A lot of people have asked me what models we use for recommendations at Spotify so I wanted to share some insights. Here’s benchmarks for some models. Note that we don’t use all of them in production. Performance for recommender models  This particular benchmark looks at how well we are able to rank “related artists”.

Erik Bernhardsson

A 2013 benchmark comparison of recommender system models used at Spotify for ranking related artists. Models evaluated include a proprietary latent factor method (vector_exp) trained on 50B+ events via Hadoop, word2vec on playlist data, RNNs on session data, Koren's collaborative filtering for implicit feedback, LDA, Freebase-based latent factors, and PLSA. Key findings: PLSA and LDA underperform, sequence-aware models (RNN, word2vec) add significant value, but the top performer is a bag-of-words latent factor model trained on massive log data.

Model benchmarks