Abstract: This talk will focus on how we can use machine learning models, and choose interventions and metrics that balance the needs of the pros and customers on both sides of the marketplace, in addition to considerations to ensure the ongoing health of the fraudulent customer detection system. We will cover techniques to deal with imbalanced and biased labelled data at the time of training the model, as well as approaches to ensure that the ML model does not become biased over time. We will also discuss how we can develop the optimal experience for both customers and pros by using appropriate model metrics, and quantifying the effectiveness and costs of different types of intervention.
After this talk, attendees will have an understanding of how to think of fraud detection as a system, not just an individual ML model. Attendees will gain insight into how to use a simple binary classification model to trigger multiple types of intervention depending on likelihood of a customer being fraudulent, as well as techniques to ensure the ongoing performance of the system.