I was super impressed on the recommendations from FB watch where I just kept it auto-playing the next and the next video. Their tech isn't as good, their videos are choppier than youtube and netflix, but their recommendations were insanely good.
I've never used FB Watch, but based off this comment I decided to go and see what it played for me.
Video 1: Sheet Pan Chicken Fajita Crunchwrap (recipe video).
My comment: I basically never cook and I would never watch a cooking video.
Video 2: "THE NUN PRANK" (some young idiots doing a prank in an underground parking lot).
My comment: I find this totally moronic.
Video 3: Banana Bread on a stick (another recipe video).
My comment: Zero interest in this.
Video 4: Brave Boy Kills Giant Tiger (looks like a trailer for a cheesy Bollywood kind of movie)
My comment: I don't watch this kind of thing at all and have zero interest.
Video 5: "Lilwin" (looks like an amateur video set in Africa with two people speaking an unknown language)
My comment: I have no fucking clue why this was recommended to me.
Are you kidding? Are you somehow suggesting Facebook shouldn't be able to know anything about me given the amount of data they've sucked up over the last decade?
Given I first opened my Facebook account 10 years ago and I've been posting status update and liking various bands and TV shows, I would expect it to be able to have a fair guess of things I might like.
It's recommending videos that aren't even in ENGLISH! I mean, I have "liked" TV shows on Facebook, artists/bands, I've posted about books I've enjoyed, movies I've liked. It knows my age, gender, location, friends, job, etc. You'd think if they were as good at recommending as some people make out they could do a decent job.
There isn't a single piece of content in my top 20 videos I'd want to watch.
It's a joke. If this is the state of their AI recommending tech they need to fire their data scientists.
I like how Sabrience got downvoted on first comment then made a super comeback with shockingly good insights.
It seems smart people expect others to read between the lines but when average people can't read between lines, same average people resort to downvoted then finally once smart person explains himself, they take his side
That doesn't sound like the argument at all. I read it as, given the vast amount of data they have, the initial suggestions should have been closer to what they might enjoy watching.
I wonder how much of that has to do with how much FB knows about you and your interests, and your friends and their interests? With YouTube, I honestly don't get any sense that it uses what Google surely knows of me via Search and GMail, e.g. anything related to Chicago, or germane to my political inclinations or personality. Instead, it seems wholly based on my viewing history. Which has some benefits, of course (I get what I click on), except when it makes inferences on outlier activity. e.g. if I click-through to a flat earth conspiracy video for a few curious laughs, it naively thinks I must have an appetite for more.
> e.g. if I click-through to a flat earth conspiracy video for a few curious laughs, it naively thinks I must have an appetite for more.
Every time I'm about to click on a youtube link / watch an embedded video I ask myself if it's worth influencing my future recommendations - which are not great, but not too far off either. If the content is outside my core interests or music tastes I opt to watch in incognito mode.
I once made the mistake to play Peppa the Pig for my daughter while signed in, and my recommendations became a mess for weeks.
I tried last.fm years ago and had to give up. I was on my phone and read about a Swedish prog rock band, thought about how I could get to hear them and signed up for last.fm there and then. The service used that seed to give me a really fun scandanavian hippy rock experience exploring related bands from Finland, Norway and Denmark. It was great for a while, but then I was trapped. No matter what I did to push the service towards a more balanced assessment of my musical tastes, it kept casting me back into nordic psychdedelia hell.
> Every time I'm about to click on a youtube link / watch an embedded video I ask myself if it's worth influencing my future recommendations
I do this all the time. Often, I click on "smart" videos, which I have no intention of watching just in the hope I get better quality content in my "feed". But it's hard to find high quality content, yet it's easy to find mindless dribble.
> it's hard to find high quality content, yet it's easy to find mindless dribble.
+1 to this. I watched a few Super Smash Bros Ultimate compilations at around the time it was released, and got some recommendations. I don't own a Switch so I don't know the game, plus its a novelty, so they caught my curiosity.
I went down a such rabbit-hole of mindless dribble --endless videos of videogame characters punching each other-- I actually tired myself and am no longer considering buying the game/console. It's such mindless consumption.
> Often, I click on "smart" videos, which I have no intention of watching just in the hope I get better quality content in my "feed".
It's crazy to me how Google almost telepathically guesses what I will search for in all of the web, but it doesn't know what videos I will watch.
YouTube's recommendation algorithm is one of the worst I use. I have to actively manage my subscriptions and many people I know mass click "not interested" on videos to try to improve it. Spotify, Amazon, Netflix, and pretty much every other recommendation works like magic by now.
Google Search, Amazon, Netflix-- my guess is there's just have more structured data for these recommendation systems. Amazon and Netflix get high-quality descriptions and information tables about their videos/products, Google Search has to do a bit more digging, but it has the whole text corpus link structure, etc of a site.
Spotify probably has it a bit tougher, but my guess is playlists play a big role as they are curated clustering of similar music (as opposed to just a "folder" to put all videos in a series, like they are on YouTube). Music itself is also probably much easier to analyze than video: extract the BPM and frequencies and you can tell a lot about the composition of type of song it's from. Video on the other hand-- you can do very computationally intensive CV, but that won't tell you much about the mood, meaning, genre, etc. of video something is.
YouTube many times will have to go off a title, a very vague description and a ton of irrelevant comments.
Edit: Actually, they have captions. I'd assume dialogue could give them a lot of relevant information, who knows ¯\_(ツ)_/¯
I had a hilarious incident of this with Google Search being on the money and YouTube being so far off the mark it appeared broken. Try searching for "GPL3" and I assume you'll see something similar.
On the other hand, searching for "GPL v3" actually gets relevant YouTube results, so they are out there and YouTube knows how to find them if you ask the right way.
I've been using FB for more than 10 years and the Facebook Watch recommended videos are absolutely terrible and of no interest to me. So they're not doing a particularly great job as far as I'm concerned.
Ooh that's interesting. I'm not surprised by how great their recommendations are; Facebook's AI research team is pretty top notch. I wonder what they're doing for recommendations.
Yeah the recommendations are absolutely horrible for me. They must know so much about me, yet two of the top ten videos were in foreign languages (which I do not speak). There wasn't one video in the top 20 that I would want to watch.