Problem
A shopper on a fashion or home item page wants a whole outfit or room, not one product. Walmart's complementary-item recommendations rank single items by co-purchase and co-view, so they do not guarantee that the items match in style, brand, price or color. Stylists can write good looks but cannot cover the whole catalog.
Example from the article: for an orange dress as the anchor item, the model shows several sets of matching items (Fig. 1 of the post, not reproduced here).
Solution in brief
Build a look as items from 4 to 5 super product types (groups of similar product types). Fill each slot from the complementary-item recommender, re-rank items with CLIP-based style embeddings trained with triplet loss, and rank looks by price, brand and color scores. Expand a few human-made looks to the whole catalog by visual search, then drop incompatible combinations by attribute rules.
Questions this case answers
- How to recommend a coherent multi-item set around an anchor item when pairwise complementary recommendations do not guarantee that the items fit together
- How to extend a small set of expert-curated item groups to a large catalog while keeping the groups internally compatible
- How to build a complete, style-coherent set of products around the item a customer is viewing?
- How to make the items of a generated outfit match each other visually, not just co-purchase statistics?
Tools and methods
- Complementary-item (CI) recommender — candidate source for each anchor item.
- Super product types (SPT) — groups of product types with the same function that define the slots of a look.
- CLIP image embeddings — input for style embeddings and for visual search.
- Triplet loss with a feed-forward network — learns the style embedding.
- FAISS or ScaNN — approximate nearest-neighbor search over image embeddings.
- Rule filters — attribute compatibility checks.
Google or a one-time email code. No password.