From "Who Is This User?" to "What Does This Purchase Mean?": A Deployed Pipeline for Semantic User Profiling at Bank Scale

The paper presents a three‑phase pipeline that shifts semantic user profiling from per‑user to per‑transaction‑pattern inference, dramatically reducing LLM inference costs. The Resolve phase maps item names to abstracts, Profile generates categorical labels, free‑text attributes, and prevalence estimates per pattern, and Tag clusters these attributes into a searchable database. Evaluated on a public e‑commerce corpus, the pattern‑based approach matches the performance of a per‑user LLM while enabling deployment at a major Japanese bank with tens of millions of users and a three‑order‑of‑magnitude reduction in inference targets.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms
• agents