An empirical investigation of personalization factors on TikTok | Hacker News Reader

An empirical investigation of personalization factors on TikTok

arxiv.org

65 pointsasmy35 comments

I started using TikTok and wanted to check how it would suggest me videos if I didn't like anything, as it would make it easier for it and I wouldn't be able to see how sophisticated it really is.

So it obviously would start in a very broad content but already target at my demographics (male 30-40 y/o, which BTW gave a worrying amount of very young girls dancing in a sexualized way) but with time it would certainly adapt and be really spot on on the videos I liked (mainly stand up/comedy stuff). Everything based in signals other than liking of following.

They obviously track all interactions from the user (time watching, repeats, shares, OP profile checks, etc) to make this targeting. Also its ML for visual recognition used to identify key aspects on the video (people, location, etc...) is very powerful as it would guess my favorite comedians and start exploring with other videos.

Also they certainly use bandit/lookalike model there as it would recommend some wildly different videos for me.

And I can bet that that last video it shows when you click to leave the app (on Android at least) is probably the highest rated recommendation for you. To really lock you in there.

> Through this approach we were able to test and analyse the affect of the language and location used to access TikTok, follow- and like feature, as well as how the recommended content changes as a user watches certain posts longer than others. Our results revealed that all tested factors have an effect on the way TikTok’s RS recommends content to its users. We have also shown that the follow-feature influences the recommendation algorithm the strongest, followed by the video view rate and like feature; besides, we found that the location is a stronger influential factor than the language that is used to access TikTok. Of course, this analysis is not exhaustive and includes only the most explicit factors, while the algorithm without a doubt can be influenced by many other aspects such as, for instance, users’ commenting or sharing actions

Subjectively that follows my experience on TikTok: everything you do feeds back into the recommendation algorithm, including how long you watch a video; but the explicit signals (like and follow) have enough weight to make it feel like you have control over where the current takes you.

I assume being able to get at the very least an implicit like/dislike signal out of every view is key to TikTok's Exploration/Exploitation tradeoff, which is what makes the platform feel much less stale than e.g. YouTube (which seems to only recommend things it's very sure about).

I wonder if there is a multi-arm bandit hidden in there somewhere.

I guess the short form video and the "skip past it" culture lowers the risk of exploration -- it's kinda the opposite of that notorious Netflix optimization challenge.

It's hidden behind the 'live' button top left. Viewers buy emoji with real money and send them to creators, creators get money, and Bytedance takes a cut. Creators do all sorts of things, some show them working, at a farm, at a shipping warehouse, whatever. Others play bingo or scratch-off tickets. Still others sit and talk or play music with their viewers. Sending an emoji is highlighted and so are easier for creators to acknowledge the emoji/give a shout out to that particular username. Thing is, creators only sometimes shout out, encouraging multi-arm bandit behavior on the part of the viewer.

Shorter duration videos = more samples to train the recommender.

You also have clear negative examples for the recommender, because not watching something is an active choice (you have to swipe away).

I wonder if it listen to audio or use the video feed to record noises you make, or your pupils dilatation and evaluation the autonomous nervous system arousal that emerges from the reaction to the content.

Because time spent, scrolling rhythm and likes seem like too weak of a signal.

I don't understand how it works for people.

I tried for what felt an eternity, and it kept showing me terrible content.

So I removed it.

Is there something that is not made to turn your brain into a sponge on tik tok ?

Your experience, while worth noting, is clearly atypical.

I understand that I probably have a niche taste, been a nerd and all, but still.

Maybe I'm not the target?

Do they have a lot of millennial users or is it mostly for the younger generation?

I think it's just broken in some cases for whatever reason. Like 9 months ago I mentioned trying out Tik Tok for some weeks and despite lots of "not interested", likes etc, I got swamped by 99% garbage, including videos I specifically said I don't want to see.

TikTok has lots of content for nerds (though I think they prefer the term neurodivergent). But there is certainly a trend that TikTok content is comparatively shallow. For example Hank [1] and John Green [2] have quite nerdy content, but it's much shorter less developed thoughts than their youtube channel [3]. Or NileRed: his most recent TikTok[4] (eating capsaicin) is in no way comparable to his most recent YouTube[5] (synthesizing cotton candy from cotton balls). In terms of nerdy content it probably shines more with life advice than with science facts, simply because of the short video format.

Not sure about the age group, most of TikTok is probably targeted at ages 12-40.

1: https://www.tiktok.com/@hankgreen1?lang=en

2: https://www.tiktok.com/@literallyjohngreen?lang=en

3: https://www.youtube.com/c/vlogbrothers

4: https://www.tiktok.com/@nilered/video/7063145462794145030

5: https://www.youtube.com/watch?v=vHuFizITMdA

I found that my initial foryou feed was pretty terrible and I ended up in political content pretty quickly. Once I nudged the algorithm by searching for my own hobbies, my foryou feed has been pretty great. I think probably every now and then it gets stuck in some local optimum, but it's pretty responsive to when you search for a specific niche going forward IME. Ultimately though like other commenters have mentioned, the content can be pretty shallow compared to other media

> Ultimately though like other commenters have mentioned, the content can be pretty shallow compared to other media

My one time experience felt like browsing the youtube trending page continuously. Which I know must be something a lot of people like, because the videos in there have hundred of millions of views.

But this kind of content is the opposite of everything I enjoy. Does tik tok show a different face deep inside the algo tree?

Speaking as a tiktok-using millennial nerd, it only took about a week or so for my feed to exclusively send me content on D&D, tech advice, science fiction skits, history lessons, and the occasional book review.

The content is there, but sometimes you do have to help the algorithm a little by using the search feature and liking the sort of videos that you actually want to see.

My experience exactly, and my friend's experience too. I'm inclined to think every "tiktok is soo great after you put time into it" comment I see is some form of Stockholm syndrome..

The content you like might not be very available on tik tok. From personal experience, it’s able to hone down on niches you like pretty quickly.

Most impressively, without having an account or liking specific videos, it can narrow down the information. Especially since they have to use some kind of algo to contextually understand videos without hashtags or subtitles.

> Most impressively, without having an account or liking specific videos, it can narrow down the information. Especially since they have to use some kind of algo to contextually understand videos without hashtags or subtitles.

I think they use some kind of object recognition to sort videos . Like if you watch videos with one particular object, you are more likely to see videos with that object.

You have to actively engage and be focused or you'll get lost..

You have to actively like content you like and follow people you want to see more of. Have a motive and plan of what type of content you want to consume and you get that kind of stuff.. I like and followed people who are nearby and visit beautiful places close by... This is like 80% of my feed and I like it...

For the longest time I didn't download it cause of its reputation. It's pretty good, no cruft of YouTube. 'youre watching this, that.. like subscribe.. word from of sponsers' none of that nonsense - pure content. Both FB and YouTube are copying the features button for button. IMO, Not bad

I haven't managed to find non-meme content on tiktok.. But I found that SponsorBlock and uBlock make youtube bearable

I wonder if the recommendation engines are different per country. For example in the US, you usually can't be recommended videos from accounts outside the US. Even with a VPN, spoofed geo location, system language, added video language you will be served only mostly videos based on your IMEI.

I think it's not the IMEI as on my imported phone I would see mostly local content. (and very bad content at that.. uninstalled after a while and never looked back)

That's odd. I get recommended Japanese content despite living in Canada. Much of this content is welcome too.

Have you tried without a SIM card? I think they use the SIM Card Country Code too.

Set more languages from your profile or account settings. I put it to English, I live in a small euro country and 80% of the content is American to be honest. And at least football related content has various people from all over the world who speak in English in their videos. That comes from top of my head but I don't keep track of it obviously.

I only wanted English content but you can add more languages.

I was searching some information on what are the principles behind the TikTok algorithm/recommendation engine and I found this very recent preprint. I wanted to have some opinions on its validity.

It's probably valid but on a quick reading I don't see a lot of depth or a clear understanding of "what the magic is".

Yeah, I pretty much understood some of it because it is very clear. I was researching some explanation or first-principles approach to the TikTok algorithm but going through ResearchGate/Arxiv but found few technical reports and the best one was this one (and very recent). And going through Google is a mine-field because of marketing companies SEO bombing the words "Tiktok" and "algorithm"

I'd like to see them address videos that game the system. I'll see videos structured so that the beginning is very slow, and so you must wait to get to the meat of the content. People do wait, and so the video is algorithmically rewarded, but that doesn't mean the video is good. Some videos even have a "fake glitch" effect where the first few frames are from later in the video. It's as if to say "here's what the video's about, but first wait through this boring part so that my video gets rewarded for engagement."

I once watched a video over several times because I couldn't make out what was supposed to be "off" about it at the end. I finally realized that there was nothing strange or anything, but the video led you to believe there was. Therefore many people, like me, stupidly rewatched it in order to figure out that there was nothing interesting. And that's why it had a boosted signal!

In Figure 6 (described in section 4.4), the content seen by "the control user 50" and "the active user 49" are both in large part by content creator "miakhalifa" (32.4% for the control, 68.6% for the active). The "experimental scenario" is "scenario 28", described as "Follow a random content creator".

Mia Khalifa is a porn actress. A quick search indicates that her TikTok posts are at least sexually suggestive.

I wish I understood what "control" meant in this paper, and what exactly this scenario tested. Does it really mean that a fresh user profile is going to see 1/3 of its feed coming from a porn star? I've never used TikTok so I don't know!