As with many e-commerce companies, recommender systems are widely used at Wayfair to show the most relevant products to customers at every stage of their shopping journey. Machine learning models used in recommender systems typically optimize for metrics related to customer satisfaction: actions such as clicks, conversions and Add-to-Cart. My talk at the 2022 Annual Meeting of the Institute for Operations Research and the Management Sciences (INFORMS) was related to how Wayfair is utilizing inputs from our supply chain to influence the products we showcase to consumers as they search and browse our storefront.
Let’s take an example of a customer that resides in San Francisco. She is interested in buying a couch with light gray polyester fabric. However, this couch is currently stored at a Wayfair warehouse in Boston. Wayfair has two couches in a warehouse close to the customer’s residence in San Francisco. Each of them differ from the customer’s ideal preference in minimal ways – one has a light-gray cushion fabric, while the other has fabric made out of linen.
The models that we are building at Wayfair would utilize a key supply chain input – the fact that these items are closer to the customer – and surface these two items in the search results. The customer benefits because they are able to shop for items that can be shipped faster to their residence. In addition, Wayfair could ship items from a nearby warehouse at a lower cost: these are savings that can be passed on to the customer. Additionally, shipping items over shorter distances helps us reduce our carbon footprint, and helps us achieve our goal of reducing Scope 1 and Scope 2 carbon emissions by 63% from a 2020 baseline.
The model we are building balances inputs from the supply chain along with criteria related to customer satisfaction to influence the output of the recommender system. These models measure the opportunity cost that takes real-world and ideal scenarios into consideration. The model is not exclusively designed to match the customer with the “best case” item for every instance – rather it also takes into account factors related to supplier success. Overall the model considers a wide range of criteria that include optimal supplier investment, lower prices for customers and the supply chain cost efficiency over a longer period of time.
This is the first project at Wayfair where we are building machine learning models to influence customer demand by improving supplier incentives – as lower shipping costs translates directly into increased sales for suppliers. Over the longer term, as the multi-objective optimization methods mature, we expect to discover even more methods.
One approach that shows promise is utilizing reinforcement learning techniques that turn learnings from customer interactions into measurable performance indicators related to cost savings or profitability. For this approach, we are considering a greater variety of supply chain inputs. To give just one example, in addition to shipping costs, we also take incident rates into account, which among other factors, are related to the percentage of items that arrive damaged at a customer’s residence. Our recommendation engines effectively surface items from suppliers that have a lower incident rate. This will result in a reduction in supply chain cost, and will allow customers to get products on Wayfair at even more affordable prices, while giving suppliers the opportunity to make optimal investments.
My career has largely focused on the field of operations research. As a result, my first instinct is to primarily consider perspectives from the world of supply chain management. However my dissertation was focused on consumer behavior at the intersection of marketing management and operations research in a quantitative modeling world, which allows me to also factor in an applied economist’s perspective. At Wayfair, I’m excited to work with a team that blends multiple perspectives from both the supply and demand sides to help people create uniquely personal spaces at home in an affordable way.