How Pento modeled aesthetic taste with Qdrant - Qdrant

What this post added

Introduces a novel approach to modeling aesthetic taste by representing users not as single points but as collections of evolving taste clusters. This involves converting user-art interactions into signed weights, generating artwork embeddings, clustering positive and negative interactions independently using HDBSCAN, scoring clusters by recency using an exponential decay function, and representing users as multivectors (positive and negative). The system leverages Qdrant's Recommendation API with multivector search to find artists with overlapping preferences while minimizing similarity to rejected content, demonstrating a flexible and universal recommendation pipeline.