The secret sauce of TikTok’s recommendations

https://www.shaped.ai/blog/the-secret-sauce-of-tik-toks-recommendations

by DashAnimal • 3 years ago

126 138 3 years ago

PaulHoule3• 3 years ago

I like those charts showing they get an AUC of 0.8 or so, my own content-based recommender gets 0.72 or so on a typical day with a very simple model, I think it could get 0.75 with mainly data-oriented tweaks. I was thinking a better model could get me into the high 70's but the fuzzy nature of the recommender problem means I am never going to get into the 90's.

I would like to see more about how they are formulating the problem, I know there is work lately in "sequential recommendation" that is focused on generating a sequence rather than scoring content items, I'd like to learn more about that.

O__________O2• 3 years ago

AUC being “area under curve” or something else?

acomms• 3 years ago

Correct, it is a performance measure for ML algos (https://www.sciencedirect.com/science/article/abs/pii/S00313...)

PaulHoule3• 3 years ago

Yep. See https://en.wikipedia.org/wiki/Receiver_operating_characteris...

If you had to pick one number for judging classifiers that is the one.

"Accuracy" is a really bad one because it is not meaningful w/o careful thought about the problem. For instance if you have a diagnostic system for a disease that has a .01 prevalence, you get 99% accuracy if you say nobody has a disease.

Practically a person might want to operate a classifier with different thresholds depending on how bothered they are by false positives vs false negatives, the ROC shows what choices are have available with a particular and the AUC of the ROC gives you one number that characterizes the quality of the classifier overall. A system with an AUC under say 0.55 hasn't really learned anything at all, at 0.6 it is showing signs of life, looking at those curves I'd say Tik Tok is beating me solidly. In this case it is limited by the problem being fuzzy (something I thumbs up today might get a thumbs down tomorrow), something up in the 0.95 level could be attained for something like "is this article about the outcome of a sports game?"

mochomocha2• 3 years ago

the fuzzy nature of the recommender problem means I am never going to get into the 90's

Right. If your AUC is in the 90's you usually either have a bug, or a predictive task that isn't very challenging.

PaulHoule3• 3 years ago

Some tasks are hard, others are easy. I get in the low 60’s trying to predict if a headline will get a lot of votes on HN. Of course this is going to be bad because the same article can get submitted several times, hit it big once and not hit it big the other times. The content of the article matters too but the model can only make a wild-assed guess. I get in the high 70’s trying to predict if a post with a lot of hits gets a lot of comments, it is a cleaner problem. I get about 90% trying to predict if a post winds up ‘dead’, I found up HN is cleaning up a lot of spam posts when they go in but I know it is terrible overfit because it says an arXiv paper which has no chance of being flagged has a 30% chance of flagging.

For some other like “is this an astrophysics or materials science paper?”, getting 0.95 or better was straightforward in 2005.

droopyEyelids• 3 years ago

In these articles I never see reference to the 'objective algorithm' aka what the business processes at TikTok are optimizing for. I think that's the real secret sauce.

TikTok is optimizing for user generated content that connects deeply with their users. Thats what they reward money which is largely derived from TikTok's "gift economy", and recommendations and therefore views are largely based on the gifts. This is in stark contrast with something like YouTube, which is optimizing for 'time spent on YouTube' which drives their creators to make longer clickbait type videos, regardless of how dissatisfying they are.

The recommendation algorithm described here could be tuned to serve any 'objective algorithm', like time spent, profit to Bytedance, or whatever- but because of TikTok's choices it is making a really fun an enjoyable experience for users.

hgsgm• 3 years ago

regardless of how dissatisfying they are.

Why would people spend time watching dissatisfying videos?

content that connects deeply

I'd say "parasocially".

But how does the app measure that? It makes decisions based on a few seconds of watch time. It could optimize for "show videos that trigger donations", but that's kind of trivial.

thomasahle• 3 years ago

There might not be a simple topline objective everything is optimized for. More likely it's a linear combination of all kinds of dynamic business needs.

rightbyte2• 3 years ago

I believe their secret sauce is mainly that they don't manipulate the feed for other factors than if it fits.

The Facebook feed used to work like that before 2013 something. You could clearly see the difference when they started optimizing for ads or what ever metrics they used. Posts could not go "viral" anymore in the way it used to. Zuckerberg capped the vitality so that companies had to pay I guess.

nikolay2• 3 years ago

Can't agree more! Facebook is useless to me today and replaces meaningful to me stuff with ads and other things I don't care about and never engage with, i.e. there's for sure other weights and the feed is not organic!

InitialLastName• 3 years ago

I've tested in the past and counted streaks of 30+ consecutive posts in my Facebook feed without any posts from friends, groups I'm in, or pages I've followed. Just "recommendations" and ads for miles.

matesz• 3 years ago

The most impressive Facebook's feature was hands down Graph Search [1] which was released around this time - 2013. Basically it was an interface to the entire social graph!

Eventually they had to pull those features out. Privacy concerns which lead to people sharing less data? Not sure.

Even now a lot of people get surprised when I put their full name into search box and show them what photos are available via simple search - they completely forget or don't know even that when they are tagged their photos are searchable by their name. I mean now most people I know don't share much anymore, facebook is not that cool anymore.

[1] great demo of graph search https://www.youtube.com/watch?v=DySRn7c736c

pfisherman• 3 years ago

Probably because open access to an entire community’s / nation’s social graph is a huge security risk. I mean if the NSA having access to all your phone call metadata isn’t already bad enough, why would you want any and all malicious state actors, criminal organizations, terrorists, or run of the mill scammers to have access to that type of information too?

hellomyguys3• 3 years ago

Generally true, but not entirely accurate...

"TikTok’s Secret ‘Heating’ Button Can Make Anyone Go Viral" https://www.forbes.com/sites/emilybaker-white/2023/01/20/tik...

They also don't just optimize for consumption, but also for new users. Don't have concrete proof, but it seems they help new users get more views to give them that taste of "15 minutes of fame" as early as possible.

vuln4• 3 years ago

Yeah I tend to agree regarding the 15 minutes of fame for new users. It's doesn't even need to go insanely viral.

How many people have created and published videos on YouTube and got maybe a few tens or a couple hundred views over years.

TikTok new users gets a couple THOUSAND views and the user is like "people like my content I MUST make more!!" And now they're hooked trying to get the same high view count as before.

I would assume (as I have never used the platform) it's also a way for TikTok to grab you financial information "you earned x$ from you first video, please supply you financial information and a scan of your ID for payment.

It's an addiction.

rightbyte2• 3 years ago

Ye I mean there is a limit on how much you can manipulate before the feed gets crappy. A small amount you won't notice at all.

Surely the temptation to do some for profit is too big.

It would not surprise me if TikTok goes the same way as Facebook on this when they get comfy, to milk som short term dollars.

lrvick• 3 years ago

Casinos use similar tactics.

Why do we let kids use this again?

waboremo• 3 years ago

You don't have to look that far for proof, there are plenty of people who made videos complaining about how they don't get good views anymore, they aren't pushed out anymore. Search "why isn't my tiktok account being pushed anymore" and you'll find similar outcry on various forums. You see it by scrolling through their profile enough, they go from 300k views and 50k likes to 1k views and 100 likes. Massive drops, where they haven't reached "self-sustaining" status yet like a Selena Gomez account.

Also as another potential source of proof, tiktok is really effective at pushing alt pages to people's FYP. Which is interesting because their old account will stagnate, but their alts will get pushed quite aggressively despite the content being very similar. So obviously it's not a content issue, tiktok is evaluating them equally and thinking you'll like it, but their old account(s) don't show up as much anymore.

Tiktok isn't as overtly aggressive with this method though, it seems like there's a cooldown period of sorts where it might start pushing you out again if you remain committed to the platform. Still, it's severe enough for people to make note of it and feel seriously demoralized by it.

andirk4• 3 years ago

Similarly, the various dating apps I have used clearly front-loaded the matches with the objectively attractive women in the area. This has been especially hilarious to me as a female friend will often start the process for me (search store, download, sign up, start searching) and more than once they have gotten this skewed version of what's out there, thinking that every nearby woman looks like a movie star.

jackson1442• 3 years ago

I think a key part of their new user optimization is putting very little weight on your previous performance. So if you make a great video as a user with 0 followers, you still have similar odds to a 10M follower user to making it big on the FYP.

There might, of course, still be additional new-user weight. But I think this is at least a part of it.

randomdata3• 3 years ago

Why is TikTok's feed so addicting?

What's addicting about it? Admittedly it kept me entertained the first day or two that I used it, but from that point forward it never wavered from the content that entertained me that first day and soon it became really boring to see the same type of thing over and over, leaving no remaining appeal.

Because everyone talks about how great the recommendation engine is, and me wanting to find great content, I went back to it after a while to see if they fixed the problem. But no, still the same content that bored me away from the app the first time.

willio58• 3 years ago

I believe you but you should know you are in the minority here. Part of why TikTok has taken over the social media space is because of how addicting it is compared to the older players.

jdminhbg• 3 years ago

TikTok is massively popular but it's not yet the case that not using it puts you in the minority, like Facebook does.

randomdata3• 3 years ago

I can completely understand how it keeps your attention longer than looking at photos of your neighbour's cat like Facebook gives you. It's just the lack of variety that makes addiction hard to fathom. Do others not get bored of seeing the same thing over and over and over?

criddell• 3 years ago

I had the same experience. I’ve given TikTok three different tries over the same number of years and it didn’t connect with me at all.

I’ve spent several hours using it and there seems to be very little of interest to me (a 53 year old Canadian living in Texas). It did show me a video of a guy doing some great welding (stack of dimes!) and I watched the whole thing. The algorithm decided that I must want to see every welding video ever uploaded and I really don’t. One or two is enough.

I concluded that TikTok isn’t for my demographic, maybe by design.

jquery• 3 years ago

Same experience here.

Maybe TikTok appeals more to teenagers who have an underdeveloped sense of self. I remember at that age conflating popularity with accuracy and authority. Now when I watch TikTok it's just an exercise in frustration that nothing is cited and it's all about presentation over substance. And that's when the algorithm isn't showing me strange videos in a desperate attempt to keep my attention, a la spiderman/elsa YouTube mashups of yore.

valarauko• 3 years ago

I found it took some curating to get the algorithm to what you're interested in - like videos you enjoy, and aggressively mark videos you don't (hold and select "not interested"). In my experience, by the end of day, the recommendations were more on target.

AstixAndBelix• 3 years ago

For me, the moment I see two videos in a row of the same format and with the same music under it I get a sense of existential dread and close the app. Something stops feeling human and that's why I deleted my tiktok and might delete my Instagram next

paulpauper2• 3 years ago

Same here. It's too short. Nothing there. like watching a bunch of gifs but with audio.

fatcat500• 3 years ago

All that brilliance and knowledge, used to get vulnerable, suggestible teenagers addicted to mind-numbing, anxiety-inducing media. The shame of software engineers is that we have been used to create 1000s of these inhuman products.

jbaczuk• 3 years ago

The shame of those software engineers

JohnFen2• 3 years ago

I do think it brings shame on the whole field. It's not right, but it seems to be the case.

blakebreeder• 3 years ago

idk, my TikTok feed is full of cats and dogs doing funny things.

falcolas• 3 years ago

And birds and sneks. Good, wholesome stuff a vast majority of the time.

tester457• 3 years ago

My question is why teenagers gravitate to horrendous content en masse.

Young girls have feeds full of eating-disorder content, and young boys have feeds mixed with misogyny like Tate. Not all teenagers watch either, but these recommendations are a disease plaguing the platform.

ok123456• 3 years ago

My feed is full of mostly educational content, lawyers and other professionals telling stories, language lessons, old computers and tech, and old commercials.

Most of the videos seem to be 1.5 minutes long.

karaterobot2• 3 years ago

Agreed, but I don't want to excuse anybody. I think every one of those engineers signed up to do it in exchange for money. They used their talents, they weren't forced into it.

imiric• 3 years ago

The problem is not just with the addiction, and it doesn't affect only teenagers.

These systems show paid content mixed with user-generated content, promoting harmful advertisements, political propaganda, or any other ideology anyone with enough resources and desire to influence large segments of society is able to take advantage of.

Let's not forget the Cambridge Analytica scandal, and not doubt that there are _many_ more such companies operating without public knowledge. This is all done legally, and within the terms of service.

This is even without mentioning the rampant data siphoning and shady multi-billion dollar data markets that exist in the background, that generate most of their revenue. After all, all this technology and optimization is serving this end goal.

These platforms are not just harmful to individuals, but to society as a whole. I hope one day governments catch up to the harms they're doing, and regulate them just as they did for Big Tobacco and many other truly evil industries.

MattDemers• 3 years ago

I hope one day governments catch up to the harms they're doing, and regulate them just as they did for Big Tobacco and many other truly evil industries.

But how does that happen when our government officials (at least in North America) seem technologically illiterate, and rely on Silicon Valley for their re-elections and donations (therefore becoming more susceptible to lobbying)?

gruffin• 3 years ago

"The best minds of my generation are thinking about how to make people click ads"

- Jeff Hammerbacher

ipaddr2• 3 years ago

This is reductionist. Why does anyone want you to click ads? Because they want you to see what they shared with the world. The best minds are trying to get people to pay attention to those who wish to pay for that.

A video game programmer might sat.. "all of the best minds are working on silly games" Why are so many bright minds wasting trying to get us into space when we have problems that need to be solved here.

The best minds solving problems is only part of any solution anyways. You need the best salesmen, best leader and right moment to bring in change that mat make lives better.

lostmsu• 3 years ago

Maybe they are not the best minds after all.

dhimes• 3 years ago

I wonder what may happen in the future. I'm sure neither Purdue Pharmaceuticals nor the Sackler family expected to be seen as criminal in what they were doing (opioid epidemic). But when the outcomes were shown to be a result of their actions, they were convicted— or at least forced into settlement.

luckylion• 3 years ago

Nobody has "been used" to create those things. They do it knowingly and willingly, because they're paid huge sums of money in return.

Don't turn it into "we just followed commands" or "we were tricked", own whatever you're working on.

eimrine• 3 years ago

The shame of software engineers

What decisions are on these guys? How to name some variable? How to receive mole salary? Or what manager's/marketologists' decision not to code?

londons_explore3• 3 years ago

The main reason TikTok has awesome recommendations is that all the other big players (Facebook, Youtube, etc) realised back in 2017 that training machine learning models to persuade people to spend more time on their platform was unreasonably effective.

Ie. tell ML to trick someone to spend 14 hours watching youtube every day, and for some small percentage of users, it will actually succeed!

For those people, it's as addictive as drugs. They spent all day on youtube rather than going to work, going to school, caring for their kids, eating or even sleeping! Can you imagine the size of lawsuits that would be heading youtubes way when those people realise they've effectively been enslaved by an algorithm??

Leadership of the big companies put an end to that, instead trying to focus on other metrics, and trying to get more users to each spend some time on the platform.

Well it seems TikTok didn't get the memo...

londons_explore3• 3 years ago

I know someone, mid 30's, jobless, who spends 12+ hours a day on TikTok. They don't even post to it - just watch hours and hours of mindless videos about mental health issues.

justinator• 3 years ago

Maybe not 12+ hours a day, but a good chunk of time and this is my house mate. They watch more Tik Toks about ADHD than they do actually trying to live with ADHD.

fragmede3• 3 years ago

Which is sad, but sans TikTok, would they just be on Reddit or Facebook or Instagram or YouTube or Netflix or Hulu or Disney+ or Paramount+ or HN for 12+ hours instead?

sirsinsalot• 3 years ago

Why expend effort for the dopamine of long-term, delayed gratification when you can just tickle your brain all day for free?

vuln4• 3 years ago

What do they do for income?

colechristensen• 3 years ago

Well Facebook failed because almost no one i knew using it in 2017 still is. It’s like the normal reaction to an addictive thing is to get captured by it for a while, realize the downsides, and drop it after a time. Some people are bad at this and stay addicted but most people escape and go on with their lives. Optimizing for short term addiction sabotages your product.

My facebook feed is a ghost town. Almost zero people i know posting, a couple of straggling groups that remained relevant, and just bs ad driven nonsense.

WhatsName• 3 years ago

Is there a source to that claim? Seems to me that much rather they started dilluting their feeds with ads to the point where they lost all appeal.

Facebook and Youtube on mobile are so packed with unappealing scams that I stop scrolling after 5min

dragonelite2• 3 years ago

People will pick their own poison to fill their free time, in the past it was reading books then it became watching TV all day im sure most millennials are using apps to fill their free time.

I used to fill my free time with world of warcraft spending more time in Azeroth than i did with friends or at school or at work. But i came to the realisation it was just escapism from a boring country/region and a waste of my time. But this is also made me realise how lucky i was to be born in such a boring country/region.

fdgsdfogijq• 3 years ago

The secret sauce:

Let grass roots content creators actually distribute their content. Let the best content go to the top. Give everyone a chance.

Facebooks algorithm:

Support influencers, corporations, and whatever pumps the ad machine.

fragmede3• 3 years ago

I hate to break it to you, but if you think TikTok has "grass roots content creators" vs Facebook having influencers, I have a bridge, as well as a TikTok content house to sell you.

https://www.kapwing.com/resources/tiktok-houses-list/

nerdponx• 3 years ago

But they do seem to be optimizing more explicitly for "user enjoyment" than other platforms, perhaps realizing that this is what keeps and attracts users and their precipitation eyeball-minutes in the long term. Youtube also (not always) does a pretty good job at this, while Instagram can be rough.

cainxinth• 3 years ago

I'm with you, and I'm of the opinion that the current TikTok craze has more to do with its brand than its algo. Facebook is old and lame to kids. Even Insta and Snap don't have as much brand cache as TikTok right now because TikTok is the newest major player with the most buzz. However, it's the nature of all hype cycles for high-flying, shiny new things to eventually fall back to Earth:

https://en.wikipedia.org/wiki/Gartner_hype_cycle

paulpauper2• 3 years ago

Not the same thing. Facebook algos are recommendations from your own friends, not random people.

ssnistfajen2• 3 years ago

It's way easier to find "smaller" content creators on TikTok than Meta's apps. Youtube has been improving a bit in that regard but my recommendations still have too many videos I've already watched. More than half of the accounts I follow on TikTok have <25000 followers, many with average views per video under 10000, with a diverse topic range including: woodworking, software dev, comedy sketches, book reviews, history trivia, exercise tips, urban design, etc. Those who complain about seeing only unintelligent garbage are unwittingly revealing themselves to be the target audience for such content.

MonkeyMalarky• 3 years ago

Clearing out old and stale IDs from the model is interesting. It makes sense for a platform serving a high volume of viral / short lived content. I can't imagine the same trick working at a streaming service.

ludoro• 3 years ago

I also think that paper is super interesting, I talked about it here: https://www.machinelearningatscale.com/how-tiktok-recommenda...

O__________O2• 3 years ago

Relate HN comments:

- https://news.ycombinator.com/item?id=33494796

The secret sauce of TikTok’s recommendations

Related Stories