Hacker News Reader

Comment by fmakunbound | Hacker News Reader

With all of this technology applied, I am still disappointed by Netflix's recommendations – to the point of just giving up and doing something else.

View on HN

Do you think part of this is that Netflix has assumed zero effort from user model? My experience has been that Netflix does an ok job of recommendations, but fails at overall discovery experience. There is no way for me to drive or view content from different angles easily. I end up googling for expert opinions or hitting up rotten tomatoes to get better reviews. Netflix knows a ton about me and their content, but seems to do a poor job of making their content browseable/discoverable overall. I do like their "more like this" feature where I can see similar titles.

Perhaps its because its a niche that isn't worth investing the resources into? Sometimes narrow problem spaces are harder for a company to justify because of the cost to reward ratio. I agree with what you want and would like it myself for music (Spotify, Amazon Music, etc.) but it's a complex problem (recommenders, custom UI and the glue between) that is hard to justify compared to incremental small improvements to existing general purpose recommendations.

It just seems like there's a paucity of signal from which Netflix could come up with anything intelligent. Movies are many hours long, and there are many reasons I could be watching something. What does it mean that I allowed a movie to play to completion? Was I even paying attention? Did I decide I hated it 3/4s of the way through, but finished it just because I cared about the plot?

TikTok, on the other hand, has way more data. Things like time-to-swipe, shares, comments presumably form the basis of some sentiment metric.

Google TV has the best content discovery I've come across so far. Recommendations across most streaming services based on overall similar movies, different slices of the genre, and movies with similar directors/cast members. Plus as soon as you select another movie, you can see all the same "similar" recommendations for that movie.

>Do you think part of this is that Netflix has assumed zero effort from user model?

Talking w/a friend who works at Netflix, it sounds like this is a warranted assumption. The way he told it, they were tearing their hair out at one point b/c users wouldn't put much into it.

What I don't understand about their response is: why not make it configurable? Admittedly this is my philosophy for almost every product I work on - "make it maximally configurable, but make the defaults maximally sane" – but I'm baffled every time I hear someone talking about this 'dilemma'.

You just keep your simple interface, but allow the power users to, say, click through to a particular menu and change their setting – the setting in this case being ~"let me provide feedback / configure how recommendations work". For that kind of user, finding a 'cheat code' is actually a gratifying product experience anyway.

I think its because the complexity of allowing configurability isn't always worth it. Verifying it works for all configurations becomes exponentially harder.

I believe it can also have performance implications especially for things like recommender systems where you are depending a lot on caching, pre computation and training.

I agree, but as aleksiy123 suggests there is an additional complexity burden and it is a long journey to teach users to make use of a new technology. I think a lot of "advanced" features get de-prioritized as not many people use them and it seems like resources could be better spent helping the masses. I think that the importance of "advanced" features is often under rated by traditional engagement models. Wikipedia is a great example of where less than 1% of users click on the edit button, but that 1% adds all the value for the other 99%.

I don't disagree!

Maybe you're just disappointed with Netflix's inventory, not their recommendations.

I think it's both. I'm usually able to find decent stuff by searching "best on netflix" with some modifiers, but I almost never find new stuff I like by scrolling on netflix.

In some ways it seems like a classic case of trying to solve the wrong problem because the wrong problem potentially has a technical solution. The real problem is making lots of interesting content for people to watch. If you can solve that problem then a simple system of categories is perfectly sufficient for people to discover content. But that’s not a technical problem, and all those engineers have to be given something to do.

This indicates that the problem is difficult to solve at scale and customized per person. Maybe the issue is with our expectations - I find other people are pretty bad at recommending things for me as well.

Maybe. Recommendation systems definitely seem to get worse as they scale. Amazon's was incredible circa 2000. Pandora seems to be getting worse and more repetitive. Netflix kept getting better and better until they ended their contest and since then they seem to have only become worse.

Rotten Tomatoes works fine as a recommendation system. It lists all of the new content coming out in a given week. I just read that every week, file down to what looks interesting based on the premise, and read a few reviews. I can usually tell pretty easily what I'll like. No need for in-app recommendations from any specific streaming service at all. Good old-fashioned human expert curators.

I was actually pretty impressed the other day when searching for "shiloh" (which they didn't have) because it showed a bunch of "related" queries to other dog movies (they also didn't have). The available search results were a little lacking though.

Reply on news.ycombinator.com