Recommendation Systems

• Collaborative Filtering: Finds users with similar behaviour and recommends what those similar users liked. "Users like you also engaged with..." Best for products/content with rich user-item interaction history. Strong for cross-sell and surprise discovery. Data requirement: High needs sufficient user-item interactions.
• Content-Based: Recommends items similar to what the user has previously engaged with, based on item features (category, tags, description embedding). Works for new items without interaction history. Good for "similar to this item". Data requirement: Low item features only, no interaction data required.
• Hybrid: Combines collaborative + content-based signals collaborative for users with interaction history, content-based for cold-start. Best overall performance when both interaction and feature data exist. Data requirement: Moderate benefits from both.
• Matrix Factorisation: Decomposes user-item interaction matrix into latent user and item vectors. ALS (Alternating Least Squares) for implicit feedback (clicks, views without explicit ratings). Classical production baseline. Efficient at scale. Data requirement: High dense-enough matrix for factorisation.
• Two-Tower Neural: Two separate neural networks one for users, one for items each producing an embedding. Recommendations = approximate nearest neighbour search in embedding space. Large-scale retrieval (millions of items). YouTube, Pinterest, TikTok-style recommendation at scale.
• LLM-based (Semantic): Uses LLM embeddings of item descriptions/content for similarity. Items with semantically similar descriptions are recommended. No interaction data required. Best for cold start new platforms with no interaction history.

The cold start problem refers to the difficulty of making recommendations for new users or new items that have no interaction history. For new users (user cold start), the system cannot rely on their personal interaction history it must fall back to popularity-based recommendations, onboarding questions that capture explicit preferences, or content-based recommendations based on item features. For new items (item cold start), collaborative filtering cannot recommend the item until enough users have interacted with it content-based approaches using item metadata (description, category, tags) are used to recommend new items alongside established ones. Two-tower neural models mitigate item cold start by representing items through their features rather than learned interaction embeddings.

Recommendation quality is measured offline (using held-out interaction data) and online (through A/B testing on real users). Offline metrics: Precision@K (of the top-K recommendations, what fraction did the user actually engage with?), Recall@K (of all items the user engaged with, what fraction appeared in the top-K recommendations?), NDCG@K (Normalised Discounted Cumulative Gain weights hits higher when they appear earlier in the ranked list), and Coverage (what fraction of the item catalogue is recommended to at least one user low coverage means the model only recommends popular items). Online metrics: CTR (click-through rate on recommendations), conversion rate (purchases from recommendations), and revenue lift (measured against a control group in an A/B test). Offline metrics are fast and cheap; online metrics are the business-relevant ground truth.