So, it’s just likes-based recommendations. And the selection not from all users, but from the active users and users I follow. An old way was not best for sure. New one is good, but can be done better. There are two ways actually. As a user I like some photo cause I like it color palette(strong bw or some particular color), geometry(lines or curves) and clearity(which can be based on pulse for example). Palette and geometry we can get during upload. And from this data we can build recommendations. Another way(without changing db sctructure and whole process) is to get info not about what photos I liked but recommendations based on adding to collections and leaving comments, cause if you like photo you’ll give it a like, but if you Love it you will comment or add to some collection. I think these approaches could give better results in addition to your new experimental algorithm. Anyway, I like what you are doing as a user and as a developer. Thank you a lot.