The secret sauce of TikTok’s recommendations

I believe their secret sauce is mainly that they don&#x27;t manipulate the feed for other factors than if it fits.<p>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 &quot;viral&quot; anymore in the way it used to. Zuckerberg capped the vitality so that companies had to pay I guess.

I was behind a woman today on an escalator. I wasn&#x27;t shoulder-surfing, she held her phone up so high I couldn&#x27;t help but see it. She was on Tik-Tok. She watched like twenty videos in one minute. One second, click. One second, click. What good this does anybody at all, anywhere, is beyond me. It was all trash.

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.

The secret sauce:<p>Let grass roots content creators actually distribute their content. Let the best content go to the top. Give everyone a chance.<p>Facebooks algorithm:<p>Support influencers, corporations, and whatever pumps the ad machine.

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 <i>unreasonably effective</i>.<p>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!<p>For those people, it&#x27;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&#x27;ve effectively been enslaved by an algorithm??<p>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.<p>Well it seems TikTok didn&#x27;t get the memo...

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&#x27;s but the fuzzy nature of the recommender problem means I am never going to get into the 90&#x27;s.<p>I would like to see more about how they are formulating the problem, I know there is work lately in &quot;sequential recommendation&quot; that is focused on generating a sequence rather than scoring content items, I&#x27;d like to learn more about that.

To be fair tiktok helped me find a lot content (series, movies, etc) and artists (specially standup comedians) that it would take a much longer time to cross-paths with. So I would see a short of that, and then looked around to watch the whole thing.<p>But in the 1 year I&#x27;ve interacted with the app I&#x27;ve already seen the effects of the content creators learning how to game the algorithm and get presented with some contents that were made in a format only to get engagement, similar to a clickbait article.<p>The result? It has spoiled the whole experience for me and I uninstalled the app.<p>And this is how those platforms slowly crumble to their own weight as the crowdsourcing becomes a business for creators and, little by little, content becomes more and more superficial, fake, commercial and uninteresting for the audience, who migrates to another app and everything starts again...

In these articles I never see reference to the &#x27;objective algorithm&#x27; aka what the business processes at TikTok are optimizing for. I think that&#x27;s the real secret sauce.<p>TikTok is optimizing for user generated content that connects deeply with their users. Thats what they reward money which is largely derived from TikTok&#x27;s &quot;gift economy&quot;, and recommendations and therefore views are largely based on the gifts. This is in stark contrast with something like YouTube, which is optimizing for &#x27;time spent on YouTube&#x27; which drives their creators to make longer clickbait type videos, regardless of how dissatisfying they are.<p>The recommendation algorithm described here could be tuned to serve any &#x27;objective algorithm&#x27;, like time spent, profit to Bytedance, or whatever- but because of TikTok&#x27;s choices it is making a really fun an enjoyable experience for users.

As a person who occasionally consumes TikTok videos, I feel like it taps into the same thing that the original StumbleUpon did; show the user fairly quickly a number of related things and give them the option to either consume more of that same thing or find something else that&#x27;s related but also new. Same goes for Pinterest. They all share that core element of quick consumption of related things in common.

Pretty click-baity title. This article is a good &quot;ELI5&quot; summary for non-ML people of some of the techniques &amp; infra described in the arxiv paper, but it has nothing to do with TikTok &quot;secret sauce&quot;.

Relate HN comments:<p>- <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=33494796" rel="nofollow">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=33494796</a><p>Related paper:<p>- <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2209.07663" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2209.07663</a><p>Related GitHub:<p>- <a href="https:&#x2F;&#x2F;github.com&#x2F;bytedance&#x2F;monolith">https:&#x2F;&#x2F;github.com&#x2F;bytedance&#x2F;monolith</a>

<i>&gt; Why is TikTok&#x27;s feed so addicting?</i><p>What&#x27;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 <i>really</i> boring to see the same type of thing over and over, leaving no remaining appeal.<p>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.

This blog post is unlikely to contain TikTok&#x27;s secret sauce, or even anything that is actually used by TikTok.<p>ByteDance runs TikTok, but they also sell recommendations as a service. The paper that this blog post is based on explicitly talks about that service in the abstract, and never even mentions TikTok.<p>Now, if I were to advertise a service that contains some of the ingredients of my other, very popular product, I would surely mention that fact ... but this is not the case here.

Another analysis I&#x27;d really like to see is how Amazon does their forecasting. A friend that works there told me it&#x27;s crazy.

I also think that paper is super interesting, I talked about it here: <a href="https:&#x2F;&#x2F;www.machinelearningatscale.com&#x2F;how-tiktok-recommendation-algorithm-scales-to-billion&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.machinelearningatscale.com&#x2F;how-tiktok-recommenda...</a>

<a href="https:&#x2F;&#x2F;www.npr.org&#x2F;sections&#x2F;codeswitch&#x2F;2022&#x2F;02&#x2F;14&#x2F;1080577195&#x2F;tiktok-algorithm" rel="nofollow">https:&#x2F;&#x2F;www.npr.org&#x2F;sections&#x2F;codeswitch&#x2F;2022&#x2F;02&#x2F;14&#x2F;108057719...</a>

Clearing out old and stale IDs from the model is interesting. It makes sense for a platform serving a high volume of viral &#x2F; short lived content. I can&#x27;t imagine the same trick working at a streaming service.

Pro-tip: On Instagram, interact with ads that you don&#x27;t mind. That way the algo picks those or similar ads which makes a more pleasant experience. For example, now I mostly get ads of nice pictures of watches.

Curious: every terms of service says you are committed to not reverse engineering their tech. Does this count, and if so, what does that mean?

this is a wall of text citing a single paper which itself is entirely unverifiable.<p>regardless, if this described how drug dealers engineer the distribution of their product to optimise addiction, we would not be celebrating it as an engineering marvel.