Hirotake Ito

This work re-cast attribute inference from per-user to per-transaction-pattern, which results in close to a three-order-of-magnitude reduction in LLM inference targets versus a per-user pipeline.