Telfer Research Seminar Series - Sheng Liu

In this talk, I will discuss the value of proxy information in decision-making problems where data may not be sufficient and/or reliable. The first use case is motivated by sales prediction in retail, where detailed SKU demand data can be too volatile and limited to enable robust prediction performance. We propose to use proxy information from aggregate sales as a regularization for fitting the SKU-level prediction model. The resulting approach also combines ideas from transfer learning and gradient boosting to yield superior performance. Then, I will briefly discuss a use case for locating physician-led prehospital critical care teams (PHCCTs) in response to major traumas, where the proxy information can help improve decision quality with limited historical data. (This talk is based on joint work with Timothy Chan, Dazhou Lei, Max Shen, JD.com, and Rachel Stephenson.)