Machine learning terms/Recommendation Systems

See also: Machine learning terms

What is a recommendation system?

A recommendation system (also called a recommender system) is a class of machine learning software that predicts the preference a user would give to an item, then surfaces a ranked list of items the user is most likely to enjoy. Recommendation engines power product suggestions on Amazon, video feeds on YouTube and TikTok, music playlists on Spotify, social posts on Facebook and Instagram, and pins on Pinterest. A 2013 McKinsey report estimated that "35 percent of what consumers purchase on Amazon and 75 percent of what they watch on Netflix come from product recommendations based on algorithms" ^[25], a pair of figures that is still the most widely quoted evidence of the technology's commercial impact.

A recommender takes three core inputs: a set of users, a set of items, and a record of interactions between them (ratings, clicks, watches, purchases, dwell time, likes). The output is a personalised ranking of items for each user, usually computed as a scoring function followed by re-ranking. At industrial scale the pipeline is split into candidate generation, scoring, and re-ranking because exhaustively ranking every item for every user is infeasible when catalogues contain hundreds of millions of items.

This page surveys the dominant families of recommendation algorithms, the milestone systems that defined each era, the evaluation metrics that practitioners use, and the open source libraries that make these techniques accessible.

Why do recommendation systems matter?

Catalogue size has exploded: Amazon lists hundreds of millions of products, Spotify hosts more than 100 million tracks, and YouTube receives more than 500 hours of video uploads per minute. Without algorithmic filtering, users cannot navigate this long tail. Attention is scarce, so platforms compete on relevance per second of user time. Recommenders also create a self reinforcing data flywheel: better recommendations drive engagement, which produces more interaction data, which trains better models.

The business stakes are large and well documented. Netflix's own researchers estimate that "the combined effect of personalization and recommendations save us more than $1B per year" ^[12], and YouTube has said that recommendations drive more than 70% of the time people spend watching on the platform ^[27]. Figures like these explain why every major consumer platform now treats recommendation quality as a core product metric rather than a background feature.

What are the classical recommendation algorithms?

Collaborative filtering

Collaborative filtering (CF) makes recommendations using only the user item interaction matrix, with no information about the content of items or the demographics of users. The core idea is that users who agreed in the past tend to agree in the future. The term itself was coined by David Goldberg and colleagues at Xerox PARC in the 1992 Tapestry system ^[1]. CF was popularised soon after by the GroupLens research group at the University of Minnesota in their 1994 paper on Usenet news filtering ^[2], and it remains the conceptual foundation of the field.

There are two flavours of memory based CF:

Scroll sideways for more →

Amazon's 2003 paper Item to Item Collaborative Filtering by Linden, Smith, and York pushed item item CF into industrial use ^[3]. Item similarities can be precomputed offline, so serving cost stays low even as the user base grows, which is why the approach scaled to Amazon's catalogue independently of the number of customers.

Content based filtering

Content based filtering recommends items whose features resemble items the user already liked. A film recommender might encode each movie as a vector of genres, actors, and keywords, then compute cosine similarity to the user's profile. Pandora's Music Genome Project, launched in 2000, is the textbook example: human analysts tag each song with up to 450 musical attributes, and the recommender finds songs whose attribute vectors lie close to those the listener has thumbed up. Content based methods avoid the pure cold start problem for new items but cannot exploit the wisdom of the crowd and tend to produce homogeneous suggestions.

Hybrid recommenders

Hybrid systems combine collaborative and content signals. Common blending strategies include weighted hybrids (linear combination of scores), switching hybrids (use content based when collaborative data is sparse), and feature combination (feed both signals into a single ranker). Robin Burke's 2002 survey Hybrid Recommender Systems enumerates seven classes.

How does matrix factorization work?

Matrix factorization (MF) was the dominant family of recommendation algorithms from the mid 2000s through the mid 2010s. It decomposes the sparse user item interaction matrix R of shape (m users by n items) into two low rank matrices: a user matrix U of shape (m, k) and an item matrix V of shape (n, k), where k is the embedding dimension, typically 16 to 200. The predicted rating of user u for item i is the dot product of their latent vectors.

Funk SVD and the Netflix Prize

In October 2006 Netflix launched the Netflix Prize, a one million dollar competition for any team that could improve the company's Cinematch rating predictor by 10% ^[26]. Simon Funk (the pen name of Brandyn Webb) published a blog post in December 2006 describing an SGD variant of SVD that handled missing entries gracefully ^[4]. His method, retroactively named Funk SVD, became the seed for almost every leading entry. The loss was squared prediction error plus L2 regularisation on the latent factors.

ALS and weighted ALS

Alternating Least Squares (ALS) fixes one factor matrix and solves the other in closed form, then alternates. Each subproblem is a regularised least squares regression and parallelises trivially across users or items, which made ALS the algorithm of choice on Hadoop and Spark clusters. The weighted alternating least squares (WALS) variant adds a confidence weight to each observation, useful when the dataset combines explicit ratings with weaker implicit signals.

SVD plus plus

SVD++, proposed by Yehuda Koren in 2008, augments standard SVD with implicit feedback by adding a term that sums embeddings of all items the user has interacted with, regardless of rating ^[5]. SVD++ was a building block of BellKor's Pragmatic Chaos, which won the Netflix Prize in September 2009 with a test RMSE of 0.8567, a 10.06% improvement over Cinematch ^[26].

Bayesian probabilistic matrix factorization

Bayesian Probabilistic Matrix Factorization (BPMF), introduced by Salakhutdinov and Mnih in 2008, treats user and item factors as random variables with Gaussian priors and uses MCMC (Gibbs sampling) for inference ^[7]. BPMF is robust to overfitting and gives calibrated uncertainty estimates at higher compute cost.

The Netflix Prize blend

The winning entry was a linear blend of a large ensemble of predictors, including SVD++, restricted Boltzmann machines, asymmetric SVD, neighbourhood blends, and time aware models; the full BellKor's Pragmatic Chaos solution drew on hundreds of individual predictor sets ^[26]. Ensembles of diverse complementary models almost always outperform any single algorithm.

What is implicit feedback in recommenders?

Most real world interaction data is implicit: clicks, plays, dwell time, purchases, scroll depth. Implicit signals are abundant but noisy and one sided, since users do not produce explicit dislikes for items they ignore. The seminal paper Collaborative Filtering for Implicit Feedback Datasets by Hu, Koren, and Volinsky (IEEE ICDM 2008) reframed the problem as predicting binary preference (interacted or not) weighted by a confidence c that grows with the count of interactions ^[6]. The objective is solved with a weighted ALS step that scales linearly in observations. The Hu Koren Volinsky formulation underpins the open source Implicit library and remains a strong baseline.

How do deep learning recommenders work?

From the mid 2010s, deep learning reshaped recommendation, replacing hand tuned similarity functions with learned nonlinear interactions over user and item features.

AutoRec

AutoRec (Sedhain et al., WWW 2015) was one of the first neural recommenders to gain traction ^[9]. It treats CF as an autoencoder problem: the input is a partially observed row or column of the rating matrix, and the network learns to reconstruct it. AutoRec outperformed plain MF on MovieLens benchmarks.

Neural collaborative filtering

Neural Collaborative Filtering (NCF) by He et al. (WWW 2017) replaced the inner product in MF with an MLP applied to the concatenation of user and item embeddings ^[14]. The paper introduced the NeuMF architecture, a fusion of generalised MF and an MLP tower. NCF was hugely influential, although a 2020 reproducibility study by Rendle et al. argued that well tuned MF baselines match NCF on standard datasets ^[22].

Wide and deep

Google's Wide and Deep Learning for Recommender Systems (Cheng et al., 2016) was deployed in the Google Play Store ^[10]. The wide arm is a linear model over crossed categorical features that excels at memorisation; the deep arm is an MLP over dense embeddings that excels at generalisation. The two arms are jointly trained. Wide and Deep set the template for hybrid architectures and inspired DCN, DeepFM, xDeepFM, and DIN.

Factorization machines and DeepFM

Factorization Machines (Rendle, 2010) generalise MF to arbitrary feature interactions ^[8]. DeepFM (Guo et al., 2017) plugs an FM layer in alongside a deep network so that low order and high order interactions are modelled jointly, without manual feature engineering ^[15].

What is sequential and session based recommendation?

Order matters. The next song you play depends on the last three you played, not on a static profile. Sequential recommenders model user behaviour as a sequence and predict the next item.

Scroll sideways for more →

Sequential models are now the workhorses of short video and music feeds.

What is a two tower recommender?

At very large scale (think YouTube, with billions of users and billions of videos) you cannot afford to score every (user, item) pair with a deep network. The two tower or dual encoder architecture trains a user tower and an item tower that each output a fixed dimensional embedding, then defines the score as a dot product or cosine similarity. Item embeddings can be indexed in an approximate nearest neighbour structure such as ScaNN or Faiss, and retrieval becomes sub linear in catalogue size.

The canonical reference is Deep Neural Networks for YouTube Recommendations by Covington, Adams, and Sargin (RecSys 2016) ^[11]. The paper frames web scale recommendation around three challenges, scale, freshness, and noise, and splits the system into a candidate generation network followed by a ranking network, the two stage blueprint that most large platforms still follow ^[11]. Sampling negatives correctly is the central engineering challenge: in batch negatives, mixed negative sampling, and sampled softmax with logQ correction are all standard tricks. Google's 2019 paper Sampling Bias Corrected Neural Modeling for Large Corpus Item Recommendations derived a streaming logQ correction that lets the system train on fresh logs.

How do graph neural networks power recommendations?

User item interaction data forms a bipartite graph, and richer graphs include item item co occurrence, social ties, and knowledge graph relations. Graph neural networks (GNNs) propagate embeddings along edges so that an item's representation reflects its multi hop neighbourhood.

PinSage

PinSage (Ying et al., KDD 2018) was the first graph convolutional recommender deployed at web scale ^[17]. Built at Pinterest on a graph of three billion nodes (pins and boards) and 18 billion edges, and trained on 7.5 billion examples, PinSage uses random walk based neighbour sampling and importance pooling, sidestepping the memory cost of full graph convolutions ^[17]. In offline evaluation the authors report "outperforming the top baseline by 40% absolute (150% relative) in terms of the hit rate and also 22% absolute (60% relative) in terms of MRR" ^[17].

LightGCN

LightGCN (He et al., SIGIR 2020) stripped GNNs down to their essentials by removing feature transformation and non linear activation ^[21]. The remaining operation is iterative neighbourhood aggregation followed by averaging across layers. LightGCN beat NGCF on all three standard benchmarks and became the strongest GNN baseline.

Knowledge graph aware recommenders

KGAT, RippleNet, and KGCN integrate external knowledge graphs (DBpedia, Wikidata) so that recommendations exploit factual relations like (movie, directed by, director). These methods help with cold start because new items inherit signal through their knowledge graph neighbours.

How does a multi stage recommender pipeline work?

Industrial recommenders are pipelines, not monolithic models. A typical stack has four stages:

Scroll sideways for more →

Each stage trades model expressiveness for latency. The candidate generator must be cheap but high recall, the ranker must be accurate but can afford more compute, and the re-ranker enforces business constraints.

What is the cold start problem?

The cold start problem arises when the system lacks interaction data for a new user, a new item, or an entire deployment. Standard mitigations:

• Content features. Use side information about the new item (text, image, metadata) so the model can place it in embedding space without interaction history.
• Demographic features. For new users, use age, location, device, or onboarding survey answers as a fallback profile.
• Transfer learning. Pretrain on a large auxiliary corpus and fine tune on the target domain. Foundation model embeddings (CLIP for images, sentence transformers for text) are popular item encoders.
• Exploration. Insert randomised or epsilon greedy items so the system gathers data on new entries. Multi armed bandit approaches such as LinUCB and Thompson sampling formalise the explore exploit tradeoff.
• Meta learning. MAML style few shot learners adapt quickly to new users.
• Prompt based generation. LLM based recommenders zero shot rank new items from textual descriptions.

How do industrial recommendation systems work?

Netflix

Beyond the original ratings recommender, Netflix personalises artwork (which thumbnail to show), row ordering on the home page, search results, and autoplay trailers. The 2016 paper The Netflix Recommender System by Gomez Uribe and Hunt is required reading ^[12]. The company runs continuous A/B tests, and its researchers estimate that "the combined effect of personalization and recommendations save us more than $1B per year" ^[12], largely by reducing subscriber churn; the same paper attributes roughly 80% of the hours streamed on Netflix to the recommender rather than to search ^[12].

YouTube

YouTube's chief product officer Neal Mohan said at CES in January 2018 that recommendations drive more than 70% of the time users spend watching on the platform ^[27]. The architecture (Covington 2016) is two stage: candidate generation via deep neural networks, then a ranking model combining hundreds of features including video age, click context, and predicted watch time ^[11]. Recent developments include reinforcement learning for long term satisfaction (REINFORCE recommender, 2019) and large scale transformer rankers.

Spotify

Spotify's Discover Weekly playlist, launched in July 2015, generates a personalised 30 song mix every Monday ^[28]. By the fourth quarter of 2025 Spotify had reached 751 million monthly active users, so the playlist is regenerated at enormous scale ^[29]. The system blends three signals: collaborative filtering over playlist co occurrence, NLP analysis of music blogs and reviews, and audio analysis using convolutional networks on raw spectrograms ^[28]. Spotify also operates Daily Mixes, Release Radar, and AI DJ (launched February 2023), an LLM voiced personalised radio host ^[30].

Pinterest

Pinterest pioneered PinSage, the first web scale graph convolutional recommender, and runs related pins, search, and home feed off of Pin and Board embeddings. The Homefeed Personalization talks describe a multi objective ranker that balances engagement, diversity, freshness, and advertiser content.

TikTok

TikTok's For You feed is credited with raising the bar for engagement. Public statements describe a ranking model that scores candidates by predicted likes, comments, completion rate, and watch time, with strong emphasis on negative feedback signals (skips, not interested taps), and surfaces niche creators quickly through aggressive exploration.

Amazon, Alibaba, Meta

Amazon's item to item collaborative filter (Linden et al., 2003) has evolved into a sophisticated pipeline that drives product recommendations across the storefront ^[3]. Alibaba's Deep Interest Network (DIN, 2018) and Deep Interest Evolution Network (DIEN, 2019) introduced attention mechanisms over a user's historical behaviour, weighting past items by their relevance to the candidate ^[18]. Taobao runs TDM (Tree based Deep Models) for billion scale retrieval. Meta's feed and Reels recommenders use multi task DNN rankers; the company open sourced DLRM (2019) and the TorchRec library ^[20].

How are large language models changing recommendation?

Large language models have begun to reshape recommendation in three ways.

First, large language models serve as item encoders. The text describing an item (title, tags, reviews) is fed to a pretrained model such as Llama or a sentence transformer, and the resulting embedding is consumed by a downstream ranker. This sidesteps cold start because new items inherit semantic meaning from text alone.

Second, LLMs power conversational recommenders. Spotify's AI DJ, Amazon's Rufus (a generative shopping assistant launched in 2024), and Meta AI inside Instagram all let users describe what they want in natural language ^[32].

Third, recent research treats recommendation as language modelling. Approaches such as P5 (Geng et al., RecSys 2022) ^[23], LLaRA, RecLLM, GenRec, and TallRec serialise user histories into token sequences and fine tune an LLM to predict the next item identifier. Generative retrieval with Semantic IDs (Rajput et al., NeurIPS 2023) replaces hashed item IDs with hierarchically clustered semantic codes the model can decode token by token ^[24]. Apple's MLX framework (released December 2023) has been used to fine tune small recommender LLMs that run on device ^[31].

Whether LLM recommenders will displace classical two tower retrieval at scale is unsettled, since latency and serving cost remain orders of magnitude higher. As of 2026, most production deployments use LLMs for offline candidate enrichment and explanation rather than online ranking.

How are recommendation systems evaluated?

Recommender quality is measured along several axes. Offline metrics use a held out test set; online metrics come from A/B tests in production.

Scroll sideways for more →

NDCG is the most reported ranking metric in academic papers because it handles graded relevance and position bias gracefully. A/B tested click through rate, watch time, and retention drive production decisions, since offline metrics correlate imperfectly with online behaviour.

How do recommenders handle diversity, serendipity, and fairness?

Maximising predicted relevance can produce filter bubbles and echo chambers, and it can amplify popularity bias. Counterweights include:

• Maximal marginal relevance (Carbonell and Goldstein, 1998) re-ranks results to balance relevance against novelty.
• Determinantal point processes (DPPs) sample diverse subsets by maximising the determinant of a kernel matrix (Kulesza and Taskar, 2012).
• Calibrated recommendations (Steck, RecSys 2018) ensures the topic distribution of recommendations matches the user's historical distribution.
• Fairness aware ranking constrains exposure across protected groups; works include FA*IR (Zehlike et al., 2017) and Equity of Attention (Biega et al., 2018).
• Provider side fairness ensures that creators or sellers receive a fair share of impressions, an active research direction at Spotify, Etsy, and Airbnb.

The European Union's Digital Services Act requires very large online platforms to offer at least one recommender option that is not based on profiling (Article 38); the regulation became fully applicable in February 2024 ^[33].

What are the best recommender system libraries?

Scroll sideways for more →

A brief timeline

Scroll sideways for more →

Index of recommendation system terms on this wiki

• candidate generation
• collaborative filtering
• item matrix
• items
• matrix factorization
• recommendation system
• re-ranking
• scoring
• user matrix
• Weighted Alternating Least Squares (WALS)

References

1. Goldberg, D., Nichols, D., Oki, B. M., and Terry, D. (1992). *Using collaborative filtering to weave an information tapestry.* CACM 35(12). ↩
2. Resnick, P. et al. (1994). *GroupLens: An open architecture for collaborative filtering of netnews.* CSCW. ↩
3. Linden, G., Smith, B., and York, J. (2003). *Amazon.com recommendations: item to item collaborative filtering.* IEEE Internet Computing 7(1), 76-80. ↩
4. Funk, S. (2006). *Netflix Update: Try This at Home.* sifter.org. ↩
5. Koren, Y. (2008). *Factorization meets the neighborhood.* KDD. ↩
6. Hu, Y., Koren, Y., and Volinsky, C. (2008). *Collaborative filtering for implicit feedback datasets.* IEEE ICDM. ↩
7. Salakhutdinov, R., and Mnih, A. (2008). *Bayesian probabilistic matrix factorization using MCMC.* ICML. ↩
8. Rendle, S. (2010). *Factorization Machines.* IEEE ICDM. ↩
9. Sedhain, S. et al. (2015). *AutoRec: Autoencoders meet collaborative filtering.* WWW. ↩
10. Cheng, H. T. et al. (2016). *Wide and Deep Learning for Recommender Systems.* DLRS, RecSys. ↩
11. Covington, P., Adams, J., and Sargin, E. (2016). *Deep Neural Networks for YouTube Recommendations.* RecSys. ↩
12. Gomez Uribe, C. A., and Hunt, N. (2016). *The Netflix recommender system: algorithms, business value, and innovation.* ACM TMIS 6(4). ↩
13. Hidasi, B. et al. (2016). *Session based recommendations with recurrent neural networks.* ICLR. ↩
14. He, X. et al. (2017). *Neural Collaborative Filtering.* WWW. ↩
15. Guo, H. et al. (2017). *DeepFM.* IJCAI. ↩
16. Kang, W. C., and McAuley, J. (2018). *Self attentive sequential recommendation.* IEEE ICDM. ↩
17. Ying, R. et al. (2018). *Graph convolutional neural networks for web scale recommender systems (PinSage).* KDD. ↩
18. Zhou, G. et al. (2018). *Deep Interest Network for click through rate prediction.* KDD. ↩
19. Sun, F. et al. (2019). *BERT4Rec.* CIKM. ↩
20. Naumov, M. et al. (2019). *DLRM.* arXiv:1906.00091. ↩
21. He, X. et al. (2020). *LightGCN.* SIGIR. ↩
22. Rendle, S. et al. (2020). *Neural collaborative filtering vs. matrix factorization revisited.* RecSys. ↩
23. Geng, S. et al. (2022). *P5: Recommendation as language processing.* RecSys. ↩
24. Rajput, S. et al. (2023). *Recommender systems with generative retrieval.* NeurIPS. ↩
25. MacKenzie, I., Meyer, C., and Noble, S. (2013). *How retailers can keep up with consumers.* McKinsey and Company. ↩
26. Netflix (2009). *Grand Prize awarded to team BellKor's Pragmatic Chaos.* netflixprize.com. ↩
27. *YouTube's recommendations drive 70% of what we watch.* Quartz (2018), reporting remarks by YouTube chief product officer Neal Mohan at CES. ↩
28. Spotify Engineering (2015). *What made Discover Weekly one of our most successful feature launches to date?* engineering.atspotify.com. ↩
29. Spotify Technology S.A. (2026). *Spotify Reports Fourth Quarter 2025 Earnings.* newsroom.spotify.com. ↩
30. Spotify (2023). *Spotify Debuts a New AI DJ, Right in Your Pocket.* newsroom.spotify.com. ↩
31. Apple Machine Learning Research (2023). *MLX: An array framework for Apple silicon.* github.com/ml-explore/mlx. ↩
32. Amazon (2024). *Amazon announces Rufus, a new generative AI powered conversational shopping experience.* aboutamazon.com. ↩
33. European Union (2022). *Regulation (EU) 2022/2065 (Digital Services Act), Article 38.* Fully applicable 17 February 2024. ↩