Pinterest Engineering introduces Pinner Progression, a program that reframes home feed recommendations around retention rather than pure engagement. The core innovation is User Interest Clusters (UICs) — a stateful, personalized representation of user interests built via Complete Linkage Agglomerative Hierarchical Clustering over a user's last 500 engaged Pins in OmniSage embedding space. UICs improve on predecessors (PinnerSage, OmniSage) with personalized clustering over engaged content only, dynamic cluster counts, and lifecycle metadata per cluster. UICs are integrated across the full recommendation stack: replacing followed-interests in retrieval (UIC-conditioned CLR), enforcing use-case diversity in L1 Utility via penalty-based scoring, making ranking utility weights lifecycle-dependent, and adding UIC-aware penalties to the SSD blending layer. Online experiments showed engagement gains, more diverse content interaction, and longer sessions. Part 2 will cover predicted UICs and reinforcement learning for systematic interest exploration.
