Evolution of Multi-Objective Optimization at Pinterest Home feed

This article discusses Pinterest's evolution of the multi-objective optimization layer in their home feed recommendation system.

• •Pinterest's recommendation system uses cascaded stages: retrieval, pre-ranking, ranking, and multi-objective optimization
• •Feed diversification is critical for long-term engagement; removing DPP reduced session time by over 2% after one week
• •V1 used Determinantal Point Process (DPP) with GraphSAGE embeddings to balance item relevance and similarity
• •V2 replaced DPP with Sliding Spectrum Decomposition (SSD) in early 2025, offering lower computational complexity and simpler PyTorch implementation
• •Added Unified Soft-Spacing Framework in mid-2025 to penalize low-quality content while maintaining diversity