| CPC G06Q 20/4016 (2013.01) [G06N 3/04 (2013.01); G06N 3/08 (2013.01)] | 16 Claims |
|
1. A computer-implemented method comprising:
accessing, by a server system, a set of payment transaction features corresponding to a plurality of historical payment transactions from a transaction database;
determining, by the server system via a fraud model, a first set of rank-ordered payment transaction features based, at least in part, on a first set of shapley additive explanations (SHAP) values, each of the first set of rank-ordered payment transaction features indicating a contribution of a corresponding payment transaction feature in predicting whether a payment transaction from the plurality of historical payment transactions is fraudulent or non-fraudulent;
determining, by the server system via an approval model, a second set of rank-ordered payment transaction features based, at least in part, on a second set of SHAP values, each of the second set of rank-ordered payment transaction features indicating the contribution of a corresponding payment transaction feature in predicting whether a payment transaction from the plurality of historical payment transactions is approved or declined;
computing, by the server system, a difference in ranks of payment transaction features based, at least in part, on the first set of rank-ordered payment transaction features and the second set of rank-ordered payment transaction features;
determining, by the server system, a set of utilized transaction features and a set of unutilized transaction features based, at least in part, on the difference in the ranks of payment transaction features;
generating, by the server system, a simulated authorizing model, based at least in part on the set of utilized transaction features;
computing, by the server system, a simulated transaction approval rate and simulated fraud transaction rate for the simulated authorizing model based, at least in part, on the plurality of historical payment transactions;
generating, by the server system, a plurality of proxy authorization models based, at least in part, on the set of unutilized transaction features;
computing, by the server system, transaction approval rates and fraud transaction rates for each of the plurality of proxy authorization models based, at least in part, on the plurality of historical payment transactions;
computing, by the server system, an increase in transaction approval rate and a change in fraud transaction rate for each of a plurality of proxy transaction approval models based, at least in part, on comparing the transaction approval rates and the fraud transaction rates with the simulated transaction approval rate and a simulated fraud transaction approval rate;
determining, by the server system, one or more recommended transaction features from the set of unutilized transaction features upon determining that the increase in the transaction approval rate for the one or more recommended transaction features and the change in fraud transaction rate for the one or more recommended transaction features are more than a predetermined threshold;
transmitting, by the server system, the one or more recommended transaction features to an authorizing entity;
utilizing, by the authorizing entity, the one or more recommended transaction features to determine whether a transaction is fraudulent and disapproved, or is approved;
receiving, by the server system, a payment authorization request for a real-time payment transaction between a cardholder and a merchant;
accessing, by the server system, payment transaction attributes based, at least in part, on the payment authorization request;
determining, by the server system via a machine learning model, a set of transaction approval rates for the real-time payment transaction based, at least in part, on the payment transaction attributes, each of the set of transaction approval rates indicating a transaction approval rate when one or more authorizing components are applied to the real-time payment transaction, wherein the machine learning model is a Deep Factorization Machine (FM) model including
an FM layer that generates FM outputs by modelling low-order feature interactions based on the payment transaction attributes,
a deep neural network (DNN) that generates DNN outputs by modelling high-order and low-order feature interactions based on the payment transaction attributes, and
an output layer that combines the FM outputs and the DNN outputs to generate a final output;
generating, by the server system, an authorizing component recommendation based, at least in part, on a set of transaction approval rates; and
transmitting, by the server system, the authorizing component recommendation to the authorizing entity.
|