One thing I wish you could do with Spotify is adjust the temperature of its audo generated playlists. For a certain genre I might want the playlist to rely either more or less on the well known top songs in that genre versus more esoteric choices.
Working on weighted averages for each song feature embedding. This might be the fix for that. Thinking about comparing the query embedding to sentences like "This query relates to musical genres" or "This query relates to the key and modality of songs" etc. This could be add importance to the "popularity" feature I have in my dataset for queries that contain words like "obscure" or "very popular".
Just an idea right now, I'll get to it when I can!