University of Southern California

Real-Time Optimization of Personalized Assortments

By Negin Golrezai, Hamid Nazerzadeh,  and Paat Rusmevichientong, Marshall School of Business

Online commerce has been growing at an astronomical rate over the past decade, led by Amazon.com, partly at the expense of brick and mortar retailing. An important advantage of online retailing is the ability to change the assortment of products displayed (offered) frequently and instantaneously. In fact, the assortment offered can be unique and tailored to a customer’s browsing history, demographic characteristics, etc. Furthermore, on the supply side, the assortment offered can be altered in real-time depending on the availability of different SKUs. Going one step further, an online retailer may want to fine-tune the assortment of SKUs not only based on current inventory but also based on the likelihood of future demand for those SKUs. A procedure that offers a real-time, personalized assortment of SKUs to online customers with unique preferences that maximizes the retailer’s revenues may seem like a utopia. In fact, Marshall faculty and a doctoral student (Negin Golrezai, Hamid Nazerzadeh, and Paat Rusmevichientong) have gone one step further and developed a procedure that determines such an assortment which not only maximizes revenues but also takes into account inventory availability. In recent research, a relatively simple but provably powerful and robust algorithm has been developed to solve this important problem faced by online retailers.

To better understand the key idea behind the procedure developed by the Marshall researchers, imagine that you had two products 1 and 2 and you have the same inventory of both products. Customers are of two types A and B, with type A customers indifferent between products 1 and 2 while type B customers will only buy product 1. Furthermore, suppose type A customers are more likely to arrive first for shopping while type B customers are more likely to come later. In this case, if product 1 is more profitable than 2, then it is better to display only product 2 to type B customers as otherwise some of them will purchase product 1 whose inventory may run out and we may have no inventory left to sell to type A customers that come later. More generally, the idea behind the algorithm developed by the researchers is – it might be better to sell a product with lower marginal revenue but a high inventory level, than to sell a product with high marginal revenue but a low remaining inventory. This is because the future customers might only be interested in the products with low (or no) inventory, and if we have already sold those products, we would lose on these profits. To implement this idea, the authors propose a simple index for each product, which balances the nominal revenue with the value of each unit of remaining inventory. These indices are easy to compute, and they serve as a simple mechanism that balances the benefit of displaying or offering a unit for sale now with the benefit of not displaying it and keeping it in inventory for future sale.

The authors not only propose such a balancing algorithm but also show rigorously that it performs extremely well even if a retailer has no clue about when customers with different preferences will arrive for a purchase transaction. So, there is no need to forecast demand of different customer types, which was required in prior work and is a serious because such forecasts often turn out to be inaccurate. Furthermore, this work also does not assume any restrictions on the nature of consumer’s product choice decision given an assortment, which was again a limitation of prior work in this area. Finally, the researchers show that their model can be tweaked to learn about customer choices over time as customers make purchases and thus make even better assortment decisions. In addition to proving that their algorithm works extremely well under a wide range of demand and supply scenarios, the authors show that their approach works extremely well on real-world data. They test the algorithm on DVD sales data from an online retailer and show that personalizing assortments can increase revenues significantly, and this work shows that real-time optimization of personalized assortments can be done efficiently and robustly using their algorithm.

Online retailers can use the ideas and the algorithm developed in this work to improve their assortment decisions and profits. But this work is not restricted to online retailers of physical products. Airlines that are selling a limited number of seats in their flights, cruise lines selling limited number of cabins can use this research to improve their profits. Finally, sites that offer a limited number of coupons and deals to certain unique customers can also use the approach proposed in this work to improve the effectiveness of their promotion strategies.

For additional details about this research, the interested reader can access the research manuscript here. This is an example of the cutting edge research in supply chain management by Marshall faculty that can help managers improve their operations.