System, method, and computer program product for interpreting black box models by perturbing transaction parameters

A system comprising a transaction processing system programmed or configured to:

• train a machine-learning model of the transaction processing system on historical authorization decisions of a first issuer system, the transaction processing system different from the first issuer system;
• receive a transaction request associated with a payment transaction initiated by a payment device of a user, the payment transaction between the user and a first merchant system, the payment transaction comprising a plurality of transaction parameters;
• generate an authorization request associated with the payment transaction;
• communicate the authorization request to the first issuer system that issued the payment device;
• determine that the first issuer system failed to return an authorization decision in response to the authorization request, wherein the first issuer system failed to return the authorization decision due to the first issuer system being at least temporarily unable to communicate with the transaction processing system;
• in response to determining that the first issuer system failed to return the authorization decision, automatically generate, with the machine-learning model, a first risk score by inputting the plurality of transaction parameters of the payment transaction to the machine-learning model, the machine-learning model programmed or configured to generate the risk score without providing at least one interpretable reason for the risk score;
• based on the risk score, generate a decline decision as the authorization decision, the decline decision generated on behalf of the first issuer system;
• the system further comprising a model interpretation network programmed or configured to:
• receive an inquiry request message identifying the payment transaction;
• input, to the model interpretation network, the plurality of transaction parameters of the payment transaction;
• for each transaction parameter of the plurality of transaction parameters of the payment transaction, perturb, with the model interpretation network, a value of the transaction parameter and reanalyze the payment transaction with the machine-learning model to generate a perturbed risk score based on the perturbed transaction parameter;
• determine at least one impact parameter from the plurality of transaction parameters by comparing the perturbed risk scores generated for each of the plurality of transaction parameters;
• and
• generate an inquiry response message based on the at least one impact parameter, the inquiry response message comprising at least one interpretable reason for the decline decision associated with the determined at least one impact parameter;
• the system further configured to:
• based on the at least one interpretable reason for the decline decision, determine that the decline decision was based on an error of the machine-learning model;
• and
• modify the machine-learning model to remedy the error, the modified machine-learning model used for making authorization decisions for future payment transactions associated with the first issuer system.

A computer program product comprising at least one non-transitory computer-readable medium including program instructions that, when executed by at least one processor, cause the at least one processor comprising a transaction processing system and/or a model interpretation network to:

• train a machine-learning model of the transaction processing system on historical authorization decisions of a first issuer system, the transaction processing system different from the first issuer system;
• receive a transaction request associated with a payment transaction initiated by a payment device of a user, the payment transaction between the user and a first merchant system, the payment transaction comprising a plurality of transaction parameters;
• generate an authorization request associated with the payment transaction;
• communicate the authorization request to the first issuer system that issued the payment device;
• determine that the first issuer system failed to return an authorization decision in response to the authorization request, wherein the first issuer system failed to return the authorization decision due to the first issuer system being at least temporarily unable to communicate with the transaction processing system;
• in response to determining that the first issuer system failed to return the authorization decision, automatically generate, with the machine-learning model, a first risk score by inputting the plurality of transaction parameters of the payment transaction to the machine-learning model, the machine-learning model programmed or configured to generate the risk score without providing at least one interpretable reason for the risk score;
• based on the risk score, generate a decline decision as the authorization decision, the decline decision generated on behalf of the first issuer system;
• receive an inquiry request message identifying the payment transaction;
• input, to the model interpretation network, the plurality of transaction parameters of the payment transaction;
• for each transaction parameter of the plurality of transaction parameters of the payment transaction, perturb, with the model interpretation network, a value of the transaction parameter and reanalyze the payment transaction with the machine-learning model to generate a perturbed risk score based on the perturbed transaction parameter;
• determine at least one impact parameter from the plurality of transaction parameters by comparing the perturbed risk scores generated for each of the plurality of transaction parameters;
• generate an inquiry response message based on the at least one impact parameter, the inquiry response message comprising at least one interpretable reason for the decline decision associated with the determined at least one impact parameter;
• based on the at least one interpretable reason for the decline decision, determine that the decline decision was based on an error of the machine-learning model;
• and
• modify the machine-learning model to remedy the error, the modified machine-learning model used for making authorization decisions for future payment transactions associated with the first issuer system.