So ... 30 years ago when we were starting Amazon (I was employee #2 there), I had an idea for a feature I wanted in the initial store that is related to what yo

So ... 30 years ago when we were starting Amazon (I was employee #2 there), I had an idea for a feature I wanted in the initial store that is related to what you're trying to do here, but quite different. While I value human curation, and it probably has to come first, what I wanted to do was to take the LoC classification for a book (typically a triple that has 3 levels of specificity) and build a multidimensional search "space" where you could find books close to the one you were looking at. In 1994, alas, the LoC data was only available via an insanely slow modem-based download, which made it unusable. The main data source we were using back then (Baker & Taylor / Books in Print) sadly fucked up their record of the LoC classification to that it was impossible to infer where the gaps between the triple's parts were. The crux of the idea was to leave book classification to librarians, who've been studying this for years and likely doing a much better job than anything us silly software types could come up with. So to use an example from your list above, in my mental image of the interface I wanted: you'd pick "Sapiens" and see it surrounded by a cloud of other books sharing the same LoC triple (could use other library classification methods too). You could zoom out, and now see even more books sharing only the first two elements of the triple, or even further and end up in "History". Think of it as a "word cloud" made of books, rather than words. You could meander over to another book, and from that book zoom in (and out) of other aspects of the LoC's classification space. I still wish I had been able to build this, since it mirrored the best part of my experience in large physical bookstores. But it was not to be. Not sure if the dataset(s) you're using would permit this either, just wanted to mention it.