Abstract: Systems and methods are disclosed for tuning a generative artificial intelligence (AI) model based on a knowledge base. Instead of manually generating questions relevant to the knowledge base, providing those questions to the generative AI model, and manually reviewing the answers generated by the generative AI model in order to tune the generative AI model over many iterations, a natural language processing model may be configured to leverage the knowledge base to automatically generate questions and answers based on the knowledge base. In this manner, the natural language processing model is able to generate tuning data that may be used to automatically tune the generative AI model. The systems and methods also disclose automatic tuning of the generative AI model, including testing and feedback that may be used to improve tuning of the generative AI model.

Abstract: Systems and methods are disclosed for managing a generative artificial intelligence (AI) model. Managing the generative AI model may include training or tuning the generative AI model before use or managing the operation of the generative AI model during use. Training or tuning a generative AI model typically requires manual review of outputs from the model based on the queries provided to the model to reduce hallucinations generated by the generative AI model. Once the model is in use, though, hallucinations still occur. Use of a confidence (whose generation is described herein) to train or tune the generative AI model and/or manage operation of the model reduces hallucinations, and thus improves performance, of the generative AI model.

Abstract: This disclosure provides systems, methods and apparatuses for identifying fraudulent accounts associated with an electronic payment service. In some implementations, a computing device may retrieve a data set including a number of attributes for each of a multitude of accounts, and may construct a plurality of different graphs each based on a unique set of the attributes. Each graph may include a plurality of nodes linked together by a multitude of edges, where each node identifies a corresponding account and each edge indicates one or more of the corresponding attributes that are common to a pair of accounts. The computing device may determine a likelihood of each graph correctly identifying fraudulent accounts by analyzing groups of nodes connected to each other by corresponding groups of edges using historical account data, and may select the graph having the greatest determined likelihood to predict whether any of the accounts is fraudulent.

Abstract: This disclosure relates to systems and methods for identifying risky merchants associated with an electronic payment service. In some implementations, a risk assessment system receives a set of features indicative of one or more risks posed by a merchant enrolled in the electronic payment service, where each feature of the set of features indicative of one or more financial attributes of the merchant. The risk assessment system determines a risk score for the merchant based on the set of features using a trained machine learning model, determines a Shapely additive explanation (SHAP) score for each feature of the set of features, and then divides the set of features into multiple groups of features based on a mapping between the features and their respective indicated financial attributes.

Abstract: This disclosure provides systems, methods and apparatuses for identifying fraudulent accounts associated with an electronic payment service. In some implementations, a computing device may retrieve a data set including a number of attributes for each of a multitude of accounts, and may construct a plurality of different graphs each based on a unique set of the attributes. Each graph may include a plurality of nodes linked together by a multitude of edges, where each node identifies a corresponding account and each edge indicates one or more of the corresponding attributes that are common to a pair of accounts. The computing device may determine a likelihood of each graph correctly identifying fraudulent accounts by analyzing groups of nodes connected to each other by corresponding groups of edges using historical account data, and may select the graph having the greatest determined likelihood to predict whether any of the accounts is fraudulent.

Abstract: Techniques are disclosed herein for determining variable importance on a predictive model on a case level. Modeling data associated with a case is received. The modeling data provides input variables, each having a corresponding value for input to a predictive modeling technique associated with the case. A measure of impact for each of the variables is determined using an input shuffling method. Variables having a measure of impact that exceeds a specified threshold are identified. A summary that includes the identified variables is generated.