Eventually, the algo is going to figure out I don’t like Mark Ruffalo – but it’s still too dumb

If I have to pick my least favorite Avenger actor, it’d be him, though the other contenders in his role are hard to live up to (I was a Bill/Lou fan). But generally, if you want me to watch a movie, don’t put Ruffalo on the cover card.

I spend a lot of time thinking about algos and machine intelligence modeling (the pretty little thief machines that current “AI” represents) and thinking through how companies use them. It really puzzled me when Netflix nuked the five star Likert scale for a simple “thumbs up / thumbs down” system. Where would they get useful, measurable data? Maybe five stars were too noisy data wise. Maybe they preferred another way of figuring out what we wanted to watch.

I planned a date night a few weeks ago with my kids: The Barbie Movie. They’d seen it in theaters, and in preparation (no spoilers) I went through Netflix and started watching Ryan Reynolds movies. It’d let me compare his range and what direction he took Ken. Turns out I picked the wrong Ryan. Oh, well.

From there I got into some discussion of movies that had me looking for Ryan Reynolds and Ryan Gosling movies on Netflix, and I started getting really weird results. Turns out if you search for an actor, they’ll show you the movies, and a whole bunch of other random movies you might like … but not tell you the actor you want only has, say, three movies on their platform right now.

So their algo is weighted for “similar” movies based on the recommendations of their users, presumably on the number of times users have watched movies “together”, or something else over and over. It makes me wonder how the double thumbs up works. I guess we’ll see.

I repeated the “search by actor” test with Ruffalo. The movie I just watched appeared in position one, positions two three and four were held by Kevin Bacon connections back to top billed actors in the movie I just watched, then positions five and eight were Ruffalo with connections in between in the gooey centre.

Which is great. Most people on Netflix are probably browsers, not people like me who want what they want and not something “similar”.

But back to Ruffalo. I don’t avoid him like I do, say, Woody Allen (and why can’t we block content by people, Netflix?). But if I see his face on a card, I’ll keep cruising by. Held up by a solid ensemble he’s not bad, but alone, pretty meh, not my cup of tea.

Netflix does more than just mix up suggestions — they make their own browsing cards. When I finally saw a new iteration of the cover/browsing card for The Adam Project featuring Ryan Reynolds instead of Mark Ruffalo, I figured maybe I’d seen it wrong? Maybe Ruffalo had been a minor character and they thought his star power would bring in more viewers? I’d seen it right, and he took up minor space in the movie, heavily supported by the child actor and Reynolds. So, a little from column A, a little from column B.

An interesting way to measure response with A/B(/C?/D?) testing for browsers. More data can make it a little bit smarter, but I’d still like a super user mode to block out or highlight specific performers/directors or themes … Heck, they could partner with PR firms to sell Hollywood peeps info on who likes them that much (or dislikes, if they care) …

The algos are getting smarter, but it’s because the people are wielding the big data they’re collecting more delicately. Don’t go taking a five figure “crash course” in AI or ML to “quit your day job” and enter the lucrative world of big data. Tweaks such as varied cover cards and Bacon connections on recommendations aren’t covered there.