| One huge challenge is in identifying and measuring the axes of appreciation. Asking "What do you like about X?" is a tough way to extract good data. People usually cannot explain why they like things. And it's legitimately difficult to know sometimes. Also, tastes are often context-driven/sensitive. A book that I loved when I read it last summer in Barcelona, or on the 16-hour flight to Auckland ... does not necessarily map to what I would enjoy reading right now. Or that I should pack for my trip next week. I've tried to suss this out in music. Songs are theoretically more approachable than books/films/etc: Bite-sized consumption quanta, a fairly robust (but large) genre taxonomy, one basic grounding theory (not really, but a reasonable approximation for the culture within which I exist). Then you can split out by instrumentation, style, arrangement, tempo, etc and get some well-defined groups. This doesn't work. It's over-analytical, and under-representative of human taste spectra. The "best" engines use high-resemblance cohorts, but no one actually likes them -- they give lame obvious suggestions, and are terrible at surfacing surprises. They're OK at "good enough, sometimes" in the same way that turning on a TV for the 6pm news and sitting there on the same channel until Letterman signs off was "good enough" (i.e. horrifically bad!) back when serial TV was a thing. There remains something ineffable about taste -- "It don't mean a thing, if it ain't got that swing". (Ironically, "swing" is now probably measurable! But the point remains for other as-yet-undefined axes.) |
In order to be able to really recommend something as multi-faceted as a book, movie, or song, you have to know a person on pretty much every level. I suppose seeing a person's entire social graph, search history, LLM history, media consumption history, and browser history might get you close, but it's still a Hard Problem™.