Problem
ASOS sells nearly 50,000 products at a time and adds hundreds of new ones every week. Its recommender ranks products at many touch points and serves 5 billion requests a day. The current model is an asymmetric matrix factorisation model. It treats every past interaction as equally important, whatever its position in the history or its relative importance. It misses the customer's style and how that style changes over time.
Example from the article: two customers both own Salomon trainers. One also bought a waterproof jacket and insulated trousers (hiking), the other a Carhartt jacket and Dickies trousers (streetwear). Their alternative shoe recommendations should differ.
Solution in brief
ASOS built a transformer recommender with NVIDIA's Transformers4Rec library and compared it with the matrix factorisation model. Self-attention lets each product in the history take on a meaning that depends on the other products around it. Positional information lets the model weigh recent interactions differently from old ones. In offline evaluation the transformer beat the baseline by over 20% on hitrate.
Questions this case answers
- How to improve a production recommender that weighs every past interaction equally, so that it accounts for interaction order and the context of other items in the history
- How to make product recommendations depend on the order and context of a customer's past interactions?
- How to capture a customer's style when the same product means different things in different baskets of interactions?
- How to replace a matrix factorisation recommender that treats all past interactions as equal?
Tools and methods
- Transformers4Rec — NVIDIA library for sequential and session-based recommendation with transformers (works with PyTorch).
- Self-attention — gives each product a representation that depends on the rest of the history.
- Positional awareness — uses the order of interactions to separate long-term style drift from short-term intent switches.
- Asymmetric matrix factorisation — the production baseline.
Google or a one-time email code. No password.