If You Liked This, You're Sure to Love That (2008) | Hacker News Reader

If You Liked This, You're Sure to Love That (2008)

nytimes.com

26 pointsmpiedrav21 comments

Just based on the experiences I have with them, I feel like the Big Problem with recommendation engines is that they're bad at teasing out why somebody liked something and end up falling back on an obfuscated equivalent of P(likes A | likes B), which ends up dominated by people obsessed with a particular genre, franchise, or theme. Or worse, it relies on extremely generic tags. If you buy, say, Steins;Gate and Rosenkreuzstilette on Steam, it's going to keep recommending the most garbage VNs and PG-13 nukige on the planet until the sun explodes because you obviously like VNs and anime (as opposed to quirky sci-fi stories and Mega Man).

There's a paragraph about this in the article that discusses automated identification of clusters, but it makes it sound like there was a lot of human judgment/guesswork in figuring out why any given set of movies clusters the way it does.

> Each new algorithm takes on average three or four hours to churn through the data on the family’s “quad core” Gateway computer.

Huh. I don't think I would have guessed that Gateway was still around in the days of quad-core desktops.

Also why is quad core in quotes lol. Is it really like four old Pentium IIIs in a compute cluster?

The article is from 2008. Multicore was still pretty novel at the time.

The Core2 Quad series released early 2007. I'm not 100% that was the first Intel consumer quad chip (memory is pretty fuzzy by now) but it was close.

Could this be solved by weighting things that are generally popular less while learning? E.g. if you like a super popular Anime, say DragonBall, it doesn't mean that you necessarily like Anime. But if you like some super niche Anime it's more likely that you would enjoy more of that. Because why would you have found and enjoyed that without some specific interest?

What we need is to monitor people's brains in some way while they are watching entertainment to see specifically what gets evoked and when. I would do it if I had the means. Then we can create original stories again, and run them through our models of human amusement and catharsis to know who will enjoy what and why and when.

I remember briefly reading about how Pandora classified music -- by getting people to analyze each song for different qualities (e.g. fast paced, melancholy, etc. etc.), and that was a key part of their "recommendation engine."

You could ask people "What did you like about this movie?" and they could tag various things they liked about it, e.g. "was smart," "cheesy humor" or "had Jennifer Lawrence."

Heck, that sounds like something that might be worth building, a tag-based movie recommendation engine.

No need for all that. There are no original stories, only variations on the ancient ones. Here's one summary:

https://en.m.wikipedia.org/wiki/The_Thirty-Six_Dramatic_Situ...

As for who will enjoy what etc., there are entire industries that spend their efforts on knowing that, like Nielsen and Rotten Tomatoes.

IIRC it was further simplified to just combinations of rise and fall of characters.

These are extremely broad categories. It's a bit like saying, "there are no original books, they're all just collections of the same 26 letters", or "all stories are the same, they're just characters doing things."

One category listed here is Pursuit:

> the fugitive flees punishment for a misunderstood conflict. Example: Les Misérables

By this logic, 'Les Miserables', 'Mission Impossible', and 'Catch Me if You Can' are all just rehashing the same basic idea. On an existential level, sure, but is there any practical use to a category that broad?

Yeah. I avoid rating anything unless my feelings about the movie generally extend to the entire genre.

Like 'Shrek'? Enjoy recommendations for kids movies for the next 3 months.

Like 'Before Sunrise'? You're going to love The Tall Girl!

Seems like any one data point would be worthless. If I like "Sleepless in Seattle" (which I didn't) it might be because of Romcom (no) or because of Ryan (no) or because of Hanks (yes). I wonder if there's a cluster of 10 movies, which if you knew the answer about all 10, it would predict how you would feel about a specific #11.

Yeah that was my thought too. Like 'signpost' movies that split the audience in 2 and would mean with a few key movies you could make predictions for everyone.

This! This is what drives me bananas as an Apple Music user: you can create radio stations based on a single album, song, or artist, but not on a playlist. I have come to think of it as the "Vince Guaraldi" problem, which is that if you tell Apple Music to create a radio station based on Vince Guaraldi it'll very quickly veer off into soundtracks from children's films (because of the Peanuts soundtracks) rather than what I want, which is piano jazz trios. This could also be because Apple relies on metadata (soundtracks, family entertainment) more than the attributes of the music (which instruments are being played, tempo) but it feels like a problem that could be solved by allowing radio stations based on playlists and it boggles my mind when each new release of Apple Music doesn't include that feature.

Google Play has this feature and it's great, I'll often put a group of songs I like into a playlist and start radio from it. Best way to "discover" new music I might like, and far better than their curated radio stations.

Spotify blows Apple Music out of the water on these kinds of reccomendations, the playlist radios and user discovery playlists are pretty impressive.

I was involved with design and planning for a solution to predict unscheduled maintenance and warranty claims for vehicles. I think that we had a pretty similar problem. Unfortunately I did not stay with the company I was working for, so I can't speak to the effectiveness of the approach. A simplified example is below.

There is tons of scheduled maintenance. And there are also lots of common unscheduled maintenance items. These were basically worthless in any naive prediction. Basically "Ah, I see that an engine failure is always preceded by dozens of oil changes." There's a tiny predictive value, because number of oil changes is a proxy for age/use; older engines are more likely to fail.

But that sort of prediction wasn't helpful. We ended up proposing a strategy that I shorthand to "without which, not". Basically we wanted to classify the events which precede a specific type of failure, but which we don't see in similar vehicles without the failure. This is grossly oversimplified.

An analog would be the idea of necessary vs sufficient pre-conditions. We were trying to identify the sufficient pre-conditions.

To draw the analogy to a recommendation engine, a common failure mode is that popular things get recommended a lot. Another failure mode is that they basically become genre filters. This is discussed elsewhere.

So in a movie context, we wouldn't be looking for "what do other people who like this movie watch"? Because that's a genre popularity contest. We would ask "Given that a user likes Movie A, what did they watch before that other Movie A likers didn't watch?" And then look for clusters there.

Again, I'm grossly oversimplifying this, because I didn't get to implement the solution I mentioned. The approach seemed promising, though. You might consider this a nuance on "down-weighting popular things", but it's not quite that.

If I remember correctly, there was a winner for this but then Netflix just ended up scrapping the whole contest/project and creating a new recommendation engine.

They paid the winners and didn't use the entry I think. Which is a shame, because their engine BEFORE the "winning" one, which was presumably better, was great. Whatever they're using now is subjectively much worse.

Could be I'm misremembering, but if I am I think I'm in good company.

Do they actually use an engine now, or just throw up a random mish-mash of Netflix originals all over the screen? Most of them are in genres I've never watched with actors I've never expressed interest in, and yet there they are. "Netflix Original" is not a category that you should be using to link other movies to me Netflix.

I heard a third-hand rumor that when Netflix switched from 1-5 star ratings to thumbs up/down, they stopped even using your ratings to personalize recommendations. They were simply using watch time. I've since read someone else on HN say that is incorrect. There are old films I really like but have not streamed on Netflix. If I give them a thumbs up, it would be a pity of Netflix didn't take incorporate that data.

A big question is how much these recommendation engines have led to us surrounding ourselves with the "same old" stuff, from music to political opinions.

Reply on news.ycombinator.com