Evolution of Multi-Objective Optimization at Pinterest Home feed

Pinterest's engineering team details the evolution of their Home Feed multi-objective optimization layer, which sits at the final stage of their cascaded recommendation funnel. Starting with Determinantal Point Process (DPP) in 2021 for feed diversification, they migrated to Sliding Spectrum Decomposition (SSD) in early 2025, implemented in PyTorch for lower latency and easier feature integration. SSD enabled richer similarity signals including visual, text, and graph embeddings. They also introduced a 'soft spacing' framework to penalize clustering of lower-quality content without hard filtering. Subsequent upgrades added PinCLIP multimodal visual embeddings and Semantic ID signals for more explicit semantic diversity control. Ablation studies showed removing DPP reduced time-spent impressions by over 2% after one week. Future work includes a unified generative post-ranking model and reinforcement learning-based value models.