How Pento modeled aesthetic taste with Qdrant - Qdrant

This post details how Pento modeled aesthetic taste using Qdrant. It describes a system that converts user-art interactions into a semantic vector space. Key technical contributions include: 1. Interaction acquisition and conversion into signed weights (positive/negative affinity). 2. Artwork embedding using image encoders. 3. Interaction clustering with HDBSCAN to identify aesthetic themes, using medoids for cluster representation. 4. Scoring taste clusters by recency using an exponential decay function and keeping the top 50% of clusters. 5. User representation as multivectors (positive and negative). 6. Retrieval via Qdrant's Recommendation API, leveraging its scoring strategy for multivectors to find artists that overlap with positive preferences while avoiding negative ones. This demonstrates a novel application of vector databases for personalized, nuanced recommendation systems beyond traditional collaborative filtering.