Press and hold on TikTok shows a popup for this.
But yeah I think this makes it pretty clear that TT is not doing anything qualitatively different (except possibly a massive army of human annotators). They just do a good job.
But TikTok is definitely special in this regard.
If you actually use TikTok for any length of time, you can see how remarkable it is. It learns extremely quickly what you want to see (and what you don't want to see), and provides it accurately and in apparently infinite quantity.
The only difference being TikTok uses a lot more human moderation compared with YouTube.
I don't find their recommendation system any special as opposed to their specific operational model.
On other platforms like YouTube for example, there are a lot of factors beyond the video itself that influences if a person even clicks on it (like the thumbnail, the title, the number of views, who the creator is), so if someone doesn't click on a video, that doesn't tell you if they would have liked it or not. Because of that your data to make recommendations isn't as accurate.
That's the main difference I see with TikTok. They make recommendations proactively, instead of reactively.
>I don't work at Twitter, but I doubt that matters very much.
>Also regarding "negative sentiment" feedback from users. TikTok does not have a UI tool to say "I don't like this", but other apps do. For example, you can ask Instagram to "show me fewer posts like this".
There's a big difference in the usefulness of implicit and explicit feedback for recommenders, and there's a long history of research trying to address the shortcomings of each, or combine them to avoid their weaknesses (e.g. research into presentation biases, augmented matrix factorization, MF with implicit data, ensembles)
Implicit feedback (dwell times etc) are particularly prone to noise and biases. On the other hand, the amount of explicit data (likes, downvotes) companies can get is so small compared to implicit data, that many companies don't even bother with their explicit data. E.g. even YouTube claim not to use likes/dislikes, in order to better recommend to the long tail of items/users. So it very much does matter, being able to get good implicit data.
The interesting thing implied by OP is that TikTok seems to have completely sidestepped the pitfalls of each form of data. They in essence seem to have made implicit data be a gold standard judgment of relevance, which is really cool I think.
> I don't want to sound overly dismissive, but it's pretty clear reading this that the author has not worked on recommendation or personalization systems recently.
This is something that hasn't been done in any other large company's recommender system as far as I'm aware. It seems like a paradigm shift to me, albeit, is only possible because of the addictive nature of the app. So at least for similar UIs, this concept makes a lot of recsys research irrelevant and is highly novel. As such, I don't think you're giving OP or TikTok enough credit.
For me at least, tiktok.com is an infinite scroll.
In fact, imgur's desktop website comes to mind as the rare design pattern where you can (or, could, back when it was only gifs) consume the content in a gallery view where the application doesn't know what holds your attention.
I don't want to sound overly dismissive, but it's pretty clear reading this that the author has not worked on recommendation or personalization systems recently.
The high level description of how he believes TikTok works is probably accurate (I don't work there so I can only speculate), but it's not unique to TikTok. Also most of the details are probably not accurate.
Some examples:
> For its algorithm to become as effective as it has, TikTok became its own source of training data.
^This is true of nearly every recommendation system.
> If you click into a text post by someone on Facebook but don’t comment or like ... That negative sentiment is difficult to capture
You can do the same thing with comments that he claims TT does with video (dwell time threshold feature)
> As you scroll up and past many stories, the algorithm can’t “see” which story your eyes rest on.
I don't work at Twitter, but I doubt that matters very much.
Also regarding "negative sentiment" feedback from users. TikTok does not have a UI tool to say "I don't like this", but other apps do. For example, you can ask Instagram to "show me fewer posts like this".
I agree with the claim that the "one thing at a time" UI is nice for the user in a lot of ways. But as a person who builds these things, I don't think this makes much of a difference for machine learning.