Recommendation Systems Language Exercises

Collaborative filtering is a recommendation technique that generates suggestions by finding patterns in the behaviour of many users—what items similar users liked or interacted with. Key terms include user-user filtering, item-item filtering, the user-item interaction matrix, implicit feedback (clicks, dwell time), explicit feedback (star ratings), and matrix factorisation. Engineers often discuss sparsity, scalability, and the nearest-neighbour lookup when describing collaborative filtering architectures.

Content-based filtering recommends items based on features of the items themselves and the historical preferences of the individual user, without relying on data from other users. Key vocabulary includes feature vectors, TF-IDF representation, item embeddings, user profiles, and cosine similarity. The main advantage is that it avoids the cold-start problem for new items, but it can lead to filter bubbles where users only see content similar to what they have already consumed.

The cold start problem occurs when a recommendation system lacks sufficient data to generate accurate suggestions—either for a new user who has no interaction history (user cold start) or a new item that has received no ratings or clicks (item cold start). Common mitigation strategies include onboarding surveys, popularity-based fallback recommendations, content-based bootstrapping, and knowledge graph embeddings. This problem is one of the most discussed challenges in recommender system design.

Recommendation engineers use A/B testing terms such as treatment and control groups, traffic split, statistical significance, p-value, confidence interval, and minimum detectable effect (MDE). They also discuss online metrics like click-through rate (CTR), conversion rate, and dwell time versus offline metrics like precision@k and recall@k. Understanding the difference between short-term engagement signals and long-term user satisfaction is critical when interpreting A/B test results in recommendation contexts.

Common offline metrics include Precision@K, Recall@K, Mean Average Precision (MAP), Normalised Discounted Cumulative Gain (NDCG), and Hit Rate. For diversity and novelty, teams measure intra-list diversity and catalogue coverage. Online evaluation relies on A/B testing with business metrics such as conversion rate, session length, and revenue per user. The choice of metric depends on the business goal—a streaming service optimises differently from an e-commerce product recommender.

A personalisation platform is a system that delivers individualised content, product, or experience rankings to users at scale. Core components include a feature store (for serving real-time user and item features), a model serving layer, an experimentation framework, and a feedback loop that retrains models on fresh interaction data. Engineers discuss concepts like retrieval (candidate generation) and ranking (scoring candidates) as the two main pipeline stages.

Recommendation fairness concerns whether a system treats different user groups or item providers equitably. Key terms include provider-side fairness (ensuring smaller content creators receive adequate exposure), consumer-side fairness (ensuring recommendations are equally relevant across demographic groups), and popularity bias (the tendency of systems to over-recommend already-popular items). Mitigation vocabulary includes re-ranking, exposure constraints, and fairness-aware loss functions.

A feature store is a centralised repository that manages the creation, storage, versioning, and serving of machine learning features. In recommendation systems, it ensures that the same features used during model training are available at inference time with low latency—a property called training-serving skew prevention. Engineers discuss online stores (for real-time feature retrieval) and offline stores (for batch training) as two complementary components.

The retrieval stage (also called candidate generation) narrows the item catalogue from millions to hundreds of candidates using approximate nearest-neighbour (ANN) search, embedding lookup, or rule-based filters. The ranking stage scores each candidate using a more expensive model—often a gradient-boosted tree or neural network—and returns the top-N items. Practitioners describe this two-tower architecture as balancing recall (retrieval) with precision (ranking).

Explicit feedback is a direct signal from the user indicating preference, such as a star rating, a thumbs up, or a review. Implicit feedback is inferred from observed behaviour—clicks, purchases, watch time, or add-to-basket events—without an explicit rating. Most large-scale recommendation systems rely primarily on implicit feedback because explicit ratings are sparse and subject to response bias. Engineers often discuss how to weight different implicit signals and handle negative implicit feedback (e.g., skipped content).