Explainers

E-commerce and multimodal AI,

Recommending a product nobody has clicked yet

  1. New product 0 clicks, 0 purchases photodescriptionattributes
  2. Step 1 Find its closest look-alike among products that already have a sales history
  3. Step 2 Borrow the look-alike's history recommend what is usually bought with it
  4. Result +7% hit rate for new items, no retraining
A brand-new product borrows the history of the existing product it most resembles in photo, text and attributes. Tested on fashion data from H&M. From our ICIS 2024 paper.

Most recommendation systems run on behaviour: people who bought this also bought that. A product that went on sale this morning has no behaviour yet, so it is never recommended, so it gets no clicks. This is the cold-start problem, and in fashion, where collections turn over every season, it never goes away.

With Anat Goldstein, we treat a new product the way a shop assistant would: by what it looks like and what it is. Its photo, description and attributes become one combined representation, and we find the existing product it most resembles. That look-alike already has a history, so we recommend what is usually bought with it. Combining all three sources beat any one of them alone.

On data from the H&M store, adding this step to a standard collaborative-filtering recommender improved hit rate by 7% for new items, and it needs no retraining when products arrive. The same multimodal similarity also prices products: our ILAIS 2026 paper with Eylon Yehiel, which won second place for best paper, uses image, text and price history to suggest a price for an item.