4 ms·

I actually think it's got much more to do with the training data. Given that you are watching a video, you need to either swipe away from it, or finish watchin

I actually think it's got much more to do with the training data. Given that you are watching a video, you need to either swipe away from it, or finish watching it. This provides either a 1 or 0 for the video classification model. The important contrast here is with FB/IG feed where you can scroll aimlessly without engaging, leaving you with perhaps 1 engagement out of 10 (or whatever). The attached doc suggests that they are only using unsupervised learning, which I find very hard to believe, to be honest.

pb7 6y ago
What surprises you about the unsupervised learning?
- disgruntledphd2 6y ago
  Normally supervised learning is much, much more effective.
  - wenc 6y ago
    I wonder if for random content discovery, unsupervised learning methods are less likely to be overconstrained by independent variables, and hence are freer to make better open-ended recommendations. Supervised algorithms almost assume too much about a user -- they assume the correlation structures that are true in the past are also likely to hold in the future. This assumption holds in deterministic environments, but is false or unnecessarily constraining in stochastic domains, where it is widely known simple models and heuristics tend to perform much better. I feel "new content discovery" is more of a stochastic problem rather than a deterministic one (which is the environment most conventional recommendation engines dwell in, hence most rec engines only make conservative recommendations). For all we know TikTok's algorithm could simply be a combination of rules of thumb + simple clusters + randomization that happen to work well. I've seen so many instances (in real life) where simple models vastly outperform complex models in stochastic situations. (p.s. in my opinion, YouTube's recommendations tend to be rather on the deterministic side. My recommendations seem to be mostly based on what I've watched and liked in the past and so the recs tend to be a bit boring. It clearly works great for YouTube from a monetization perspective, but it doesn't unearth a lot of interesting new content for me -- I have to search for those.)
jakear 6y ago
Reducing engagement data to 0/1 would lose a lot of precious data, it treats getting 50s through a 1m video worse than watching the whole of a 5s clip, and watching a 2s clip the same as watching a 2s clip on loop 20 times because it's so incredible. Given the point of the app is to get people on it for a long time, it would make much more sense to track number of seconds watched and train to maximize that.
- disgruntledphd2 6y ago
  Like, this probably isn't how I would implement such a system. But it's an important explanation for how TikTok does such good recommendations. It's a HN comment, not an in-depth blog post :)
benreesman 6y ago
Dwell is a very powerful ranking signal in a binary classifier. Pairwise association of videos watched by the same user consecutively or even just sampled pairs from their last N videos will get you a video embedding. Pairwise sampling of users who watch the same video to the end will get you a user embedding. Turking category tags will prime the pump for other types of embeddings. These things can be ensembles, stacked, force learned jointly, etc. All of this comes out of the box in Keras (though it’s up to you to feed the data in fast when you’ve got a lot office). You can argue that getting latent representations/factorizations without explicit “user clicked show me more” is semi-supervised I guess, but if so all the recommender stuff since the Netflix Challenge meets that criteria to one degree or another.
- londons_explore 6y ago
  What surprises me is all of what you've said is the way pretty much all recommenders work - YouTube, Facebook, Twitter, etc. will all be doing this. Yet people don't spend hours per day browsing tweets. What is it about tiktok which is so much more effective?
  - lalos 6y ago
    You don't have that equivalent signal, that's what matters. i.e. Twitter doesn't know if you even finished reading the tweet which is what is providing a strong signal in the case of TikTok, skipped vs watched once vs watched multiple times.
  - hooande 6y ago
    I spend hours per day browsing tweets. Nothing about this algorithm seems to be groundbreaking. TikTok has a strong brand and is benefiting from that
  - SergeAx 6y ago
    Facebook and Twitter are social-graph subscription-based services. TikTok just gives recomendations without subscriptions while knowing nothing about my social graph.
nl 6y ago
> The attached doc suggests that they are only using unsupervised learning, which I find very hard to believe, to be honest. Where are you seeing that? I see: > Recommendations are based on a number of factors, including things like: > User interactions such as the videos you like or share, accounts you follow, comments you post, and content you create. > ... and > A strong indicator of interest, such as whether a user finishes watching a longer video from beginning to end... which indicates the opposite https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you https://newsroom.tiktok.com/en-us/how-tiktok-recommends-vide...

pb7 6y ago

What surprises you about the unsupervised learning?

disgruntledphd2 6y ago
Normally supervised learning is much, much more effective.
- wenc 6y ago
  I wonder if for random content discovery, unsupervised learning methods are less likely to be overconstrained by independent variables, and hence are freer to make better open-ended recommendations. Supervised algorithms almost assume too much about a user -- they assume the correlation structures that are true in the past are also likely to hold in the future. This assumption holds in deterministic environments, but is false or unnecessarily constraining in stochastic domains, where it is widely known simple models and heuristics tend to perform much better. I feel "new content discovery" is more of a stochastic problem rather than a deterministic one (which is the environment most conventional recommendation engines dwell in, hence most rec engines only make conservative recommendations). For all we know TikTok's algorithm could simply be a combination of rules of thumb + simple clusters + randomization that happen to work well. I've seen so many instances (in real life) where simple models vastly outperform complex models in stochastic situations. (p.s. in my opinion, YouTube's recommendations tend to be rather on the deterministic side. My recommendations seem to be mostly based on what I've watched and liked in the past and so the recs tend to be a bit boring. It clearly works great for YouTube from a monetization perspective, but it doesn't unearth a lot of interesting new content for me -- I have to search for those.)

jakear 6y ago

Reducing engagement data to 0/1 would lose a lot of precious data, it treats getting 50s through a 1m video worse than watching the whole of a 5s clip, and watching a 2s clip the same as watching a 2s clip on loop 20 times because it's so incredible. Given the point of the app is to get people on it for a long time, it would make much more sense to track number of seconds watched and train to maximize that.

disgruntledphd2 6y ago
Like, this probably isn't how I would implement such a system. But it's an important explanation for how TikTok does such good recommendations. It's a HN comment, not an in-depth blog post :)

benreesman 6y ago

Dwell is a very powerful ranking signal in a binary classifier. Pairwise association of videos watched by the same user consecutively or even just sampled pairs from their last N videos will get you a video embedding. Pairwise sampling of users who watch the same video to the end will get you a user embedding. Turking category tags will prime the pump for other types of embeddings. These things can be ensembles, stacked, force learned jointly, etc. All of this comes out of the box in Keras (though it’s up to you to feed the data in fast when you’ve got a lot office). You can argue that getting latent representations/factorizations without explicit “user clicked show me more” is semi-supervised I guess, but if so all the recommender stuff since the Netflix Challenge meets that criteria to one degree or another.

londons_explore 6y ago
What surprises me is all of what you've said is the way pretty much all recommenders work - YouTube, Facebook, Twitter, etc. will all be doing this. Yet people don't spend hours per day browsing tweets. What is it about tiktok which is so much more effective?
- lalos 6y ago
  You don't have that equivalent signal, that's what matters. i.e. Twitter doesn't know if you even finished reading the tweet which is what is providing a strong signal in the case of TikTok, skipped vs watched once vs watched multiple times.
- hooande 6y ago
  I spend hours per day browsing tweets. Nothing about this algorithm seems to be groundbreaking. TikTok has a strong brand and is benefiting from that
- SergeAx 6y ago
  Facebook and Twitter are social-graph subscription-based services. TikTok just gives recomendations without subscriptions while knowing nothing about my social graph.

nl 6y ago

> The attached doc suggests that they are only using unsupervised learning, which I find very hard to believe, to be honest. Where are you seeing that? I see: > Recommendations are based on a number of factors, including things like: > User interactions such as the videos you like or share, accounts you follow, comments you post, and content you create. > ... and > A strong indicator of interest, such as whether a user finishes watching a longer video from beginning to end... which indicates the opposite https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you https://newsroom.tiktok.com/en-us/how-tiktok-recommends-vide...