Artsy

Artsy uses The Art Genome Project, a classification system that assigns over 1,000 descriptive characteristics called 'genes' to artworks and artists, powering personalized recommendations and artwork discovery for users of the platform. When a user views or follows an artist, the system surfaces related artists and artworks based on shared genes.

The Art Genome Project assigns each artwork and artist a scored set of 'genes' (characteristics such as art movement, subject matter, or formal quality) on a scale of 0–100, capturing how strongly each attribute applies. These scored gene vectors are used to compute similarity between artworks and artists. Artsy also applies collaborative filtering at the artist level — using behavioral patterns across users (e.g., 'people who follow Andy Warhol also tend to follow Roy Lichtenstein') — to broaden recommendations. Artsy's engineering blog confirms that Elasticsearch powers real-time artwork similarity features across all front-ends. Gene assignment has historically been performed by a team of human art historians, with machine assistance for more visually computable characteristics.

Artsy uses an automated data pipeline to compute an artwork similarity graph that underpins The Art Genome Project's discovery features, processing large batches of artwork data offline to generate similarity scores used in search and browse results.

According to Artsy's engineering blog, the artwork similarity graph that powers The Art Genome Project is processed offline by a generic job engine written in Ruby or by Amazon Elastic MapReduce. The system takes data snapshots from MongoDB, runs computation jobs, and exports results back to the production database. Artsy also uses Jupyter Notebooks with pandas and scikit-learn for more in-depth data analysis work. This pipeline feeds the similarity scores that surface related artworks throughout the platform.