At Wayfair, we recommend millions of products across all styles and budgets. Because of the scale of our catalog, we constantly ask ourselves: how can

Two Birds, One Stone: Hedwig, A Random-Walk Based Algorithm for Substitutable and Complementary Furniture Recommendations

submited by

Style Pass

2021-06-18 19:30:08

At Wayfair, we recommend millions of products across all styles and budgets. Because of the scale of our catalog, we constantly ask ourselves: how can we help our customers find the right products efficiently while discovering exciting complements to fill out their space?

In this article, we want to shed some light on a robust model called Hedwig used as a standalone recommendation system in multiple places where customers interact with Wayfair. Our Hedwig recommendation model helps customers throughout their entire journey, from email recommendations to suggestions on the home page, retargeting ads, and even push notifications.

The Hedwig model supersedes a legacy model trained on customer co-clicks to find related products using the Jaccard Similarity Index . Hedwig offers three crucial advantages compared to our legacy approach.

Hedwig is a recommendation model for fetching related furniture and home goods, given an example item. The Pixie Random Walk algorithm inspired the creation of our model. We create a bipartite graph by incorporating data on how customers interact with products to power the model. The two entities in the bipartite graph are SKUs and customer curations. A customer curation is essentially a set of SKUs curated by our customers, including session-level clicks, favorite lists, registries, carts, and orders. We simulate random walks on the graph to build the desired recommendations.