Abstract
A system (200) for point-in-time reconstruction of time-decayed behavioral embedding accumulator states for payment entities includes an append-only accumulator delta log (202) that records, for each accumulator update event, an entity identifier, an exact authorization timestamp, a D-dimensional embedding delta, a pre-computed exponential decay factor, and the resulting accumulator weight. A checkpoint store (204) periodically snapshots the accumulator state including a decayed sum, a total weight, and a timestamp for each entity. A reconstruction engine (206) reads the latest checkpoint prior to a query timestamp, replays all delta log entries for the entity with timestamp strictly less than the query timestamp using the pre-stored decay factors, and returns the exact accumulator state as it existed at the query timestamp. A training set construction application programming interface applies the reconstruction engine (206) to an entire training transaction set (208), associating each transaction with the embedding for its entity at the instant before the transaction's authorization timestamp, enforcing temporal causality in training feature construction and eliminating the embedding-feature component of training-serving skew.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kala, Durga S.; Huang, Xinrong; Telukunta, Nishanth; and Kanichetty, Sreekanth, "SYSTEM AND METHOD FOR TEMPORALLY-CORRECT POINT-IN-TIME EMBEDDING RECONSTRUCTION FROM APPEND-ONLY ACCUMULATOR DELTA LOGS", Technical Disclosure Commons, (September 16, 2026)
https://www.tdcommons.org/dpubs_series/11748