How would you explain the concept of 'embedding' in the context of recommendation systems?

Example answer

An embedding represents a user, item or context as a vector. Its meaning depends on the objective and scoring rule used to learn it. A matrix-factorization model can learn user and item vectors whose dot product approximates observed feedback. A feature-based encoder can instead construct vectors from item content or context. Low dimensionality alone does not establish useful recommendations.

For an illustrative dot-product model, let the user vector be (1, 2), item A be (2, 0), and item B be (0, 2). A scores 2 and B scores 4, so this model ranks B first. These scores are not probabilities. Changing to cosine similarity or Euclidean distance can change rankings; use the metric the training and retrieval design actually supports.

Implementation starts with the interaction definition, permitted data and target outcome. Distinguish exposure from clicking or an explicit rating: an unobserved interaction is not proof of dislike. Choose sampling and loss deliberately. Keep encoder versions, item vectors and the retrieval index compatible; provide a fallback for users or items without usable representations.

Evaluate retrieval recall or ranking quality on a temporally appropriate held-out set, including new-user and new-item slices. Compare latency and memory with a suitable baseline. Then test product outcomes with a controlled experiment where feasible, monitoring errors and user control as well as engagement. A platform login increase after release does not show that embeddings caused better recommendations.

Sources: Google matrix factorization, Google candidate scoring.