Online Assortment Planning with Ranking-based Choice Models

Description

Assortment planning selects the set of products to be sold and is among the most important decisions to retailers. The traditional approach to assortment planning first forecasts demand for candidate products and then optimizes the assortment. This approach suffers from a few drawbacks. (1) Historical data may be too limited to allow for reliable forecasting; in case of new products, no historical sales data are available. (2) In practice, demand forecasting for different products is handled separately, and mutual influences among products are not captured. Therefore, the resulting forecast is often unreliable. (3) The forecast-then-optimize approach, when directly applied for dynamic assortment planning, risks being suboptimal due to ignoring the exploration-exploitationtrade-off. Motivated by our consulting experience with Alibaba, in this project, we consider an e-tailer that needs to plan its assortment over a planning horizon. Assuming no or little prior knowledge about customer preferences, we propose a learning-while-doing approach for the dynamic assortment planning problem. The envisioned approach repeats two steps iteratively. The first step maintains an effective assortment set and treats each assortment as an arm. A tailor-made multi-armed bandit algorithm will be used to evaluate the ssortments and generate revenue; thus this step involves the classical exploration-exploitation trade-off. The first step spans overmultiple periods until sufficient information has been collected regarding customer preferences. The second step is then carried out to find a new, and hopefully better, assortment. In this step, we propose to simultaneously learn customer preferences and optimize the assortment using a MaxMin optimization model. This promises to yield a high-quality assortment that is robust to sales data uncertainty. The effective assortment set is augmented by including the new assortment. The process then proceeds by repeating the two steps.We propose to model customers’ product preferences using a ranking-based choice model. This nonparametric choice model allows us to capture customers’ intricate substitution behavior while incurring a low risk of model misspecification. However, its combinatorial nature also poses great challenges in terms of optimization and computation. Therefore, we will incorporate some side information to reduce the number of rankings we need to consider, which will ease the computational challenge. We will develop efficient algorithms to solve the MaxMin optimization model, characterize the theoretical performance of the proposed framework, and design a practicallyactionable dynamic assortment planning process.For the project, we will partner with Alibaba, a leading e-commerce firm, which has provided a detailed dataset and will also test our proposed approach on their platform. This will ensure that the assortment planning approach developed in this project is both theoretically appealing and practically effective. The project will yield research findings that have wide applicability tothe online retailing industry.