Fake-Edit: I wrote all this, and then realized, I could just ask: Can someone very familiar with recommendation algorithms comment on the content of Reginald's

Fake-Edit: I wrote all this, and then realized, I could just ask: Can someone very familiar with recommendation algorithms comment on the content of Reginald's presentation? TL;DR, references to Pandora's recommendation algorithm: Describe GOOD things broadly, permanently, and personally. Describe BAD things specifically, temporally, and impersonally. Now let's think about Pandora. Specifically, Pandora's recommendation that you thumbs up (upvote) a few things, and thumbs down (downvote) many more things. Why? Let's think about Pandora's recommendation algorithm, or, recommendation algorithms in general (pandora's is good because you can give input every ~5 minutes, as opposed to Netflix, where it's ~1hr). Note: Not an expert! I just read stuff on the internet a lot. Pandora takes your upvote, and it promotes an entire region of music-space. Pandora takes you downvote, and it demotes a very narrow region of music-space. In other words, it broadly interacts with GOOD things, and specifically interacts with BAD things. Pandora isn't passing judgement on you, but it is a personal thing (or, really, per-channel), and you can un-vote anything; so it at first glance it doesn't apply on the other axis but I wouldn't put it past them if my upvotes carried across channels (so they're more personal, while the downvotes stay "impersonal" to each channel), and it might be better if downvotes decayed in some way over time. In any event, you can imagine what this looks like in state-space: You start with a broadly defined region, and you start taking holes out of it.