Bundled Product Recommendations

25 May 2024

Problem

Target.com shows bundles titled "Frequently Bought Together" or "Compatible Items" so a guest can add several matching products to the cart in one click. The first algorithm, association rule mining (ARM) on product attributes, could not tell the driver product (the one the guest views) from the go-together products. It sometimes put overly similar products in one bundle, and it gave no recommendations for a large part of the catalog. A graph built from joint purchases beat ARM but left large gaps: in a large catalog, most product pairs are rarely bought or added to the cart together.

Example from the article: a click on a car seat shows a bundle with a baby car mirror, a car seat organizer and a seat protector.

Solution in brief

Build a weighted, directed graph per product category from joint purchases and adds-to-cart. Add edges from a Similar Items graph, then train a graph neural network (GNN) with the SEAL framework to predict missing links. Show the top K predicted products per driver product. In an A/B test this raised click-based revenue by over 35% and coverage by over 50%.

Questions this case answers
  • How to recommend complementary items for most of a large catalog when pairwise co-purchase signals are too sparse to cover it
  • How to build asymmetric item-to-item relations (main item and add-ons) from sequences of cart events
  • How to recommend a bundle of complementary products (\"Frequently Bought Together\") for most of a large catalog, not only for products with many joint purchases?
  • How to tell the main product of a bundle from its add-ons when building bundles from co-purchase data?

Tools and methods

  • Association rule mining (ARM) — the baseline bundling algorithm
  • Co-purchase graph — directed, weighted, one per category
  • Similar Items graph — pads the co-purchase graph
  • SEAL framework — GNN for link prediction
Continue reading — it's free

Google or a one-time email code. No password.