Better suggestions through tagging: How Netflix deconstructs Hollywood, by Mike Hastings
Mike works at Netflix (no relation to the CEO with the same last name). He suggests that channel surfing is intellectual, not lazy. The channel surfer identifies who’s on the screen, what else they’ve been in, how colorful the screen is and what that means (comedy vs drama), whether it’s kid friendly, etc. There’s a lot (subliminally) being identified and decided.
“Common objects become strangely uncommon when removed from their context and ways of being seen.” -Wayne Thiebaud
For Netflix, they’re showing TV and movie options out of context – no scenes, no hints to help people make quick decisions. How do they make sure they’re providing people with enough info to go on, compared to the context they would get flipping channels?
They create algorithms that drive everything. But what they can’t do is make sense of all the data. They had to determine how to provide context along with these titles being recommended – an editorial plan.
And they had to figure out how to scale this editorial plan.
Stop thinking of human vs. machine, and start thinking of editorial (human) as part of machine.
To do the tagging, their requirements are:
- Experience analyzing scripts (movie reviewer, or film school, or screenwriter)
- Objectivity – it doesn’t matter if you think it’s hilarious, what matters is that it’s intended to be a thriller
- Tag by things that identify every single character, and the typical scenarios
- Some things get tagged on a 1-10, others are yes/no
- They categorize over 180 qualities in their tagging
Tangled: overall tone is “family friendly”
- This gets tricky because some things are family friendly in one country but not another
- They try to tag globally, not locally, to help with scaling
- One user might get it recommended in “Family friendly movies,” and another might get it in “Friendly Adventures”
- Business rules:
- They have to figure out how to combine genres. If “Family friendly” is combined with “Adventures” then it’s “Family Adventures”
- If “Horror” is combined with “Scary ,” just say “horror.” But it “Horror” is combined with “Suspenseful” then it’s “Suspenseful Horror”
Contraband: one tag is “heavy violence”
- One person might get it in “Violent Action Thrillers” and another might get “Violent Revenge Movies”
- Some tags are used to exclude as well as include. If you’re watching “Family Friendly,” then “Violent” movies get excluded
House of Cards: importance of location is “crucial”
- Not something you might get from the synopsis
- Location = where it’s set, not where it’s shot
- The show might appear in “TV shows set in Washington DC” or in “Political TV Dramas set in Washington DC”
- The great part about “TV shows set in Washington DC” is that it can combine dramas and comedy – things you wouldn’t think to combine, but would actually provide recommendations you really enjoy.
Personalized recommendations are only partly about the metadata. They’re also about the business rules.
Context is everything!
[had to leave early, I’m sure the rest was great!]