AI
Pinterest Replaces Clustering with 4B LLM for User Journey Detection, Lifts Engagement Metrics
The social platform switched from multi-stage clustering to a fine-tuned Qwen3 model to infer user goals across sessions, achieving measurable gains in email and push notification performance.
InfoQ
Pinterest has migrated its user-journey inference system from a traditional clustering pipeline to a fine-tuned 4-billion-parameter language model. The shift has yielded concrete improvements: a 1.1% lift in email click-through rates and 1.3% higher push notification open rates, alongside statistically significant gains in overall engagement and retention.
How Pinterest Defines User Journeys#
Rather than treating individual searches and saves as isolated interests, the platform now synthesises them into longer-term goals spanning multiple sessions. A user renovating a kitchen, for instance, might search for cabinet styles, save countertop images, and later explore backsplash or lighting options. The system recognizes these diverse actions as part of a coherent project—what Pinterest calls a "journey"—that bridges interests, intent, and context.
From Clustering to Generation#
Pinterest's previous approach extracted keywords from activity histories, clustered them using embeddings and hierarchical methods, named the clusters, and ranked them through additional models and diversity heuristics. While an improvement over interest-only signals, the method had limitations: related activities could split into separate journeys, labels became generic, or casual engagement masqueraded as intent.
The new system reframes journey inference as a text generation task. A fine-tuned LLM ingests a chronological record of up to 150 user actions over a 360-day window, paired with a compact user profile, and outputs a ranked list of journey names. The model weights searches and saves more heavily than incidental clicks; the prompt itself encodes rules to favour repeated engagement, recent activity, and specificity while discouraging unrelated mergers, suppressing noise, and permitting empty results when evidence is thin.
Training on Synthetic Data#
To avoid expensive manual labelling, Pinterest used frontier models to generate synthetic training examples from real user activity histories. These examples fine-tuned smaller open-weight Qwen3 models via supervised fine-tuning and low-rank adaptation (LoRA). Testing revealed the 4B-parameter variant struck the best balance between inference quality and serving throughput.

