Pinterest Replaces User Journey Clustering with Fine-Tuned 4B Qwen3 LLM

Transition to LLM-Based Journey Inference

Pinterest has updated its system for inferring user journeys, moving from a multi-stage clustering pipeline to a fine-tuned 4-billion-parameter Qwen3 language model. This shift aims to better understand users' long-term goals by analyzing their activity histories, rather than just grouping keywords into interests. The identified journeys are then fed into Pinterest's existing recommendation infrastructure.

Improved Engagement Metrics

Following the implementation of the new LLM-based system, Pinterest reported a 1.1% improvement in email click-through rates and a 1.3% increase in push notification opens. These metrics indicate a more effective identification of user intent and context, leading to more relevant user engagement.

From Clustering to Generation Task

The previous system extracted keywords, grouped them using embeddings and hierarchical clustering, and then ranked the resulting clusters. While effective, it could fragment related activities or produce generic labels. The new system treats journey inference as a generation task: a fine-tuned LLM processes a chronological history of up to 150 user actions over a 360-day period, along with a user profile, to generate a ranked set of journey names. Search and save actions are weighted more heavily, and the model considers repeated engagement, recency, and specificity.

Synthetic Data for Training

To train the new system without requiring extensive human-labeled datasets, Pinterest utilized frontier models to create synthetic training examples from its existing user activity histories. These synthetic examples were then used to fine-tune smaller open-weight Qwen3 models through supervised fine-tuning and low-rank adaptation (LoRA).