And I can select 10 of these things (and thereby provide 10 signals to Google) vs. 1 to Netflix.
- For some strange reason Netflix doesn’t seem to take abandonment into account. “Hey, you watched Vikings, I bet you want some more Norse-y stuff”. Well no, I started ‘The Vikings’ and found it unwatchable.
I’m sure Netflix and Amazon are both working hard on this – I wonder sometimes if they ever leave the office and just watch how people really interact with content. Book recommendations are even worse if that’s possible so this isn’t a unique issue.