Retail Assortment Planning: From Model Driven to Data Driven

零售商品品類規劃: 從模型驅動到數據驅動

Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date4 Sep 2018


This dissertation studies retail assortment planning problems from two perspectives: (1) assortment planning and updating decisions when customers have interest-decay effects on the provided products; and (2) customer preferences for online learning and dynamic assortment selection for general retail scenarios.

Based on a stylized model, the dissertation first considers assortment selection and rotation decisions under a locational choice model when customers are both variety and fashion seeking. The dissertation studies a two period assortment decision problem and captures variety-seeking behaviors by assuming that consumers have a lower valuation of products that are purchased in two successive periods. Due to this kind of variety-seeking behavior, the retailer needs to update the existing assortment after the initial planning period. When the retailer is myopic, i.e., when it is optimizing revenues in the current period, a single-period optimization model is established from which the respective optimal assortments of the two periods are derived. The results show that the retailer should retain customers by updating the most popular products and introducing more variants to precisely match customer preferences, which differ from the results in the existing literature. On the other hand, when the retailer strategically optimizes total revenue over two periods, the optimization model becomes more complicated. This work proposes two planning strategies to solve the problem.

This dissertation also studies a so-called attractiveness-decay effect, where customers' preferences for products within an assortment decay as they age on the shelf. In this case, uniform and unimodal preference distributions are considered in turn to investigate the optimal assortment decisions of the retailer. The results indicate that the optimal assortment under the uniform distribution would never contain products with overlaps in the attribute space when the cardinality constraint is loose, whereas under the unimodal distribution, the retailer has incentives to introduce more product variants, and the optimal assortment thus always consists of products with overlaps. It is also interesting to find that assortment updating is profitable only at an intermediate cost of launching a new product.

To further explore dynamic assortment decisions in a realistic context, this dissertation studies a learning-while-doing problem, where no initial preference information about customers is available, and the retailer may try different assortments to determine the optimal one. Due to the combinatorial nature of this problem, however, it is impossible to test all possible assortments to decide the optimum, let alone take into account the stochastic feedback involved. To address these challenges, the dissertation proposes a framework to efficiently guide the discovery of the optimal assortment. This framework consists of two modules: evaluation and improvement. The retailer attempts to oer the best assortment based on the existing knowledge and simultaneously collects new information about alternatives in the evaluation stage. This implies a well known exploration-exploitation trade-off. In the improvement stage, a ranking-based non-parametric model is adapted to estimate customer preferences and generate a new assortment. This non-parametric model provides a generic way to describe customer preferences without worrying about the problem of model misspecification, but it also poses great challenges in terms of optimization and computation. Fortunately, column generation and bender decomposition techniques are shown to efficiently solve the problem. The theoretical contribution of this part is twofold: (1) the regret bound in the evaluation stage is O(logT), where T is the time horizon; and (2)under certain conditions, the optimal assortment corresponding to the underlying customer preferences can be discovered with a higher probability. The experiments also demonstrate the effective and efficiency of the framework and the models.

    Research areas

  • Assortment Planning, Choice Model, Attractiveness decay, Online learning