Methods and systems for reducing false positives for financial transaction fraud monitoring using artificial intelligence

Abstract

Systems and methods for reducing false positives for financial transaction fraud monitoring using machine learning techniques. Using an original model for separating transactions into high risk and low risk categories for fraud, transactions falling into the high-risk category may be labeled as a false positive or a true positive. The labels and data associated with the transactions may be used to train two or more false positive reduction models (FPRMs) using iterative machine learning techniques. Once training is complete, a future transaction may be processed using the original model, and, if the original model indicates that the future transaction is high risk, data associated with the future transaction may be processed by the trained FPRM(s), which may determine whether the future transaction is at a high risk or a low risk of being fraudulent.

Having thus described one or more embodiments of the invention, what is claimed as new and desired to be protected by Letters Patent includes the following: 
     
         1 . A computer-implemented method performed by a server system, the method comprising:
 receiving payment transaction data associated with transactions previously identified as potentially fraudulent and accuracy labels preselected for each of the transactions, wherein the accuracy labels identify the transactions as a false positive or a true positive;   feeding the payment transaction data and the accuracy labels to a first false positive reduction model (FPRM) to classify a first subset of the transactions with a high risk of being fraudulent and a second subset of the transactions with a low risk of being fraudulent;   conducting supervised training of one or more subsequent FPRMs to produce one or more trained subsequent FPRMs by inputting the accuracy labels and the payment transaction data associated with one or more training subsets of the transactions into the one or more subsequent FPRMs in an iterative or recursive manner until a desired accuracy metric is satisfied by output of the one or more subsequent FPRMs, wherein an initial one of the one or more training subsets comprises the first subset and any subsequent ones of the one or more training subsets comprise those of the transactions classified with high risk of being fraudulent by the immediately preceding one of the one or more subsequent FPRMs; and   feeding future payment transaction data associated with a future transaction into at least one of the one or more trained subsequent FPRMs to classify the future transaction's risk of being fraudulent.   
     
     
         2 . The computer-implemented method of  claim 1 , wherein the transactions previously identified as potentially fraudulent are identified by feeding initial payment transaction data associated with a plurality of customers from a transaction database into an original model, wherein the original model classifies the initial payment transaction data into an original high-risk category or an original low risk category, wherein transactions in the original high-risk category are the transactions previously identified as potentially fraudulent. 
     
     
         3 . The computer-implemented method of  claim 2 , wherein at least one of the original model, the first FPRM, and the one or more of the subsequent FPRMs are trained using machine learning analysis. 
     
     
         4 . The computer-implemented method of  claim 3 , wherein at least one of the original model, the first FPRM, and the one or more of the subsequent FPRMs are trained using supervised machine learning algorithms. 
     
     
         5 . The computer-implemented method of  claim 2 , wherein the original model classifies transactions associated with the initial payment transaction data as at high risk of being fraudulent or at low risk of being fraudulent based on at least one of: (a) transaction amount, (b) online versus offline transactions for the corresponding one of the plurality of customers, (c) Point of Sale (POS) versus Barcode based transactions for the corresponding one of the plurality of customers, and (d) market-specific fraud patterns. 
     
     
         6 . The computer-implemented method of  claim 2 , wherein the initial payment transaction data is associated with a plurality of initial transactions over a given period of time. 
     
     
         7 . The computer-implemented method of  claim 1 , wherein the future transaction is identified as having a high risk of being fraudulent by an original model based on one or more thresholds prior to the step of feeding the future payment transaction data associated with the future transaction into the at least one of the one or more trained subsequent FPRMs. 
     
     
         8 . The computer-implemented method of  claim 7 , wherein the classification of the future payment transaction by the at least one of the one or more trained subsequent FPRMs is based on or in addition to the identification by the original model. 
     
     
         9 . A computing system for reducing false positives for financial transaction fraud monitoring using machine learning techniques, the system comprising one or more processors individually or collectively programmed to:
 receive payment transaction data associated with transactions previously identified as potentially fraudulent and accuracy labels preselected for each of the transactions, wherein the accuracy labels identify each of the transactions as a false positive or a true positive;   feed the payment transaction data and the accuracy labels to a first false positive reduction model (FPRM) to classify a first subset of the transactions with a high risk of being fraudulent and a second subset of the transactions with a low risk of being fraudulent;   conduct supervised training of one or more subsequent FPRMs to produce one or more trained subsequent FPRMs by inputting the accuracy labels and the payment transaction data associated with one or more training subsets of the transactions into the one or more subsequent FPRMs in an iterative or recursive manner until a desired accuracy metric is satisfied by output of the one or more subsequent FPRMs, wherein an initial one of the one or more training subsets comprises the first subset and any subsequent ones of the one or more training subsets comprise those of the transactions classified with high risk of being fraudulent by the immediately preceding one of the one or more subsequent FPRMs;   feed future payment transaction data of a future transaction to an original model, wherein the original model uses one or more thresholds to determine an original model output, wherein the original model output classifies the future transaction as at high risk of being fraudulent or at low risk of being fraudulent; and   automatically feed the future payment transaction data, when the original model output identifies that the future transaction is at high risk of being fraudulent, to at least one of the one or more trained subsequent FPRMs to further classify the future transaction's risk of being fraudulent.   
     
     
         10 . The computing system of  claim 9 , wherein the transactions previously identified as potentially fraudulent are identified by feeding initial payment transaction data associated with a plurality of customers from a transaction database into the original model, wherein the original model delineates the initial payment transaction data into an original high-risk category or an original low risk category, wherein transactions in the original high-risk category are the transactions previously identified as potentially fraudulent. 
     
     
         11 . The computing system of  claim 10 , wherein the initial payment transaction data is associated with a plurality of initial transactions over a given period of time. 
     
     
         12 . The computing system of  claim 9 , wherein at least one of the original model, the first FPRM, and the one or more of the subsequent FPRMs are trained using machine learning analysis. 
     
     
         13 . The computing system of  claim 12 , wherein at least one of the original model, the first FPRM, and the one or more of the subsequent FPRMs are trained using supervised machine learning algorithms. 
     
     
         14 . The computing system of  claim 10 , wherein the original model classifies transactions associated with the initial payment transaction data as at high risk of being fraudulent or at low risk of being fraudulent based on at least one of: (a) transaction amount, (b) online versus offline transactions for the corresponding one of the plurality of customers, (c) Point of Sale (POS) versus Barcode based transactions for the corresponding one of the plurality of customers, and (d) market-specific fraud patterns. 
     
     
         15 . A non-transitory computer-readable storage media having computer-executable instructions for reducing false positives for financial transaction fraud monitoring using machine learning techniques, wherein when executed by at least one processor, the computer-executable instructions cause the at least one processor to:
 access payment transaction data associated with a plurality of customers from a transaction database, the payment transaction data comprising information of transactions previously performed by the plurality of customers;   feed the payment transaction data to an original model, wherein the original model scores the transactions and classifies the transactions into an original high-risk category or an original low-risk category based on the scores, wherein the original high-risk category is associated with transactions having a high risk of fraud and the original low-risk category is associated with transactions having a low risk of fraud;   receive accuracy labels selected for each of the transactions in the original high-risk category, wherein the accuracy labels identify each of the transactions in the original high-risk category as a false positive or a true positive;   train a first false positive reduction model (FPRM) based, at least in part, on the transactions in the original high-risk category and the accuracy labels, wherein the first FPRM classifies a first training subset of the transactions in the original high-risk category with a high risk of being fraudulent and a second training subset of the transactions in the original high-risk category with a low risk of being fraudulent;   train a second FPRM based, at least in part, on the first training subset of the transactions and the corresponding accuracy labels, wherein the second FPRM classifies a third training subset of the transactions of the first training subset with a high risk of being fraudulent and a fourth training subset of the transactions of the first training subset with a low risk of being fraudulent;   generate a trained FPRM from at least one of the first FPRM and the second FPRM;   feed data associated with a future transaction into the original model, wherein the original model classifies the future transaction as having a high risk of fraud; and   based on the classification of the future transaction as having a high risk of fraud, feed the data associated with the future transaction to the trained FPRM, wherein the trained FPRM classifies the future transaction as high risk for being fraudulent or low risk for being fraudulent.   
     
     
         16 . The non-transitory computer-readable storage media of  claim 15 , wherein the trained FPRM is generated based on determining a desired accuracy metric is satisfied by an output of the first FPRM, an output of the second FPRM, or an output of a subsequent FPRM trained using output from the second FPRM. 
     
     
         17 . The non-transitory computer-readable storage media of  claim 15 , wherein the trained FPRM is generated based on one or more subsequent FPRMs receiving output from previous FPRMs in an iterative or recursive manner, beginning with an output from the second FPRM being fed into a third FPRM. 
     
     
         18 . The non-transitory computer-readable storage media of  claim 15 , wherein the transactions previously performed by the plurality of customers are collected over a period of time. 
     
     
         19 . The non-transitory computer-readable storage media of  claim 15 , wherein at least one of the original model, the first FPRM, and the second FPRM are trained using supervised machine learning algorithms. 
     
     
         20 . The non-transitory computer-readable storage media of  claim 15 , wherein the original model classifies the future transaction based on at least one of: (a) transaction amount, (b) online versus offline transactions for the corresponding one of the plurality of customers, (c) Point of Sale (POS) versus Barcode based transactions for the corresponding one of the plurality of customers, and (d) market-specific fraud patterns.