Performing LLM inference on each user’s transaction history scales linearly with the number of users, making the cost prohibitive for large‑scale finance. We therefore shift attribute inference from a per‑user to a per‑transaction‑pattern granularity and build a three‑phase pipeline:
- Resolve: disambiguate abstract item names, optionally grounding them on the web;
- Profile: for each frequent pattern a single LLM call emits predefined categorical labels, free‑text attributes, and prevalence estimates for each attribute;
- Tag: cluster free‑text attributes into a queryable database.
Since inference runs over patterns rather than users, the budget grows with pattern count instead of user count. On the public Open e‑commerce corpus the resulting database is statistically indistinguishable from an LLM that reads each user’s raw history directly, achieving comparable AUC across evaluated attributes, while prevalence estimates provide discriminative signals between positive and negative users.
The pipeline is deployed at a major Japanese bank serving tens of millions of users, achieving roughly three orders of magnitude reduction in LLM inference targets compared with a per‑user pipeline. The code is publicly available at https://github.com/CyberAgentAILab/profiling-agent-open-ecommerce.
Review: By abstracting to pattern‑level inference, the approach dramatically cuts LLM costs without sacrificing profiling accuracy, offering a practical solution for large‑scale user profiling in the banking sector.