This is how Netflix's top-secret recommendation system works

[smoll]

The idea that you can group everyone into 2k+ cohorts of differing tastes sounds like a good way of attacking the problem (though it feels a bit like spaghetti sauce taken to the extreme https://ed.ted.com/lessons/malcolm-gladwell-on-spaghetti-sau...). Personally, though, I think IMDb score is way more predictive of my enjoyment than their recommendation engine has ever been. Nine times out of ten watching a show that is below an 8.0 is a complete waste of time for me - I'll hate it even if it comes up as a 99% match for me. Conversely, I'll watch a show totally out of my wheelhouse that I know nothing about (cf. The Grand Tour) if it's good television. I figured there would be a cohort of people who watch good things of any type but maybe not - maybe there is, I just haven't been placed in it for whatever reason.

[jorvi]

It always seems like pretty much every IMDB score is between 7-8. This makes it essentially useless, because if everything is good, nothing is good. I've experienced that RottenTomatoes is a far better indicator, especially the 'Audience score'.

[s_kilk]

Except it mostly doesn't work.

My dashboard has recommended Minions "because you watched The Thick Of It". Out of the whole list, only Yes Prime Minister comes close, and even at that it misses the (more) obvious Yes Minister series.

[jetru]

I find the Netflix recommendations largely non-beneficial considering their lack of selection. Most of the movies/series I want to watch are not on Netflix at all.

Moreover - I think the internal algorithm for training their models will be to maximize "continuations" of watching series enabling binge-watching. While this may be a good measure of engagement for THEM, it hardly is for ME. I would much rather that the choice was made out of the global list of "good stuff I should be watching" versus the "list of stuff Netflix sales people could buy and negotiate for".

I do think it's cool that when you type in a title that's NOT in their system - they can recognize it and recommend similar titles - although - it's a very good bet that actually clicking and watching one of those is highly likely going to a waste of time.

[mayank]

> I think the internal algorithm for training their models will be to maximize "continuations" of watching series enabling binge-watching. While this may be a good measure of engagement for THEM, it hardly is for ME.

This seem inconsistent -- are you saying that you binge watch shows that you don't like?

There's also a well-known gap between what people say they want to watch vs what they actually watch, they should absolutely be optimizing for the latter.

[skue]

It's a shame they didn't bother to create a category for "Resists our suggestions and prefers to manage their own view list."

Every time I've tried using Netflix, I've canceled in frustration. Their movie selection is awful, and it baffles me that you have to hunt for your own curated watch list among rows and rows of uninspired TV shows they insist you might enjoy.

[krzat]

It's amusing how hostile they are to power users.

[homulilly]

I must not fit into one of their "taste groups" because I've found they're suggestions to be worse than useless.

From the article it sounds like they're more concerned with trying to get people turned on too long running original series cash cows than showing them good suggestions anyway. their Black Mirror = Luke Cage suggestion example is laughable.

[merrua]

They badly need to let their users override their system. I want to be able to mark something hidden and not see it again.

[petters]

The takeaway here is that the majority of useful data is implicit -- very different from the Netflix challenge ten years ago.

[username3]

Can the system tell when we're sick of a genre and clones?

[ENGNR]

TLDR; machine learning