How recommendation engines choose what to show users, from collaborative filtering to modern ranking models.
57 terms
Alternating least squares (ALS) factorizes a user-item matrix by solving for user and item factors in turn, a core method in recommender systems.
Bayesian personalized ranking (BPR) is a pairwise learning method that trains recommenders to rank items a user interacted with above unseen ones.
BERT4Rec is a sequential recommendation model that uses a bidirectional Transformer and masked item prediction to suggest the next item a user wants.
Beyond-accuracy metrics evaluate recommender systems on diversity, novelty, coverage and serendipity rather than on prediction accuracy alone.
Candidate generation is the first stage of a recommender system, quickly narrowing millions of items to a few hundred relevant ones for detailed ranking.
Catalog coverage measures what share of available items a recommender system actually recommends, revealing how much of the catalog users can discover.
The cold start problem is the difficulty recommender systems face in suggesting items to new users or new items that have little or no interaction data.
Collaborative filtering is a recommender technique that predicts what a user will like from the behavior of users with similar tastes, not from item content.
Content-based filtering recommends items whose features resemble those a user already liked, using item attributes rather than other users' behavior.
Context-aware recommendation adapts suggestions to the situation of use, such as time, location, device or company, rather than to user and item alone.
A conversational recommender system (CRS) elicits user preferences through multi-turn dialogue and refines its suggestions as the conversation develops.
Cross-domain recommendation transfers knowledge about users or items from one domain, such as books, to improve suggestions in another, such as films.
CTR prediction estimates the probability that a user clicks a shown item, and recommender systems use it to rank content, products and ads.
The data sparsity problem in recommender systems arises when users rate or interact with so few items that most user-item pairs have no observed signal.
Deep & Cross Network (DCN) is a ranking model for recommender systems that learns explicit feature crosses alongside a deep neural network to predict clicks.
A deep learning recommendation model (DLRM) combines embedding tables and MLPs to predict clicks from sparse and dense features.
DeepFM is a click-through rate prediction model that joins a factorization machine and a deep neural network over shared feature embeddings.
Explainable recommendation is the design of recommender systems that tell users why an item was suggested, using reasons that reflect how the model decided.
Explicit feedback in recommender systems is any preference a user states on purpose, such as a star rating, a like or a written review of an…
A factorization machine is a supervised model that captures pairwise feature interactions through latent vectors, widely used in recommender systems.
A field-aware factorization machine (FFM) is a click prediction model that learns a separate latent vector for each feature and each field it interacts with.
A filter bubble is the narrowing of what a user sees when recommender systems keep ranking content that matches past behavior and inferred preferences.
Generative recommendation is an approach in which a sequence model directly generates identifiers of items to recommend instead of scoring every candidate.
Group recommendation suggests items to several people at once, combining their preferences so that a shared choice satisfies the whole group fairly.
A hybrid recommender system combines collaborative filtering, content-based filtering and other signals to rank items more accurately and robustly.
Implicit feedback is user behavior such as clicks, views, purchases and watch time that recommender systems treat as indirect evidence of preference.
An item embedding is a dense learned vector that represents a product, video or article so recommender systems can compare items and match them to users.
Item-based collaborative filtering recommends items similar to those a user already liked, using item-to-item similarity learned from many users' behavior.
Learning to rank (LTR) trains machine learning models to order items by relevance, powering the final ranking stage of search and recommender systems.
LightGCN is a simplified graph convolution model for collaborative filtering that learns user and item embeddings by smoothing them over the interaction graph.
LLM-based recommendation uses large language models to rank, explain or generate item suggestions from text about users, items and their interaction history.
Matrix factorization in recommender systems learns low-dimensional user and item vectors whose dot products predict ratings or preferences.
MovieLens is a movie recommendation service and a family of public rating datasets that serve as standard benchmarks in recommender systems research.
Multi-objective recommendation ranks items by balancing several goals at once, such as clicks, watch time, satisfaction, diversity and fairness.
Multi-stage ranking is a recommender architecture that narrows millions of items to a short list through successively more precise and costly models.
Negative sampling trains recommender models on implicit feedback by contrasting each observed interaction with a few sampled items the user did not choose.
The Netflix Prize was a 2006-2009 open competition offering $1 million to improve movie rating prediction accuracy by 10%, shaping modern recommender systems.
Neural collaborative filtering (NCF) replaces the dot product of matrix factorization with a neural network that learns user-item interactions from feedback.
Next-basket recommendation predicts the full set of items a customer will buy in the next shopping trip from the sequence of past baskets.
Popularity bias is the tendency of recommender systems to over-recommend already popular items while niche items in the long tail stay rarely shown.
Position bias is the tendency of users to click items shown higher in a list regardless of relevance, which distorts feedback recommender systems learn from.
Rating prediction estimates the score a user would give an item they have not rated yet, a classic task in recommender systems evaluated with RMSE or…
Real-time recommendation updates suggestions within seconds of user actions, using streaming events and fresh features to rank items during a live session.
Recommendation diversity measures how different the items in a recommendation list are from each other, balancing variety against pure relevance.
Recommendation fairness is the requirement that recommender systems treat users, items and providers equitably in what they rank, show and promote.
A recommendation feedback loop arises when a recommender trains on clicks it caused itself, amplifying its own biases and narrowing what users see.
A recommender system is software that predicts which items a user is likely to value and ranks them, using past behavior, item content and context.
SASRec is a self-attentive sequential recommendation model that predicts the next item a user will interact with from the order of past interactions.
Sequential recommendation predicts the next item a user will interact with by modeling the order of past interactions rather than an unordered set.
Serendipity in recommender systems measures how often suggestions are both unexpected and relevant, helping users discover items they would not find alone.