Doubly high-dimensional contextual bandits: An interpretable model for joint assortment and pricing
The rapid growth in data availability, the vast need for decision-making, and advancements in machine learning and statistics have made data-driven decision-making possible and unprecedentedly important. In high-stake fields, such as business and healthcare, decision-makers face more challenges: managing high dimensionality of data, balancing interpretability with performance, ensuring computational efficiency and statistical accuracy, and adhering to domain-specific principles. These multifaceted challenges call for innovative approaches in modeling, methodology, and theory.
In this talk, I will focus on my work on doubly high-dimensional contextual bandits. This work is motivated by a real-world challenge: the joint assortment and pricing problem faced by an industry-leading instant noodles company, where we need to make decisions about product offerings and their pricing simultaneously. To address this problem, we propose a novel model — doubly high-dimensional contextual bandits — to capture this sequential decision-making problem. We propose an efficient algorithm for this interpretable yet flexible model. We showcase their power through theoretical guarantees, case studies, and simulation studies. If time permits, I will also talk about my related work on personalized reinforcement learning.