This document discusses Spotify's approach to generating recommendations for who a user should follow. It first generates candidate recommendations for each user using their 2-hop social network or Facebook friends. It then trains a machine learning model to rank these candidates by extracting features like the number of common connections, cosine similarity of profiles, etc. It found gradient boosted decision trees performed best. Testing recommendations on Spotify employees found over 60% liked and knew the recommended user. It optimized the process by loading data into a database instead of memory to allow for multiprocessing.