Detecting fraud using machine-learning

A non-transitory computer-readable medium having program instructions stored thereon that are executable by a computer system to perform operations comprising:

• sending, to a first recipient account, a first message containing a first link to a first electronic resource of a plurality of electronic resources, wherein the first electronic resource is associated with a first sender account of the computer system;
• receiving a request to access the first electronic resource via the first link;
• and
• before granting the request to access the first electronic resource, evaluating the request to access the first electronic resource using a fraud detection model, wherein the fraud detection model is a graph embedding model generated using machine learning, including by:
• accessing labeled training data that includes request information about a plurality of previous requests, wherein at least a portion of the request information in the labeled training data is encoded using one-hot encoding, and wherein each of the plurality of previous requests (a) is a request to access one of the plurality of electronic resources and (b) is associated with a respective sender account of the computer system and a respective recipient account;
• and
• for a number of iterative training epochs:
• for each of the plurality of previous requests,
• sequentially generating embedding values corresponding to the respective sender account and the respective recipient account associated with that previous request,
• wherein an embedding value corresponding to the respective recipient account associated with that previous request is generated using an embedding value corresponding to the respective sender account associated with that previous request,
• wherein each embedding value comprises a vector that represents a particular sender account or a particular recipient account within the fraud detection model using a reduced number of dimensions relative to a number of dimensions of the request information,
• using a classification loss function, calculating a loss value for that previous request based on the embedding value corresponding to the respective recipient account;
• and
• updating one or more parameters of the fraud detection model based on the loss value for that previous transaction;
• wherein evaluating the request to access the first electronic resource using the fraud detection model includes:
• generating a first embedding value for the first recipient account based on the request to access the first electronic resource and one or more embedding values included in the graph embedding model;
• and
• generating a prediction score for the first recipient account using the first embedding value;
• and wherein the operations further comprise:
• determining whether to accept the request to access the first electronic resource using the prediction score.

A system, comprising:

• at least one processor;
• and
• a memory having instructions stored thereon that are executable by the at least one processor to cause the system to:
• send, to a first recipient account, a first message containing a first link to a first electronic resource of a plurality of electronic resources, wherein the first electronic resource is associated with a first sender account of the system;
• receive a request to access the first electronic resource via the first link;
• and
• before granting the request to access the first electronic resource, evaluate the request to access the first electronic resource using a fraud detection model, wherein the fraud detection model is a graph embedding model generated using machine learning, including by:
• accessing labeled training data that includes request information about a plurality of previous requests, wherein at least a portion of the request information in the labeled training data is encoded using one-hot encoding, and wherein each of the plurality of previous requests (a) is a request to access one of the plurality of electronic resources and (b) is associated with a respective sender account of the system and a respective recipient account;
• and
• for a number of iterative training epochs:
• for each of the plurality of previous requests,
• sequentially generate embedding values corresponding to the respective sender account and the respective recipient account associated with that previous request,
• wherein an embedding value corresponding to the respective recipient account associated with that previous request is generated using an embedding value corresponding to the respective sender account associated with that previous request,
• wherein each embedding value comprises a vector that represents a particular sender account or a particular recipient account within the fraud detection model using a reduced number of dimensions relative to a number of dimensions of the request information,
• use a classification loss function to calculate, a loss value for that previous request based on the embedding value corresponding to the respective recipient account;
• and
• update one or more parameters of the fraud detection model based on the loss value for that previous transaction;
• wherein evaluating the request to access the first electronic resource using the fraud detection model includes:
• generating a first embedding value for the first recipient account based on the request to access the first electronic resource and one or more embedding values included in the graph embedding model;
• and
• generating a prediction score for the first recipient account using the first embedding value;
• and wherein the instructions are further executable to cause the system to:
• determine whether to accept the request to access the first electronic resource using the prediction score.