| CPC G06Q 20/4016 (2013.01) [G06N 20/00 (2019.01)] | 17 Claims |
|
1. A computer-implemented method comprising:
generating, with a machine-learning model, a risk score for each of a plurality of payment transactions processed over an electronic payment processing network, wherein the network comprises electronic communications among a merchant computing system, a transaction processing system, and an issuer computing system, wherein the plurality of payment transactions are initiated by payment device holders with the merchant computing system, wherein a first payment transaction among the plurality of payment transactions comprises an authorization decision comprising a decline authorization decision of the first payment transaction, wherein the decline decision is based on a first risk score generated by the machine-learning model without providing at least one interpretable reason for the first risk score, wherein the first risk score is based on an analysis of a plurality of transaction parameters of the first payment transaction, wherein at least one of the transaction parameters comprises a transaction category and wherein the transaction category comprises at least one of a merchant category code, the transaction type, and the transaction date;
inputting, to a model interpretation network, an inquiry request message identifying the first payment transaction and the first risk score, wherein the first payment transaction comprises the plurality of transaction parameters;
for each transaction parameter of the plurality of transaction parameters, including the transaction category:
perturbing, with the model interpretation network, a value of the transaction parameter based on the plurality of transaction parameters of the first payment transaction without input from other payment transactions, wherein the perturbing comprises iteratively perturbing a single transaction parameter of the plurality of transaction parameters at a time, wherein each iteration generates a set of perturbed transaction parameters which comprises the plurality of transaction parameters with one parameter perturbed;
inputting, with the model interpretation network, each set of perturbed transaction parameters to the machine-learning model;
generating, for each set of perturbed transaction parameters, with the machine-learning model, a perturbed risk score, wherein the generating generates a plurality of perturbed risk scores; and
generating, with the model interpretation network, a plurality of combined risk scores, wherein each combined risk score comprises the sum of each of the perturbed risk scores and the first risk score;
comparing, with the model interpretation network, each of the combined risk scores with the first risk score to analyze the magnitude of any change in the first risk score;
determining, with the model interpretation network, based on the change in the first risk score for each of the combined risk scores, at least one impact parameter from the plurality of transaction parameters, wherein a larger change indicates a larger impact on the first risk score; and
generating, with the model interpretation network, an inquiry response message based on the at least one impact parameter, wherein the response comprises at least one interpretable reason for the decline decision.
|