KOEN
Read Original
#AI

Evolution of Multi-Objective Optimization at Pinterest Home feed

Evolution of Multi-Objective Optimization at Pinterest Home feed
01

Summary

Why Pinterest Stopped Chasing "Saves" and Started Caring About Feed Harmony

The technical journey of balancing short-term metrics with long-term retention using multi-objective optimization.

Explore how Pinterest builds the final layer of its recommendation funnel to balance engagement, diversity, and content quality. This article covers the transition from complex DPP models to a scalable SSD framework and the integration of multimodal signals.

  • 01Resolving the trade-off between immediate saves and long-term session health
  • 02Transitioning from high-complexity DPP to efficient Sliding Spectrum Decomposition (SSD)
  • 03Migrating ranking logic to PyTorch-based model serving for faster iteration and experimentation
  • 04Implementing Soft-Spacing to prevent content clustering without relying on rigid hard-filters
  • 05Leveraging PinCLIP and Semantic IDs for deep semantic and visual diversity control

RECOMMENDATION

Essential reading for ML engineers focused on re-ranking, multi-objective optimization, and maintaining ecosystem health in large-scale recommender systems.

The Problem

In Home Feed recommendations, optimizing only for short-term engagement like saves leads to visually repetitive content, which reduces long-term user satisfaction and session duration.

The Solution

Pinterest evolved its re-ranking layer from Determinantal Point Process (DPP) to the more efficient Sliding Spectrum Decomposition (SSD) implemented in PyTorch, integrating a soft-spacing framework and Semantic IDs for richer diversity.

The Result

The introduction of DPP improved time-spent impression by over 2% after the first week, and the migration to SSD enabled lower serving latency and the inclusion of multimodal signals like PinCLIP for better feed composition.

Trade-off

Diversification strategies can lead to a slight decrease in immediate engagement metrics (like saves) on the first day, and more advanced algorithms require complex infrastructure for handling high-dimensional embeddings and real-time signals.

03

Key Concepts

Concept · 01

DPP (Determinantal Point Process)

A probabilistic model that uses a kernel matrix to select a diverse subset of items by balancing relevance and similarity.

  • Implemented as Pinterest's first-generation diversification component to boost user retention metrics.
Concept · 02

SSD (Sliding Spectrum Decomposition)

A position-adaptive diversification method that rebalances exposure of latent spectra within a sliding window.

  • Replaced DPP to reduce serving latency and simplify implementation using standard linear algebra blocks in PyTorch.
Concept · 03

Semantic ID

A hierarchical representation derived from discretization of embeddings that provides a stable notion of item semantics.

  • Used as a penalty term in SSD to discourage clusters of Pins with high semantic overlap, improving perceived diversity.