Recommendation Systems

A recommendation approach that predicts a user’s preferences based on the preferences of similar users (user-user CF) or based on similarities between items that the user has engaged with (item-item CF). Does not require content features — only interaction data.

"Our ‘users like you also bought’ feature uses user-based collaborative filtering: find the 50 most similar users (by cosine similarity of purchase vectors), aggregate their purchases weighted by similarity, filter items the target user already bought. Works well for popular items; struggles with new users and new items (cold start problem)."

A technique for latent factor models: decompose the user-item interaction matrix (ratings, clicks) into two lower-rank matrices (user embeddings and item embeddings). The dot product of a user’s and item’s embedding predicts the rating. ALS and SGD are common training algorithms.

"We train a matrix factorisation model on implicit feedback (clicks, dwell time). Each user and item gets a 64-dimensional embedding. To recommend for a user, compute the dot product of their embedding with every item’s embedding, rank, filter already-seen items, return top 10. The embeddings capture latent preferences: ‘users who like distributed systems books also like SRE content.’"

The difficulty of making recommendations for new users (no interaction history) or new items (no interaction data). Without history, collaborative filtering cannot find similar users or items. Requires fallback strategies: popularity-based, content-based, or exploration.

"New users see a ‘quick taste quiz’ at signup to bootstrap their profile — this addresses the new user cold start. New articles use a content-based fallback: similar to articles the user has read (using text embeddings) until the article accumulates 100 interactions. After 100 interactions, the CF model takes over."

A ranking quality metric. DCG rewards placing relevant items high in the list (discounted by log rank). NDCG normalises DCG by the ideal ordering (IDCG). Scores range from 0 to 1; higher is better. Captures both relevance and position.

"We measure recommendation quality at NDCG@10: the top 10 recommendations are scored. A relevant item at position 1 contributes more than the same item at position 5. Our baseline NDCG@10 was 0.43; after the new ranking model it’s 0.51 — an 18% relative improvement. NDCG is our primary offline evaluation metric before A/B testing."

A two-stage recommendation architecture. Stage 1 (candidate generation): fast retrieval of hundreds of candidates from millions of items using approximate methods (ANN, collaborative filtering, content similarity). Stage 2 (ranking): score candidates with a more complex model considering user context, item features, business rules.

"Our recommendation pipeline: candidate generation (ANN retrieval with user embedding + popularity-based candidates) produces 500 candidates in 10ms. The ranking model (gradient-boosted tree on 50 features) scores each candidate in 50ms. Post-processing applies business rules (no out-of-stock, diversity constraints). Total: under 100ms."

Filter bubble: the tendency of recommendation systems to show users increasingly narrow content aligned with their existing preferences, reducing exposure to diverse viewpoints. Serendipity: a metric measuring how surprising and delightful recommendations are — deliberately injecting diversity to combat filter bubbles.

"Our listening analysis showed users in a filter bubble: they were recommended only one genre after 3 months. We added a serendipity injection: 10% of recommendations must be from genres the user hasn’t seen in the last 30 days. Short-term click-through dropped 2%, but 30-day retention improved 8% — users were less likely to feel the platform had nothing new."