Applied AI/ML Research at PayPal with Vidyut Naware - #593

Causal Inference

Causal inference is a pivotal component in PayPal's machine learning models, aiming to enhance the robustness and stability of predictions. Vidyut Naware highlights the shift from predictive to treatment-effect problem formulation, particularly in marketing, to better influence customer behavior 1. This approach seeks to identify causal features, moving beyond correlation to understand why certain outcomes occur, thus improving model performance amidst data distribution shifts 1.

We believe that instead of looking at these problems as predictive problems, they are much better formulated as treatment and effect kind of problems.

--- Vidyut Naware

The exploration of causal ML techniques is ongoing, with collaborations such as the pilot with Stanford's Professor Susan Athey, which focused on identifying causal features 2.

   

Graph Techniques

Graph machine learning techniques are integral to PayPal's AI strategy, particularly in fraud detection and compliance. Vidyut Naware explains the use of Graph Convolutional Networks (GCNs) to enhance collusion detection by learning embeddings that reveal fraudulent patterns 3. The challenge lies in managing the vast scale of PayPal's graph data, which includes millions of customers and billions of transactions 4.

Graph is such a rich representation of our user interaction data.

--- Vidyut Naware

The team focuses on adapting academic advancements in graph ML to practical applications, ensuring these innovations align with PayPal's unique data attributes 4.

   

Federated Learning

Federated learning at PayPal is a journey towards enhancing AI capabilities while maintaining privacy and efficiency. Vidyut Naware outlines the phased approach, starting with on-device inferencing and gradually incorporating differential privacy measures 5. This method balances algorithmic and logistical challenges, aiming to eventually enable on-device training 5.

Federated learning, where the full ambit of federated learning actually is also eventually doing training on device.

--- Vidyut Naware

The focus is on building the necessary infrastructure to support this evolution, reflecting a commitment to privacy-preserving machine learning 5.