Interpersonalized recommendations
I just got back from a few days of workshop and backpacking but am not in much of a mind to talk about it, so I ask you this question that is very far from my current mood, but which I was reminded of by the demands of relaxing upon returning home:
Why does Netflix—a service that recommends what to watch based on one’s observed tastes, and which has separate profiles for each member of one’s household—not offer recommendations for pairs or groups of people, who individually have profiles?
I don’t see how this can be hard, if they have good recommendations for both individuals. For instance, they could just check everything suggested above a certain bar for Alice and see if it’s also suggested so for Bob. Even something very janky would seem to be better than the pair doing something equivalent manually (for instance, looking over Alice’s recommendations, and parsing what each one is well enough for Bob to judge if he would like it). But I bet they can do much better.
Group recommendations would also seem to be extremely useful. I’d guess a large fraction of the time that a person wants to watch something, they want to do it with someone else. And furthermore it seems substantially harder to figure out what two people would like to watch together than just oneself, perhaps because neither person has great access to the other person’s mind, so anything that fares well on an intuitive glance then has to be explained to the other person. I’m not sure.
Is there some good reason they don’t? Am I wrong somehow? Do people use Spotify Blend as much as I would have predicted?
Good question. There's a related, more general question - why isn't Netflix (and other streaming services) more interested in recommendation? 20 years ago Netflix offered a million dollar prize for recommendation algorithms that outperformed their existing system. Now you never hear about it. Occasionally you can get a taste of what really good recommendation feels like. Tiktok is famously good at serving up content users like, and I think sometimes YouTube can be very good as well. But the big streaming sites don't seem to care.
Is it because the problem harder than it looks? Is it intractable? Is it because high-quality recommendation algorithms are jealously-guarded corporate secrets?
I think the answer is business models. Netflix and other big streamers don't care about recommendation as a problem because they want you to watch one of the dozen or so featured shows and movies that are in the current rotation. They are no longer in the library business, they are more like studios now, serving up seasonal content, much of which they have made, or at least paid for, themselves.
Can this even be done automatically? Group dynamics of greater than two members are generally chaotic so you would expect any group recommendation algorithm to be importantly wrong most of the time. Thinking about rewatches and "so bad it's good", people watch things together for different reasons that are not simply related to their individual preferences.
What is really needed here is some kind of well designed negotiation protocol implemented with software that helps a group figure out what kind of experience they want to have together.
What does that look like? No idea. What kind of naturally elicits the true information about what experiences people want to have with others other than just doing it live? Can we do better with software than that? It seems difficult to me because most of the information is hidden even to the participants.