Instagram’s Explore Recommender System • Hacker News Board

Re: Instagram’s Explore Recommender System

#21

Post by pornel » Tue, Nov 26, 2019, 4:07 PM UTC

raverbashing wrote: post #16

gravitas wrote: post #5

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. :)

There is a "see fewer posts like this" button (as mentioned in the text) But yeah it can be off-putting.

This option is not working for me at all.

I wonder if it's because if I see a thumbnail of something unwanted, I first have to open the post in order to access the menu, so the the app counts it as a view first, adding the unwanted post to things I "engage" with.

Re: Instagram’s Explore Recommender System

#22

Post by CabSauce » Tue, Nov 26, 2019, 4:39 PM UTC

stuff4ben wrote: post #2

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.

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.

Re: Instagram’s Explore Recommender System

#23

Post by nightski » Tue, Nov 26, 2019, 5:00 PM UTC

gravitas wrote: post #5

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. :)

This is so amusing to me. I imagine their team goes to great lengths to obtain training data, clean it up and feed it to their ML algorithm.

Yet the most critical input data - direct feedback from the user on the traits that shape their interests, is completely ignored/not collected.

Re: Instagram’s Explore Recommender System

#24

Post by wsinks » Tue, Nov 26, 2019, 5:05 PM UTC

gravitas wrote: post #5

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. :)

I got hacked at one point, and I really should restart my account it's so bad. Every time I pull it up it's just big butts...

Re: Instagram’s Explore Recommender System

#25

Post by PowerfulWizard » Tue, Nov 26, 2019, 5:13 PM UTC

raverbashing wrote: post #16

gravitas wrote: post #5

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. :)

There is a "see fewer posts like this" button (as mentioned in the text) But yeah it can be off-putting.

I think it is driven by likes and bookmarks, if you browse your likes and bookmarks you might see the source of certain recommendations, and unliking and unmarking should control it. (I don't actually know, but this is how I believe it works.)

Re: Instagram’s Explore Recommender System

#26

Post by baddox » Tue, Nov 26, 2019, 5:27 PM UTC

nightski wrote: post #23

gravitas wrote: post #5

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. :)

This is so amusing to me. I imagine their team goes to great lengths to obtain training data, clean it up and feed it to their ML algorithm. Yet the most critical input data - direct feedback from the user on the traits that shape their interests, is completely ignored/not collected.

Presumably IG is optimizing for something different than what you as a user would be optimizing for. Possible candidates would be “engagement” or sad impressions or clicks. As a user you probably don’t care to maximize how many ads you see or even how much you use the app. You just want to see things that interest you. And of course there’s absolutely no reason why IG would optimize for that.

Re: Instagram’s Explore Recommender System

#27

Post by baddox » Tue, Nov 26, 2019, 5:31 PM UTC

brenden2 wrote: post #10

thosakwe wrote: post #8

Earlier quoted context omitted.

AFAIK classification, deep learning, ML, etc. all are valid subsets of AI

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.

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.”

Re: Instagram’s Explore Recommender System

#28

Post by baddox » Tue, Nov 26, 2019, 5:36 PM UTC

stuff4ben wrote: post #2

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.

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.

Re: Instagram’s Explore Recommender System

#29

Post by sbilstein » Tue, Nov 26, 2019, 5:39 PM UTC

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.

Re: Instagram’s Explore Recommender System

#30

Post by nightski » Tue, Nov 26, 2019, 5:44 PM UTC

baddox wrote: post #26

nightski wrote: post #23

Earlier quoted context omitted.

This is so amusing to me. I imagine their team goes to great lengths to obtain training data, clean it up and feed it to their ML algorithm. Yet the most critical input data - direct feedback from the user on the traits that shape their interests, is completely ignored/not collected.

Presumably IG is optimizing for something different than what you as a user would be optimizing for. Possible candidates would be “engagement” or sad impressions or clicks. As a user you probably don’t care to maximize how many ads you see or even how much you use the app. You just want to see things that interest you. And of course there’s absolutely no reason why IG would optimize for that.

You may or may not be right. I don't use instagram because of this. But how could they possibly know how relevant the data is to even their model since they do not even collect that data in the first place?