US10657525B2 United States Inventor: Pablo Ariel Duboue Current Assignee: Kasisto Inc Events: 2017-06-27 Application filed by Kasisto Inc 2018-12-27 Publication of US20180374089A1 2020-05-19 Application granted Status: Active
NOTE: Patents are a thorny topic with respect to their interference with FLOSS. In this patent and others I have invented, the decision to patent was taken by my customer (or employer) at that time. On the positive side, this research wouldn't have been made public otherwise, as it had competitive value to its assignee.
Transaction strings as generated from credit cards data are short, telegraphic strings describing the vendor. Mapping them to transaction categories is a recurring problem in the banking industry. The technique described in this patent creates signatures using subsequences of characters from the transaction strings and then further conflates them using a deep autoencoder.
Using this method, it is possible to train a system to recognize that that transaction signatures with the words 'CAFE' are more likely related to the category 'Food&Beverages'.
See the patent text at Google Patents.