09-20, 13:25–14:00 (Europe/Amsterdam), Rembrandt
When providing ML-based optimization products with multiple independent components, we can easily loose the holistic view of the problem that we are trying to solve and might end up not making optimal decisions. The easy way out - switching from a modular approach to a single unified optimization model - is often not feasible though. In this talk, I will discuss and demonstrate the implementation of an alternative method that will allow ML models to consider potential behavior of other components in a complex system using regret-sensitive loss functions. By using a simple game-theoretical formalization of the system and quantifying regret (i.e., the experience of a sub-optimal decision when information about the best action comes available after the model was already called), we can widen the optimization scope of the model and increase the overall performance of complex decision processes.
The talk will include open-source implementation examples for classical ML (e.g., XGBoost) and deep-learning (e.g., Tensorflow). Moreover, it will demonstrate how the discussed method was used to boost the performance of a model that is processing millions of payments every day. The talk is directed at everyone who is interested in model development - an understanding of basic ML concepts specifically loss functions would be beneficial.
Agenda of the talk:
• Example of a sequential ML decision process (5 min)
• Regret in decision theory (5 min)
• Implementation of regret-avoidant models - practical examples (10 min)
• Boosting conversion rates with regret models at Adyen (5 min)
• Customizing regret functions to balance different objectives (5 min)
Laura is a senior machine learning scientist at Adyen - currently working in the authentication domain. She has a background in cognitive science and completed her PhD in the field of affective computing at LMU Munich.