Hacker News Reader

Comment by Eliezer | Hacker News Reader

I noticed that YouTube's recommendations had suddenly gotten better! I wondered if they were using a new statistical approach, or had just started really optimizing at all because the old recommendations were extremely naive. I'm actually a little disappointed to find out that it might just be another deep learning thing. (Yes, it works, but I feel like you learn a little less about problem structure when what you read is mostly "we threw a generic function approximator and 30,000 hours of GPU time at it".)

View on HN

That's curious, because in my opinion YouTube recommendations haven't been any good since 2009. Instead of getting interesting, strange and niche content, I'm bombarded with videos that have >100k views, feature clickbait titles and thumbnails and are generally incredibly low effort content.

Methods that work better for a population as a whole might not work better for a large subset of that population, and might even cause users to stop using features entirely. The lack of transparency in recommendation algorithms combined with the homogenizing effect of distributing low-quality content this way is something I find somewhat depressing.

Maybe those are just the videos that you're statistically more likely to watch through to the end based on your viewing history...

Precisely. But that might be a local minimum. "Show him boobs and action trailers" is guaranteed to make him stay another 40min.

But perhaps there is a more risky strategy that takes longer to craft and actually delivers hours and hours of content to the user (but needs to fail longer before getting there).

It seems like reinforcement learning would be useful, i.e. at a high level, forming a policy for recommendations would require balancing exploration (experimenting with more risky recommendations) vs. exploitation (showing you recommendations that it knows will likely lead to clicks) and using the click-throughs, time spent watching the video, etc. as reward signals.

Does anyone know whether RL is used for recommendation in practical settings, and if so what is the current state of the art?

This is a very natural avenue and an active area of research at Google/Deep Mind. Stay tuned...

YouTube has used machine learning in recommendations for many years.

We have struggled with interpretability, both while debugging mistakes made by the system and exposing plausible "reasons" to users.

There was a fascinating discussion [1] about interpretability during a deep learning panel at KDD this year.

[1] https://www.youtube.com/watch?v=furfdqtdAvc#t=54m25s

Mine seem like they have gotten a lot more click-baity. I wonder if that's an artifact of their optimization goal.

If people fall for the click bait(they do), then they get more views from optimizing for it. It would be more surprising if an optimization algorithm for views didn't favor this.

Probably some part of it is that there's more clickbait videos being created now, which is making them get more views.

It's a vicious clickbaity cycle.

Throwing passive aggressive shade at what is current state of the art in machine learning is unwarranted.

Please. Deep learning is cool and all and has its applications but state if the art inplies that it,s the best for everything when in reality is that they are just very good for a very little subset of problems.

Reply on news.ycombinator.com