> an average of all humans likes, which will never fit your exact personal unique preferences
Well, it's not as if that's an unknown problem in machine learning. In all generative domains (and I suppose recommendations are a generative domain in a sense) there's a risk of "mode seeking", where the model produces bland/blurry/gray output in order to cover its bases, and not miss too badly. It's rarely what we want.
There's also the often quoted "Most of the volume of a high-dimensional orange is in the peel, and the average of two vectors is very unlikely" (Well, I like to quote it because it's fun to say. I forget where it originally came from.) In plain English, if you need more than two dimensions to characterize variation, then you need to use spherical interpolation to find a likely data point in between them, not regular interpolation.
It's surprising how many times this has been overlooked and rediscovered.
Well, it's not as if that's an unknown problem in machine learning. In all generative domains (and I suppose recommendations are a generative domain in a sense) there's a risk of "mode seeking", where the model produces bland/blurry/gray output in order to cover its bases, and not miss too badly. It's rarely what we want.
There's also the often quoted "Most of the volume of a high-dimensional orange is in the peel, and the average of two vectors is very unlikely" (Well, I like to quote it because it's fun to say. I forget where it originally came from.) In plain English, if you need more than two dimensions to characterize variation, then you need to use spherical interpolation to find a likely data point in between them, not regular interpolation.
It's surprising how many times this has been overlooked and rediscovered.