From Activity to Intent: Generating User Journeys with LLMs

Lin Zhu, Manan Kalra, Byung Eun Jeon, Ye Liu, Xiangyi Chen, Jaewon Yang, Jinwen Xu, Tingting Zhu, Sudarshan Lamkhede16 Oct 2026

Problem

Recommenders usually optimize the next item or a durable interest such as "Home Decor". Users pursue projects over weeks: a remodel, a garden, a wedding. Pinterest's first journey system clustered keywords from search queries, engaged Pins and boards, named each cluster by its top keyword and ranked the clusters. It fragmented one goal into several journeys, fell back to generic names ("Art", "Shopping") on messy clusters and let a few meme clicks take a top slot. Fragments of one goal competed for the same notification slot, and generic names gave nothing to personalize on.

Example from the article: cabinets, countertops and backsplash searches ended up as three separate journeys instead of one "Modern Farmhouse Kitchen".

Solution in brief

Treat journey inference as one generation step: a prompt holds a short profile plus a chronological activity log, and the model returns a ranked JSON list of journey names. A frontier model produces training targets, and a fine-tuned 4B Qwen3 student serves production. The prompt is the ranking and safety policy.

Questions this case answers
  • How to infer a user's current multi-session goals from a noisy chronological event log when clustering fragments one goal into several topics and falls back to generic labels
  • How to serve a fine-tuned small language model for batch scoring of a very large user base at acceptable cost
  • How to infer what goal a user is working on from a long activity log, not just which topics they like?
  • How to replace a multi-stage keyword clustering pipeline with a single LLM call without losing control over quality and safety?

Tools and methods

  • Qwen3 (0.6B to 8B), SFT with LoRA — student models trained on teacher outputs
  • Frontier LLM as teacher — synthetic targets, with extra reasoning tokens allowed
  • vLLM behind NVIDIA Dynamo, L40S GPUs — online serving with continuous batching and prefix caching
  • Semantic IDs (RQ-VAE) — experimental compact item tokens
  • LLM judge — scores relevance, support, diversity and granularity of emitted journeys
Continue reading — it's free

Google or a one-time email code. No password.