36 ms·

Instagram’s Explore Recommender System

109 points by YoavShapira 7y ago

stuff4ben 7y ago
Is this really AI? This seems like simple classification and ranking. Honestly I didn't see anything new in there that hasn't been around for the past 10 years. KNN? NDCG? That's entry level ML. TFA does throw around neural networks a bit, but doesn't go into any detail. EDIT: Maybe I'm just thrown off by the "Powered by AI" part of the article title. I was expecting more I suppose.
- brenden2 7y ago
  Anything moderately related to software and data is called AI now, because that's what gets the upvotes.
  - thosakwe 7y ago
    AFAIK classification, deep learning, ML, etc. all are valid subsets of AI
    - brenden2 7y ago
      It's more like complex linear regressions. The term "AI" evokes the idea of computers possessing something similar to human intelligence which it very much is not. We don't even properly understand how the human brain works.
      - brootstrap 7y ago
        Yeah i am totally with you. AI can mean linear regressions and often does. "look Ma i'm writing my own AI algorithm!!!" ... "writes linear regression by hand in python"
      - baddox 7y ago
        There’s no need for AI to actually work the same way that the human brain works, and that’s never been part of the definition of “AI.”
  - bigbadgoose 7y ago
    During a recent discussion on this topic, this emerged with regard to "what is AI": A. System should, without prompting, identify areas of improvement and innovation B. Automated collection of data and the processing thereof, combined with application towards a concrete goal — does not qualify under A. C. Part of A. is willing and unwilling discovery and exposure to both benevolent and adversarial environments and operating conditions
    - brenden2 7y ago
      This is just moving the goalposts around so that you can take advantage of the hype boost you get from calling everything "AI".
- confeit 7y ago
  Just read it as "Powered by big data" or "Powered by software" or "Powered by mathematics". It sounds fake and dumb. If it sounds fake and dumb to you, a ML practicioner, then maybe this article was not written for you. If you read the article itself, you get the feeling that AI was tacked on at the last minute, likely with an intent that has nothing to do with knowledge sharing.
- CabSauce 7y ago
  Embeddings have only really been popularized in the last few years. The 'original' word2vec paper was 2013. It's really only been maybe two years that using embeddings on other things has been popularized.
  - tanilama 7y ago
    http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf Bengio has a paper in 2003 that describes almost the same idea as word2vec (CBOW model to be exact).
- baddox 7y ago
  Why would something need to use techniques less than 10 years old to qualify as being AI? The field of artificial intelligence has been around for well over 50 years. Long before the most recent boom in AI research and marketing, AI has been a serious academic field and has been taught as part of a standard computer science curriculum. Very basic techniques like A* search are widely recognized as being AI techniques. There’s really no point in having a semantic argument, but just know that if you wish to do so you are arguing against many decades of wide usage of the term.
- huac 7y ago
  the concepts are simple but doing it at scale is hard. there are some good insights on scaling here: * scaling to more engineers/products: IGQL is an interesting way to compose ML pipelines with straightforward syntax * scaling to more ranking candidates: an active user with a large follow graph who loads the explore tab likely has millions of eligible candidates - how do you load those fast? the idea of using a "distilled" model as a first, light ranking before using a full model as the final predictor is a good intuitive idea that I haven't seen described before. * scaling KNN is hard: FB has done interesting work to make approximate nearest neighbor search fast, and opensourced it (the FAISS library which is referred to in the post). the improvements here are certainly non-trivial. * scaling to more users: creating useful general purpose user embeddings is hard! * scaling to more objectives: instagram has many business objectives, e.g. likes, follows, minimizing hides, so there is a need to have multiple models making many predictions. There is also a need to weight them intelligently, which is where the Bayesian optimization libraries come in. in some sense, nothing is truly AI, but this is useful work which you can learn a lot from.
- tanilama 7y ago
  True. It sets rightfully under the engineering category, because all this does is just engineering. At scale is challenging, but with Facebook's resources, it is more like problem waiting to be tackled, but the stakes are not high, because it CAN be tackled, just how much resources you want to pour into. What makes this mildly interesting is the IGQL, but again, without knowing the full syntax, it feels pretty restrictive. Youtube is doing much more advanced stuff, like Reinforcement learning@Scale, as comparing to Instagram in this regards.
automatoney 7y ago
It's so interesting how powerful word embeddings (or in this case account embeddings) are. This reminds me a bit of the 538 article[1] about doing math on subreddits - I'd be interested to see what sort of math you could do on instagram account embeddings. What happens when you subtract two celebrities from each other? Also, I'm curious about the tradeoffs of revealing this information - does knowing this make it easier to game the instagram algorithm? From this article I'd think that having a more narrowly targeted account (for example someone putting selfies on one account and landscape photos on another) might make their embedding more similar to others. Another thought is that maybe someone liking a bunch of things unrelated to their content would make them wrongly appear in certain explore pages. [1] https://fivethirtyeight.com/features/dissecting-trumps-most-rabid-online-following/ https://fivethirtyeight.com/features/dissecting-trumps-most-...
gravitas 7y ago
It can go pretty wrong - look at a picture of a gun one time and IG will start shoving gun pictures down your throat, it's horribly non-forgiving and can't be tuned by the end user easily. You learn pretty quick to never look even once at something you don't want a whole lot of that same thing force-fed to you - on the other hand, start looking at huskies and you'll get tons of puppers filling your explore. :)
- wastedhours 7y ago
  That is one of the things I like about the YouTube one - the option to remove from your view history seemingly does have an effect on the recommendation engine.
  - hombre_fatal 7y ago
    > seemingly It definitely does. The homepage goes back to recommending a mix of my subscriptions and content generally related to them. I like to reset my Youtube viewing history a couple times a month to see which direction my viewing habits will take my recommendations this time. I'll often ratchet into new territory. This month it was Warcraft 3: Reforged gameplay (Grubby), a game I haven't played in 15+ years. The month before it was fiction book review channels. It's fun to change it up. In Youtube's recent homepage update this month, they also rolled out a "Don't recommend this channel anymore" button on the dropdown which is very welcome.
    - leobakerhytch 7y ago
      Thanks for this suggestion — will try this as a way to escape the ‘filter bubble’. To clarify: do you reset completely, retaining absolutely no likes, subscriptions, etc., or do you find it sufficient to clear your watched videos?
- SnowingXIV 7y ago
  Agreed, it would be preferable if you could turn it off altogether and just get a wide variety of popular posts as that seems to match the title of 'explore' more closely. I'm into lifting but my explore has turned into a flood of fit instagram models.
tempodox 7y ago
In the past, when you wrote code you didn't understand yourself, you'd call it “experimental” and maybe open-sourced it. Now, when you write code you don't understand yourself, you call it “AI”, put a patent on it and sell it for a shitload of money. That's progress.
- brootstrap 7y ago
  smart. nobody questions it. corporate spends millions making shitty ADs that basically say AI will make our farmers better and therefore help the world. I love when someone asks me about AI and i'm like yeah it's a pile of shit that nobody understands but sells like hot garbage. Now before any hardcode deep learning bros get at me, sure there are cutting edge AI applications and science stuff going on. But i'm talking about commercialized shit that sells AI like a feature. It's just a pile of poop but people gobble it up.
sbilstein 7y ago
If instagram's product KPIs were more inline with what I wanted as a user, this would be great...but they're not and my explore feed is frequently filled with models, child musical prodigies, and other popcorn-esque content. Compared that to Spotify, whose goal I presume is to get me to listen to more music and buy tickets and merch through their occasional marketing. I'm a music snob but damn does Spotify get me great recommendations on new releases, my discover weekly, and more. Not only that, I've bought tickets through their frequent listener promotions probably more than 10 times at this point.
- uoaei 7y ago
  I've had terrible luck the past couple years with Spotify's Discover Weekly. Last time I remember it being good was Fall 2016. Now my "Discover" Weekly playlist has me "discovering" the same exact songs over and over. I've been pigeonholed way beyond what I thought possible. Do other users really engage with the same 10 songs over and over and over that this is the default behavior of their recommendation engine? I get "Discover" tracks which are from the same album I have downloaded to my phone!
  - volkk 7y ago
    i have a separate problem. i got a puppy and there was a time where i had to just put in headphones while i was crate training him, and listen to rain sounds to fall asleep, instead of his cries. months later, all i get on my discover feed are a bunch of mellow rain/sleepy songs. I dont actually listen to that crap, i just want it periodically for certain things. there needs to be a way to tag songs as "DO NOT CONSIDER" for the discover feature
    - squeaky-clean 7y ago
      Similar situation here, I use spotify for music during D&D games (you can find some amazing playlists on reddit / dnd forums), but now 20% of what's in my mixes is generic fantasy and halloween background ambiance.
  - sdan 7y ago
Akababa 7y ago
I've worked at a large e-commerce company before and it's surprising how many of the basic techniques like embedding, seed accounts, round robin diversification are exactly the same. I used to wonder if some of the apparent idiosyncrasies in our system were shared by other companies. It's uncanny how much of it is industry standard.
- ticktockten 7y ago
  Yeah, i have come to a similar conclusion. It is also very clear that academic papers are unless simple enough to be understood (like word2vec) they rarely make it to production.
ticktockten 7y ago
Interesting, and looks like the present choices have evolved over time. Seems like they are moving towards a structured RL implementation. There are elements of it, a follow-up post on some components would be interesting.
38932ur98u 7y ago
Interesting engineering and all, but let's keep in mind that at the end of the day, this is, once again, some of the smartest n% of the people in the world putting their brains to work making an app as addicting as possible so people click more ads.
tanilama 7y ago
Honestly...it feels pretty standard stuff.
eggie5 7y ago
So IG ins't using collaborative filtering? The whole process starts w/ simple NN search in the account embedding space. Those candidates are then passed to the ranking stack. This makes sense w/ what I see in IG recs: past behavior is strongly reinforced w/ littler diversity. Filter Bubble/Pigeon Hole problem. So in conclusion, I would argue that the IG explore tab doesn't have ANY explore at all!

109 points by YoavShapira 7y ago

Is this really AI? This seems like simple classification and ranking. Honestly I didn't see anything new in there that hasn't been around for the past 10 years. KNN? NDCG? That's entry level ML. TFA does throw around neural networks a bit, but doesn't go into any detail. EDIT: Maybe I'm just thrown off by the "Powered by AI" part of the article title. I was expecting more I suppose.

brenden2 7y ago
Anything moderately related to software and data is called AI now, because that's what gets the upvotes.
- thosakwe 7y ago
  AFAIK classification, deep learning, ML, etc. all are valid subsets of AI
  - brenden2 7y ago
    It's more like complex linear regressions. The term "AI" evokes the idea of computers possessing something similar to human intelligence which it very much is not. We don't even properly understand how the human brain works.
    - brootstrap 7y ago
      Yeah i am totally with you. AI can mean linear regressions and often does. "look Ma i'm writing my own AI algorithm!!!" ... "writes linear regression by hand in python"
    - baddox 7y ago
      There’s no need for AI to actually work the same way that the human brain works, and that’s never been part of the definition of “AI.”
- bigbadgoose 7y ago
  During a recent discussion on this topic, this emerged with regard to "what is AI": A. System should, without prompting, identify areas of improvement and innovation B. Automated collection of data and the processing thereof, combined with application towards a concrete goal — does not qualify under A. C. Part of A. is willing and unwilling discovery and exposure to both benevolent and adversarial environments and operating conditions
  - brenden2 7y ago
    This is just moving the goalposts around so that you can take advantage of the hype boost you get from calling everything "AI".
confeit 7y ago
Just read it as "Powered by big data" or "Powered by software" or "Powered by mathematics". It sounds fake and dumb. If it sounds fake and dumb to you, a ML practicioner, then maybe this article was not written for you. If you read the article itself, you get the feeling that AI was tacked on at the last minute, likely with an intent that has nothing to do with knowledge sharing.
CabSauce 7y ago
Embeddings have only really been popularized in the last few years. The 'original' word2vec paper was 2013. It's really only been maybe two years that using embeddings on other things has been popularized.
- tanilama 7y ago
  http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf Bengio has a paper in 2003 that describes almost the same idea as word2vec (CBOW model to be exact).
baddox 7y ago
Why would something need to use techniques less than 10 years old to qualify as being AI? The field of artificial intelligence has been around for well over 50 years. Long before the most recent boom in AI research and marketing, AI has been a serious academic field and has been taught as part of a standard computer science curriculum. Very basic techniques like A* search are widely recognized as being AI techniques. There’s really no point in having a semantic argument, but just know that if you wish to do so you are arguing against many decades of wide usage of the term.
huac 7y ago
the concepts are simple but doing it at scale is hard. there are some good insights on scaling here: * scaling to more engineers/products: IGQL is an interesting way to compose ML pipelines with straightforward syntax * scaling to more ranking candidates: an active user with a large follow graph who loads the explore tab likely has millions of eligible candidates - how do you load those fast? the idea of using a "distilled" model as a first, light ranking before using a full model as the final predictor is a good intuitive idea that I haven't seen described before. * scaling KNN is hard: FB has done interesting work to make approximate nearest neighbor search fast, and opensourced it (the FAISS library which is referred to in the post). the improvements here are certainly non-trivial. * scaling to more users: creating useful general purpose user embeddings is hard! * scaling to more objectives: instagram has many business objectives, e.g. likes, follows, minimizing hides, so there is a need to have multiple models making many predictions. There is also a need to weight them intelligently, which is where the Bayesian optimization libraries come in. in some sense, nothing is truly AI, but this is useful work which you can learn a lot from.
tanilama 7y ago
True. It sets rightfully under the engineering category, because all this does is just engineering. At scale is challenging, but with Facebook's resources, it is more like problem waiting to be tackled, but the stakes are not high, because it CAN be tackled, just how much resources you want to pour into. What makes this mildly interesting is the IGQL, but again, without knowing the full syntax, it feels pretty restrictive. Youtube is doing much more advanced stuff, like Reinforcement learning@Scale, as comparing to Instagram in this regards.

automatoney 7y ago

It's so interesting how powerful word embeddings (or in this case account embeddings) are. This reminds me a bit of the 538 article[1] about doing math on subreddits - I'd be interested to see what sort of math you could do on instagram account embeddings. What happens when you subtract two celebrities from each other? Also, I'm curious about the tradeoffs of revealing this information - does knowing this make it easier to game the instagram algorithm? From this article I'd think that having a more narrowly targeted account (for example someone putting selfies on one account and landscape photos on another) might make their embedding more similar to others. Another thought is that maybe someone liking a bunch of things unrelated to their content would make them wrongly appear in certain explore pages. [1] https://fivethirtyeight.com/features/dissecting-trumps-most-rabid-online-following/ https://fivethirtyeight.com/features/dissecting-trumps-most-...

gravitas 7y ago

It can go pretty wrong - look at a picture of a gun one time and IG will start shoving gun pictures down your throat, it's horribly non-forgiving and can't be tuned by the end user easily. You learn pretty quick to never look even once at something you don't want a whole lot of that same thing force-fed to you - on the other hand, start looking at huskies and you'll get tons of puppers filling your explore. :)

wastedhours 7y ago
That is one of the things I like about the YouTube one - the option to remove from your view history seemingly does have an effect on the recommendation engine.
- hombre_fatal 7y ago
  > seemingly It definitely does. The homepage goes back to recommending a mix of my subscriptions and content generally related to them. I like to reset my Youtube viewing history a couple times a month to see which direction my viewing habits will take my recommendations this time. I'll often ratchet into new territory. This month it was Warcraft 3: Reforged gameplay (Grubby), a game I haven't played in 15+ years. The month before it was fiction book review channels. It's fun to change it up. In Youtube's recent homepage update this month, they also rolled out a "Don't recommend this channel anymore" button on the dropdown which is very welcome.
  - leobakerhytch 7y ago
    Thanks for this suggestion — will try this as a way to escape the ‘filter bubble’. To clarify: do you reset completely, retaining absolutely no likes, subscriptions, etc., or do you find it sufficient to clear your watched videos?
SnowingXIV 7y ago
Agreed, it would be preferable if you could turn it off altogether and just get a wide variety of popular posts as that seems to match the title of 'explore' more closely. I'm into lifting but my explore has turned into a flood of fit instagram models.

tempodox 7y ago

In the past, when you wrote code you didn't understand yourself, you'd call it “experimental” and maybe open-sourced it. Now, when you write code you don't understand yourself, you call it “AI”, put a patent on it and sell it for a shitload of money. That's progress.

brootstrap 7y ago
smart. nobody questions it. corporate spends millions making shitty ADs that basically say AI will make our farmers better and therefore help the world. I love when someone asks me about AI and i'm like yeah it's a pile of shit that nobody understands but sells like hot garbage. Now before any hardcode deep learning bros get at me, sure there are cutting edge AI applications and science stuff going on. But i'm talking about commercialized shit that sells AI like a feature. It's just a pile of poop but people gobble it up.

sbilstein 7y ago

If instagram's product KPIs were more inline with what I wanted as a user, this would be great...but they're not and my explore feed is frequently filled with models, child musical prodigies, and other popcorn-esque content. Compared that to Spotify, whose goal I presume is to get me to listen to more music and buy tickets and merch through their occasional marketing. I'm a music snob but damn does Spotify get me great recommendations on new releases, my discover weekly, and more. Not only that, I've bought tickets through their frequent listener promotions probably more than 10 times at this point.

uoaei 7y ago
I've had terrible luck the past couple years with Spotify's Discover Weekly. Last time I remember it being good was Fall 2016. Now my "Discover" Weekly playlist has me "discovering" the same exact songs over and over. I've been pigeonholed way beyond what I thought possible. Do other users really engage with the same 10 songs over and over and over that this is the default behavior of their recommendation engine? I get "Discover" tracks which are from the same album I have downloaded to my phone!
- volkk 7y ago
  i have a separate problem. i got a puppy and there was a time where i had to just put in headphones while i was crate training him, and listen to rain sounds to fall asleep, instead of his cries. months later, all i get on my discover feed are a bunch of mellow rain/sleepy songs. I dont actually listen to that crap, i just want it periodically for certain things. there needs to be a way to tag songs as "DO NOT CONSIDER" for the discover feature
  - squeaky-clean 7y ago
    Similar situation here, I use spotify for music during D&D games (you can find some amazing playlists on reddit / dnd forums), but now 20% of what's in my mixes is generic fantasy and halloween background ambiance.
- sdan 7y ago

Akababa 7y ago

I've worked at a large e-commerce company before and it's surprising how many of the basic techniques like embedding, seed accounts, round robin diversification are exactly the same. I used to wonder if some of the apparent idiosyncrasies in our system were shared by other companies. It's uncanny how much of it is industry standard.

ticktockten 7y ago
Yeah, i have come to a similar conclusion. It is also very clear that academic papers are unless simple enough to be understood (like word2vec) they rarely make it to production.

ticktockten 7y ago

Interesting, and looks like the present choices have evolved over time. Seems like they are moving towards a structured RL implementation. There are elements of it, a follow-up post on some components would be interesting.

38932ur98u 7y ago

Interesting engineering and all, but let's keep in mind that at the end of the day, this is, once again, some of the smartest n% of the people in the world putting their brains to work making an app as addicting as possible so people click more ads.

tanilama 7y ago

Honestly...it feels pretty standard stuff.

eggie5 7y ago

So IG ins't using collaborative filtering? The whole process starts w/ simple NN search in the account embedding space. Those candidates are then passed to the ranking stack. This makes sense w/ what I see in IG recs: past behavior is strongly reinforced w/ littler diversity. Filter Bubble/Pigeon Hole problem. So in conclusion, I would argue that the IG explore tab doesn't have ANY explore at all!