Skip to content

SYSTEM Cited by 1 source

Pinterest Sliding Spectrum Decomposition (SSD)

Definition

Sliding Spectrum Decomposition (SSD) is the diversity algorithm in Pinterest's Blending layer that determines the final ordering of Pins in a user's Home Feed (Source: sources/2026-07-27-pinterest-pinner-progression-better-use-case-representation-driving-weekly-active-user-growth). It is the final integration point for UIC in the serving funnel.

The problem: visual diversity ≠ use-case diversity

Prior to Pinner Progression, SSD optimized variety at the individual Pin level — ensuring adjacent Pins looked and felt different — but had no awareness of the higher-level use-cases a user is engaged with. As a result, a feed could appear visually diverse while still being dominated by a single interest category (e.g. ten visually distinct "apartment decorating" Pins crowding out everything else).

UIC-aware penalty

Pinner Progression introduced a UIC-aware penalty term directly into SSD scoring — an instance of penalty-based diversity in the blending layer:

  1. Associate each Pin with a UIC by computing cosine similarity between its OmniSage embedding and each cluster medoid; assign the cluster with the highest similarity above a 0.85 threshold. Unmatched Pins fall into a default group.
  2. Penalize Pins proportionally to how much their UIC is already represented in the selected set: the penalty for over-represented clusters grows with each selection (UIC_coverage_i = fraction of already-selected Pins sharing cluster i).

Effect: Pins from under-represented interests are increasingly favored as the feed gets longer. Newer interests surface not by overriding relevance at the top of the feed, but by naturally winning in later positions where dominant clusters have already accumulated penalty.

Impact

In online experiments, UIC-aware SSD led to meaningful engagement gains, much more diversity of the content users interacted with, and increases in longer sessions — confirming that balancing use-case representation translates to deeper engagement.

Seen in

Last updated · 766 distilled / 2,225 read