TechPrufer
Get started
Pinterest
Pinterest logoPinterest · Feature trail

User Interest Clustering for Recommendation Systems

8 posts · May 2026 – Oct 2026

What this trail covers

Pinterest evolved its user journey inference system by replacing a multi-stage clustering pipeline with a fine-tuned 4-billion-parameter LLM. This generative approach directly infers a ranked list of journey names from chronological activity logs, improving specificity, coherence, and handling of multilingual contexts. The system leverages supervised fine-tuning and LoRA on open-weight models, with the 4B parameter student offering the best balance of quality and throughput. Serving optimizations for batch processing at notification scale include PagedAttention and continuous batching.

All Pinterest trails
  1. From Activity to Intent: Generating User Journeys with LLMsThis post details the transition from a legacy keyword clustering system to a generative LLM-based approach for user journey inference. It describes the prompt engineering for the LLM, including input formatting, action collapsing, and output constraints. It also covers the distillation process of a frontier teacher model into a 4B parameter student using SFT and LoRA, and the serving optimizations for batch processing at scale, including PagedAttention and continuous batching.
  2. Evolving Pinterest’s Embedding Retrieval PlatformThis post details the evolution of Pinterest's Manas embedding retrieval platform, focusing on three key areas: Quantization (Scalar Quantization and Product Quantization) to reduce memory footprint and serving costs, with benchmarks showing trade-offs between compression and recall, and the implementation of Linear Scaling SQ with SIMD intrinsics for compute efficiency; SSD-based Serving using SPANN with PQ quantization to enable serving ANN queries directly from SSD, demonstrating significant memory and CPU savings compared to in-memory solutions; and Multi-embedding Retrieval to move beyond the two-tower model's single-vector constraint, enabling late interaction models like ColBERT with "Sum of MaxSim" scoring logic and updates to the Manas query parser to handle multiple query embeddings.
  3. Evolving Pinterest’s Embedding Retrieval PlatformThis post details the evolution of Pinterest's Manas embedding retrieval platform. Key technical contributions include: 1. Quantization: Implementation and evaluation of Scalar Quantization (SQ) and Product Quantization (PQ), demonstrating trade-offs in compression vs. recall, and the use of SIMD intrinsics for Linear Scaling SQ to improve compute efficiency. 2. SSD-based Serving: Adoption of SPANN with PQ quantization for disk-based ANN serving, achieving significant memory reduction and improved QPS/latency compared to in-memory solutions like HNSW. 3. Multi-embedding Retrieval: Introduction of late interaction models like ColBERT, using "Sum of MaxSim" scoring, and updates to the Manas query parser to handle multiple query embeddings for richer similarity calculations.
  4. Scaling Conditional Learned Retrieval for Pinterest Home FeedThis post details the evolution of Pinterest's Conditional Learned Retrieval (CLR) system for home feed candidate generation. Key contributions include: expanding CLR to support Interest, Pin, and Board conditions; scaling model foundations with a Conditioned User Sequence Transformer and integrating the PinFM Foundation Model for deeper user sequence understanding; and redesigning the serving infrastructure with a Slot Architecture (condition_os, condition_id, condition_type_id) for condition-agnostic routing, replacing bespoke pipelines with a unified model. Pin tower improvements include using large ID embedding tables and semantic IDs derived from quantized content embeddings.
  5. Scaling Conditional Learned Retrieval for Pinterest Home FeedThis post details the evolution of Pinterest's Conditional Learned Retrieval (CLR) system for home feed candidate generation. Key contributions include: expanding CLR to support Interest, Pin, and Board conditions; scaling model foundations with a Conditioned User Sequence Transformer and integrating the PinFM Foundation Model for deeper user sequence understanding; and redesigning the serving infrastructure with a slot architecture (condition_os, condition_id, condition_type_id) for condition-agnostic routing, enabling Unified CLR models. Pin tower improvements include using large ID embedding tables and semantic IDs derived from quantized content embeddings.
  7. Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at PinterestThis post introduces User Interest Clusters (UICs) as a new stateful interest representation to augment sequential user understanding with holistic, persistent use-case representation for retention. It details the construction of UICs using Complete Linkage Agglomerative Hierarchical Clustering on engaged Pins within the OmniSage embedding space, allowing for personalized clustering, dynamic cluster counts, and stateful lifecycle metadata. The post also outlines the system-level integration of UICs across the recommendation serving stack (retrieval, scoring, ranking, blending) to drive retention by treating interests based on their maturity and fostering serendipitous discovery.
  8. Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at PinterestThis post details the introduction and implementation of User Interest Clusters (UICs) as a novel stateful interest representation for Pinterest's recommendation system. It describes the motivation for moving beyond engagement-only optimization to retention, the limitations of existing sequential models, and the design of UICs. Key technical contributions include: 1. Personalized clustering over engaged content only, making clusters semantically coherent to individual users. 2. Dynamic cluster count determination based on coherence thresholds, allowing for a variable number of interests per user. 3. Stateful lifecycle metadata (recency, frequency) for each UIC to reason about interest maturity. The core signal construction mechanism is detailed as Complete Linkage Agglomerative Hierarchical Clustering on OmniSage embeddings, with a worked example illustrating the process. The post also outlines the system-level integration of UICs across the home feed serving stack (retrieval, ranking, blending) to enable use-case-aware recommendations.
  9. Making User-Sequence Data More Cost-Efficient, Faster, and Easier to UseThis post details the redesign of Pinterest's user-sequence platform to improve cost-efficiency, speed, and usability. The core innovation is the 'one definition, many runtimes' principle, where signal and enrichment definitions are configured once in Python and then consistently applied across real-time indexing, batch indexing/backfill, and online serving. This approach uses a shared execution engine, a time-versioned store for low-latency updates, batch pipelines for historical data, columnar time-partitioned storage, and a serving API. Key design decisions include adopting configuration-as-code for sequence and enrichment definitions to ensure alignment between training and serving, and to simplify the onboarding of new event types and enrichments. The platform aims to provide a stable contract for downstream consumers, reduce storage and network usage, and support both real-time and batch production paths.
  10. Making User-Sequence Data More Cost-Efficient, Faster, and Easier to UseThis post details the redesign of Pinterest's user-sequence platform to improve cost-efficiency, speed, and usability. Key contributions include: 1. **Configuration-as-Code:** Defining sequence and enrichment logic in Python, compiled to JSON, for faster onboarding and improved reviewability. 2. **Shared Execution Engine:** A unified engine for both streaming and batch processing that reuses enrichment logic, minimizing code duplication and drift. 3. **Pluggable Executors:** Modular 'business logic' for specific event types within the shared engine. 4. **Unified Runtimes:** A 'one definition, many runtimes' approach ensuring consistency across training, batch, and serving. 5. **Columnar, Time-Partitioned Storage:** Optimized storage for efficient reads and writes. The platform aims to provide a stable contract for downstream ML consumers.
  11. Enhancing Ad Relevance: Integrating Real-Time Context into Sequential Recommender ModelsThis post details the integration of real-time context into a sequential recommender model for ad relevance. The key technical contributions include: a new model architecture with an integrated context layer in the query tower, a novel training approach using synthetic augmented data to inject pseudo-context, and a hybrid user embedding inference flow (offline for historical sequence, online for real-time context). This resulted in a 3x-10x increase in Recall@K and a ~275-300% median relevance improvement on the Related Pins surface, with a 1.08% overall ads relevance improvement and a 0.7% ROAS lift.