Music recommendations using cover images (part 1)

An exploration of using album cover images as a proxy for music recommendations on Spotify. Starting with a simple pixel-based similarity approach (resizing to 16x16, grayscale normalization, Euclidean distance via Annoy), the author finds poor results except for visually uniform genres like minimal techno. Switching to GIST image descriptors via pyleargist yields meaningfully better AUC scores. The post ends with plans to explore learning a mapping from image space to collaborative filtering space and deep learning approaches.