Hacker News Reader

Comment by ashtonkem | Hacker News Reader

As a computer program, this is probably not feasible. While in theory you are possibly capable of writing the algorithms for your own recommendation engine using whatever ML tools are popular today, the issue you'd have is data. YouTube's recommendation algorithm effectively only works because it's capable of aggregating millions of users behavior to feed into their own system. Getting that data personally would be very difficult both because YouTube doesn't want to enable competitors, and because consumers would be extra freaked out to find out about some random engineer outside of YT getting their watch history.

As a non-program, you can always depend on the means that humans have used for content discovery since time immemorial: word of mouth.

View on HN

This is maybe a tangent, but do these ML-powered recommendation algorithms actually, like, work well for people?

I use both YouTube and Netflix pretty extensively. These companies are both held up as ML success stories. But, at least in my personal experience, they suck! My front-page feeds consist of (1) content that the platform knows I have already seen because IT IS IN MY WATCH HISTORY, (2) content that is trivially related, like by having the the same channel/director/actors, and (3) random junk I have no interest in watching and always regret clicking on.

It's not that the content isn't out there -- I'm able to find stuff I like through "big spreadsheet of all the Netflix movies" websites, online recommendations from humans, and YouTube channel crossovers/cameos. But I've never had that magical "the algorithm knows you better than you know yourself" experience that every ML engineer I've ever spoken to seems to be convinced their systems achieve.

Are all ML engineers mindless drones with exactly average movie preferences, or am I some kind of incomprehensible weirdo?

> random junk I have no interest in watching and always regret clicking on

if you regret clicking on it, that means you did click on it and the recommendation did it's job; remember that the job of the recommendation engine is not (directly) to suggest content you'll like, it's to suggest content that will get you to click. Ideally those goals would be aligned, but not necessarily. YouTube is simply incentivised (via ad revenue) to get you to click on as many videos as possible.

At least in my case, it's not that I'm constantly clicking on things that it turns out I dislike. Rather, it's that the few times I have clicked on a recommendation (because hey, maybe they know something I don't, maybe Sonic the Hedgehog actually is a good suggestion for people who liked Eternal Sunshine of the Spotless Mind), I've regretted it.

Right, but my point is that from YouTube's perspective it doesn't "matter" if you like or dislike the video, if they can get you to click. And clearly this strategy works on some people, or they wouldn't be doing it.

So while it may not be overly successful on you (i.e. you only click occasionally), if it works on a significant portion of people it could still be a successful strategy overall.

Sometimes you click even when you don't want to. For example on youtube on ipad it's very difficult to click on the vertical three-dot hamburger icon on a video (that you need to use to tell them you're not interested in a video) without clicking on the video by mistake. The hitbox on that icon is all wrong.

So true. The thing I find bizarre about youtube is how bad not only their recommendations but their search is. This is a Google company but somehow if I search for a specific episode in a specific series (eg one episode in a game walkthrough or whatever) it can't get the ranking right?

I also still don't understand why yt recommendations are as poor as they are. I've tried taking the time to try removing certain things from the "recommendation history" (or whatever they call it), flag recommendations I'm not interested in etc and it does improve things a bit but still it recommends things I would never watch in a million years (eg right now I have 4 Joe Rogan recommendations on my homepage even though I've literally never watched any of his things and have zero interest in ever doing so, and click on every one and say not interested etc, for a while it was obsessed with trying to get me to watch MMA content even though I can't bear watching other people hurt each other).

And the whole issue of recommending videos where it knows I've watched the whole thing, I just don't even know where to start.

On search, I wonder if this is deliberate; if they gave you exactly what you want as the top result, you'd click that one and not look any further. But by having to scroll past a few other videos, maybe you'll see another one or two that you also decide to click on, resulting in more overall views.

This is a different strategy to Google web search, where the monetisation occurs on the search results page (via ads there) rather than on click-throughs, so there isn't an incentive to try and get you to click several tangentially related links.

Open Data is a thing, I can imagine some PeerTube instances would publicly share their data (anonymously etc), in order to build open recommendation systems.

Personally, I would gladly donate my (anonymous etc) data for this noble goal.

Open data + federated learning (somewhat ironically a Google invention) could help with this, but it has a long way to go, and a lot of opportunities for incumbents to fight back.

The problem is (as with all open/federated systems) spam.

How do you figure out which data has been faked to drive views to particular content?

Except that youtube’s recommendation engine It terrible and getting worse. I am on youtube all the time and all it ever recommends is things I already subscribe to. And it seems to be steering further into things I have already watched.

Data offers diminishing returns, and at sufficient quantity introduces vetting and validity problems --- see Amazon's reviews crisis or fake Yelp reviews, as examples.

Data quality and relevance matter far more.

Well just for me I don't care what's trending so don't need the feed of other people's views. But absolutely needing a feed of new posts might be problem.

That's really missing the point. Recommendation engines are a lot more dumb than you are thinking. They don't understand the content at all. They only know that thousands of users that watched the same videos as you in the past also watched this other new video, so you might like it as well.

A recommendation engine that works solely based on your own watch history would require a totally different approach using algorithms that are years/decades/forever away from existing.

Good points.

It does make sense that the job is simpler with everyone's history and inputs, if you can keep PII out of it. Are there any good references on recommendation aggregation algorithms?

Here's a simple proposal that keeps everyone's stuff anonymous and allows a rec service to pay bills.

Distribute the pool of everyone's watch histories/prefs/etc using IPFS, which is a DHT with static names and some storage. User clients put the share data into IPFS and register into a public pool. The client remembers their own hash. Rec services (there can be multiple) pull everyone's sets periodically and aggregate/indexe them. Users can then retrieve a batch of new recs by sending a small fee to the service of their choice along with their hash and the server will place some new recs into their bucket. If the recs are not suitable, they can try another service.

If I send you links to 2-3 rare videos and you view them, I can probably uniquely identify you (or restrict the list of potential users to ~5 users then recognize you based on other videos you would have watched). I then have access to your full video watching history.

Even without this, I could just check videos you shared on Twitter, follow the same process, and have a decent chance of identifying you.

Anonymization is very very hard. Maybe federated learning is more promising here, keeping all your data privately stored on your own devices.

> That's really missing the point. Recommendation engines are a lot more dumb than you are thinking. They don't understand the content at all.

I think you're the one underestimating modern recommender systems, especially the ones built by advanced teams in big tech companies.

They absolutely use content features. Among other advantages, this counters the cold start problem (how will you learn about a new video if it doesn't get recommended to anyone).

Have you ever seen the torrent of garbage that is new posts on e.g. Reddit or Imgur? You only see decent posts on the ‘front page’ because some people sift through that and filter out teenagers' low-effort attempts.

YouTube has hours of video uploaded every minute. A non-centralized, context-less recommendation system would mean that you are offered those crap videos together with ones that are the standard for ‘good’ now. Just get into the long tail on any search results and see if you want those.

A non-monopolized video discovery would be splendid, only who knows how to achieve that. Perhaps it would be more realistic to do the recommendations based on the numbers of ‘retweets’ in federated networks—in a more ‘social’ style instead of views (dunno if PeerTube already offers that). Also, an open system will be incessantly gamed.

Reply on news.ycombinator.com