Soft segmentation based rules optimization for zero detection loss false positive reduction
AssigneeFair Isaac Corporation
InventorsScott Michael Zoldi, Qing Lin
Life of this patent
Estimating the term…
Abstract
A system and method includes soft-segment based rules optimization that can mitigate the overall false positives while maintaining 100% true positive detection. The soft clustering allows real-time re-assignment of an account to a dominate archetype behavior, as well as rule optimization based on a logical order with more relaxation on thresholds for the most inefficient rules is performed within each archetype. The rule optimization provides false positive reduction compared to a baseline rule system. The method can be used to reduce false positives for any rule-based detection system in which the same true positive detection is required.
Claims24
A method of reducing false positives in financial crime detection while maintaining the same true positive detected in an existing financial crime detection system, the method comprising:
- defining, by at least one computer processor, a unique dictionary of words within a set of transaction data representing transactions associated with accounts, the transaction data comprising categorical field data and continuous field data, each word defined in the dictionary comprising transaction data from one or more categorical fields and one or more continuous fields for a given customer for a given transaction;
- soft-segmenting, by the at least one computer processor using at least one topic model, the dictionary of words to determine semantic structures underlying the one or more transactions, and to determine entity behavior archetypes associated with the one or more accounts based on the determined semantic structures according to the topic model, the entity behavior archetypes representing transactional behavior patterns across all of the one or more accounts based on the transaction data, the topic model being trained using only non-financial crime accounts of the one or more accounts, the soft-segmenting producing a global learned distribution for each word from the defined dictionary in an archetype space such that words in the dictionary are associated with archetypes at different probabilities;
- storing, by the computer processor, the entity behavior archetypes in a data store accessible by the computer processor, the storing including associating the entity behavior archetypes with each associated one or more accounts and associated transaction data;
- deriving, by the at least one computer processor after receiving new transaction data in an online environment, updated latent archetypes from the new transaction data and previous latent archetypes in the archetype space according to the entity behavior archetypes stored in the data store, the deriving comprising executing a Bayesian inference algorithm;
- and
- storing the updated latent archetypes in the data store, the updated latent archetypes being associated with the one of the one or more accounts being available to the system.
A system for reducing false positives in financial crime detection while maintaining the same true positive detected in an existing financial crime detection system, the system comprising
- a programmable processor;
- and
- a non-transitory, machine-readable medium storing instructions that, when executed by the processor, cause the at least one programmable processor to perform operations comprising:
- defining, by at least one computer processor, a unique dictionary of words within a set of transaction data representing transactions associated with accounts, the transaction data comprising categorical field data and continuous field data, each word defined in the dictionary comprising transaction data from one or more categorical fields and one or more continuous fields for a given customer for a given transaction;
- soft-segmenting, by the at least one computer processor using at least one topic model, the dictionary of words to determine semantic structures underlying the one or more transactions, and to determine entity behavior archetypes associated with the one or more accounts based on the determined semantic structures according to the topic model, the entity behavior archetypes representing transactional behavior patterns across all of the one or more accounts based on the transaction data, the topic model being trained using only non-financial crime accounts of the one or more accounts, the soft-segmenting producing a global learned distribution for each word from the defined dictionary in an archetype space such that words in the dictionary are associated with archetypes at different probabilities;
- storing, by the computer processor, the entity behavior archetypes in a data store accessible by the computer processor, the storing including associating the entity behavior archetypes with each associated one or more accounts and associated transaction data;
- deriving, by the at least one computer processor after receiving new transaction data in an online environment, updated latent archetypes from the new transaction data and previous latent archetypes in the archetype space according to the entity behavior archetypes stored in the data store, the deriving comprising executing a Bayesian inference algorithm;
- and
- storing the updated latent archetypes in the data store, the updated latent archetypes being associated with the one of the one or more accounts being available to the system.
Dependencies are inferred from the claim text.