Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at Pinterest

Summary

Pinterest designed the Pinner Progression initiative to shift its home feed recommendation pipeline from short-term engagement optimization toward long-term user retention. Standard ranking systems model user actions as immediate sequential features, failing to capture the evolving lifecycles of distinct user habits. To address this gap, the platform introduced User Interest Clusters (UICs), which cluster a user's engaged Pins in OmniSage embedding space using dynamic cluster counts and stateful lifecycle metadata. Pinterest integrated a UIC-aware penalty directly into Sliding Spectrum Diversity (SSD) scoring by matching Pin embeddings to cluster medoids via cosine similarity. Online experiments demonstrated that balancing use-case representation increased interaction diversity, extended session lengths, and delivered meaningful engagement gains across the platform.

Context

Pinterest found that optimizing home feed recommendations solely for short-term engagement metrics, such as clicks and saves, fails to sustain long-term weekly active user retention. Existing transformer-based ranking and retrieval systems model users through immediate action sequences, treating content in accelerating and decaying interest categories identically. Because sustainable platform growth relies on durable habit formation across distinct use-cases rather than transient curiosity, the recommendation stack lacked a stateful representation of interest lifecycles.

Approach / What changed

Pinterest developed User Interest Clusters (UICs), which cluster only a user's engaged Pins within the multi-entity OmniSage embedding space. Rather than using fixed cluster counts, UIC dynamically assigns cluster numbers based on coherence thresholds and stores stateful metadata like engagement recency and frequency. Pinterest incorporated a UIC-aware penalty into Sliding Spectrum Diversity (SSD) scoring, assigning Pins to clusters when cosine similarity exceeds 0.85 and penalizing over-represented clusters to surface under-represented interests in later feed positions.

Takeaways

• Use-case adoption exhibits a non-linear relationship with retention, where retention benefits accelerate sharply at the top decile of adoption breadth.
• User Interest Clusters partition only engaged Pins in OmniSage space with dynamic cluster counts based on coherence thresholds, pairing medoids and landmark Pins with lifecycle metadata.
• Sliding Spectrum Diversity assigns Pins to clusters using an OmniSage cosine similarity threshold of 0.85 and penalizes candidate Pins proportionally to existing cluster coverage in the selected set.

Related reading

• Pinterest · Apr 27, 2026

  From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at Pinterest

  Pinterest built a dedicated shopping candidate generation model to optimize for lower-funnel purchase conversions rather than relying strictly on engagement retrieval models. Because offsite conversion events are sparse and delayed, engineers trained a single multi-surface two-tower model using conversion data supplemented by duration-weighted click engagement and unengaged ad impressions as hard negatives. The two-tower retrieval architecture incorporates a parallel DCN v2 and MLP layer structure that allows simultaneous learning from the same inputs without information bottlenecks. Later revisions transitioned the system from a multi-head loss structure to a unified single-head multi-task architecture paired with an advertiser-level objective. Deployed across more than 600 million monthly active users, the system increased shopping conversion volume by 2.3% and boosted return on ad spend by 3.1% for United States shopping campaigns.

  Pinterest Engineering

• Grab · Jul 3, 2019