Multi-product Procurement Decisions: Feature-based Substitution, Data-driven and Preference Learning

Project: Research

View graph of relations


One of the challenging issues in developing analytical tools to assist multi-product (item) procurement decisions is how attributes(features) of products can be accommodated to address substitution effect. It is challenging because historical sales data seldomly captures whether a sale of a product was caused by its primary demand or a substitution for another product that was stocked out. This would be further exacerbated when new products are introduced into the existing assortment of a category, as there is no historical data for their demand. However, if a firm has been selling similar products of the category in the past and keeps a good record of them, which may contain rich information about the features of those products, such as colour, fabric, design style, sellingseason, price, etc., in the case of apparel. Such feature information may play an important role in shoppers' choice of one product over another. Intuitively, the prediction of a demand for a product is done by linking the “choice" of an arriving shopper to the “choice" made by historical shoppers who faced “similar" available product sets. In particular, we propose to study the problem when a retailer determines the initial order quantities of all products based on the transaction data, which records the feature information of each product, which product past shoppers bought from an assortment, and what products were available at the time of their “shopping". We will not assume a model on the distribution of shoppers' valuations, but only impose, instead, that choices are made based on their ranking orders of the available products, which in turn are determined by product features. That is, one product is purchased over others because the shopper ranks it the highest. (No purchase is alsoregarded as an option.) Hence, the individual ranking order is partially observable from the data. Then, the individual ranking orders - preferences - are learned via ranking kernels that measure similarities among products. Next, the approximate probabilities of purchasing individual products are estimated, and the average profit over the selling season is derived, upon which the optimalorder quantities will be determined. Consistency of estimation and asymptotic optimality will be established, and approximation strategies will be developed. Synthetic and real data will be used to conduct comprehensive numerical experiments to test the effectiveness of the framework and solutions and to draw managerial insights 


Project number9043422
Grant typeGRF
StatusNot started
Effective start/end date1/01/23 → …