Paypal, Inc.

Vishal Sood

Sudhindra Murthy

Divakar Viswanathan

Hong Fan

Vidya Sagar Durga

Grahame Andrew Jastrebski

Nitin S Sharma

Chirag Gupta

Ashwin Maruti Hegde

Sheena Chawla

Yegya Narayanan Gopala Krishnan

There are provided systems and methods for a compute platform for machine leaning model roll-out. A service provider, such as an electronic transaction processor for digital transactions, may provide intelligent decision-making through decision services that execute machine learning models. When deploying or updating machine learning models in these engines and decision services, a model package may include multiple models, each of which may have an execution graph required for model execution. When models are tested from proper execution, the models may have non-performant compute items, such as model variables, that lead to improper execution and/or decision-making. A model deployer may determine and flag these compute items as non-performant and may cause these compute items to be skipped or excluded from execution. Further, the model deployer may utilize a pre-production computing environment to generate the execution graphs for the models prior to deployment or upgrading.

Compute platform for machine learning model roll-out

Techniques are disclosed for time constrained electronic request evaluation. A server system receives, from a computing device, a request submitted via an account, including a first set of characteristics associated with the request. The system executes a first machine-learning model to determine a first risk score for the request by inputting the first set of characteristics into the first model. The system generates an initial authentication decision for the request based on the first score and sends the decision to the device. The system executes a second, different machine-learning model to determine a second risk score for the request, by inputting the first set of characteristics and a second, different set of characteristics associated with the account into the second model. Based on the second score, the system determines a final authentication decision. The disclosed techniques may advantageously improve computer security and operations via identification of malicious electronic requests.

Time constrained electronic request evaluation

Systems, methods, and computer program products are directed to machine learning techniques that use a separate embedding layer. This can allow for continuous monitoring of a processing system based on events that are continuously generated. Various events may have corresponding feature data associated with at least one action relating to a processing system. Embedding vectors that correspond to the features are retrieved from an embedding layer that is hosted on a separate physical device or a separate computer system from a computer that hosts the machine learning system. The embedding vectors are processed though the machine learning model, which may then make a determination (e.g. whether or not a particular user action should be allowed). Generic embedding vectors additionally enable the use of a single remote embedding layer for multiple different machine learning models, such as event driven data models.

Optimally compressed feature representation deployment for automated refresh in event driven learning paradigms

This application discusses identifying data processing timeouts in live risk analysis systems. A service provider, such as an electronic transaction processor, may provide a production computing environment that includes a risk analysis system having one or more risk models, which may be machine-learning based. These risk models may be utilized in order to determine whether incoming data processing requests are fraudulent. To test these risk models using production data traffic, an audit computing environment made of a set of machines that do not service production computing environment requests, but that utilize databases and data connections as are used by the production systems. The audit computing environment may thus mirror the risk models and functionality of the production computing environment without the drawbacks of a typical fully separate testing environment. Thus, risk model performance and execution times may be monitored to determine whether any models encounter errors with production data traffic.

Identifying data processing timeouts in live risk analysis systems

Growing community of inventors

Bengaluru, India

Vishal Sood