How to Cite Personalization Research in 2026: The Papers That Actually Transfer

TL;DR.

• The most-cited neural recommendation papers cannot beat a well-tuned ItemKNN when baselines receive equivalent hyperparameter search — Dacrema et al. documented this in 2019 and the field never fully corrected course.
• Standard benchmark datasets apply 10-core filtering, which removes cold users entirely; nearly every published NDCG number is measuring warm-user performance on a problem your product does not have.
• The offline-online gap is large and measured: a paper reporting +5% NDCG is consistent with a neutral or negative A/B result.
• In 2026, the most production-relevant research has shifted to LLM-augmented cold start, knowledge graph traversal, and contrastive representation learning — not matrix factorization variants.
• If you only remember one thing: before citing a paper to justify a design decision, check whether its evaluation protocol includes cold-start users — if it does not, you are citing warm-user benchmarks for a problem that starts cold.

Personalization research is one of the most cited and least replicated sub-fields in machine learning. Engineers building recommender systems and knowledge graph-backed personalization layers cite NCF, BERT4Rec, and SASRec the way frontend engineers cite Stack Overflow answers — quickly, with confidence, without reading the evaluation section. In 2026, with LLM-augmented recommenders entering production and day-zero personalization becoming a first-class product requirement, the stakes for misapplied citations are higher. A citation that doesn't match your evaluation conditions isn't supporting your architectural claim — it's laundering it. This post is a practitioner's guide to the recommender systems literature: which papers earned their place in the canon, which are benchmarking folklore, and which 2024-2026 results are worth building on.

Why the Most-Cited Papers May Not Support Your Claim

The short version: evaluation protocols in recommender systems research have been systematically inconsistent, and the inconsistency usually favors the proposed method.

The landmark paper here is Dacrema, Cremonesi, and Jannach's "Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Recommendation Approaches" (RecSys 2019). They retested seven top-tier neural recommendation papers from 2015-2018 against properly hyperparameter-tuned baselines. In 7 of 9 cases, simple algorithms like ItemKNN-CF matched or outperformed the neural models when the baselines received equivalent tuning effort. The neural models hadn't been wrong — the baselines had been undertuned, making improvements look larger than they were.

This is not a historical artifact. Rendle et al.'s 2022 analysis (arXiv 2110.14037) showed that iALS — implicit Alternating Least Squares, a method from 2008 — still competes on standard benchmarks when correctly regularized. The improvement curve in collaborative filtering research has been substantially flatter than the publication rate suggests.

What evaluation sections usually omit

Cold users are excluded. The 10-core or 5-core filtering step, standard across nearly all benchmark experiments, removes users with fewer than 10 (or 5) interactions before splitting into train/test. On MovieLens-1M, this removes roughly 30–40% of the original user base. On sparser datasets like Amazon Reviews, it removes 60–70%. What remains is a population of warm users with established preferences — exactly the users where any reasonable model performs adequately. Day-zero personalization is not in scope.

Negative sampling overstates precision. Most papers evaluate by sampling 100 random negative items and ranking the target positive against them. A real user is not choosing from 101 options. They are choosing whether to engage at all, against a catalog of 500,000 items. Sampled evaluation significantly inflates the apparent discriminative ability of every model tested.

Metrics favor point-in-time snapshots. NDCG@10, HR@10, and MRR are computed on a single held-out positive per user. Multi-positive evaluation — where users have multiple relevant items in test — is rarely used, even though it better reflects engagement in catalog-scale systems.

The Canonical Papers: What Earned Their Citations

Collaborative filtering foundations

Matrix factorization for recommendations was formalized in the Netflix Prize literature. Koren, Bell, and Volinsky's "Matrix Factorization Techniques for Recommender Systems" (IEEE Computer, 2009) established the decomposition framing that every subsequent model builds on. It remains the correct citation for explaining why user and item latent vectors can be dot-producted to produce relevance estimates.

Rendle et al.'s "BPR: Bayesian Personalized Ranking from Implicit Feedback" (UAI 2009) introduced the pairwise ranking objective that trains on implicit signals by contrasting observed interactions against unobserved ones. Nearly every modern recommendation model training on clicks or engagement uses a variant of BPR whether the authors name it or not. If you are training on implicit feedback, cite BPR — it is more honest and more specific than citing a neural wrapper around the same objective.

Sequential recommendation

Kang and McAuley's SASRec "Self-Attentive Sequential Recommendation" (ICDM 2018) introduced transformer-based sequential recommendation and remains a practical default for session-aware personalization. It uses causal attention, which means it can be run autoregressively at inference without future-token leakage — a practical advantage over bidirectional alternatives.

Sun et al.'s BERT4Rec (CIKM 2019) applies bidirectional masked item prediction, analogous to MLM pre-training. It generally outperforms SASRec on benchmark datasets. However, Petrov and Macdonald (RecSys 2022) showed the comparison was confounded by inconsistent evaluation protocols — different negative sampling, different test set construction, different sequence length limits. When evaluation is harmonized, SASRec and BERT4Rec are roughly equivalent across most datasets. If you are citing BERT4Rec to justify a bidirectional architecture over a causal one, the empirical justification is weaker than the original paper suggests.

Knowledge graph-backed personalization

The most practically useful papers in this subfield form a clear lineage.

RippleNet (Wang et al., 2018, arXiv 1803.03467) propagates user preferences outward from seed entities along knowledge graph edges, producing a preference ripple that captures multi-hop semantic relationships without requiring a learned graph embedding. This is the right citation when arguing for graph-based preference propagation at cold start — the mechanism is intuitive and the connection to seed-entity initialization is direct.

KGAT: Knowledge Graph Attention Network (Wang et al., 2019, arXiv 1905.07854) applies attention over graph neighborhoods, weighting relational paths by their predictive relevance to a specific user. KGAT is the standard citation when justifying attention-weighted graph traversal in a production personalization system.

For a broader survey, Guo et al.'s "A Survey on Knowledge Graph-Based Recommender Systems" (arXiv 2003.00911) covers 40+ papers with a consistent taxonomy: embedding-based methods (TransE, TransR applied to recs), path-based methods (PER, MCRec), and propagation-based methods (RippleNet, KGCN, KGAT). This taxonomy is useful when framing an architecture choice in a design doc where reviewers may not be familiar with the subfield.

One important caveat: most KG-rec papers evaluate on Yelp-2018, MovieLens-20M, or Amazon Book with 10-core filtering and clean, structured side information. If your item catalog does not have a pre-built knowledge graph, the gains from KGAT require the additional investment in graph construction — something the papers do not account for in their reported cost.

The Offline-Online Gap: Why +5% NDCG Is Not +5% in A/B Tests

This is the most practically important thing to understand when reading any recommender systems paper.

Offline metrics (NDCG, HR, MRR, Recall) measure whether a model can predict which items a user interacted with in a held-out test period. Online A/B metrics (CTR, session depth, engagement time, D7 retention) measure whether users do something differently when the model is live. These are not measuring the same thing. The correlation between them is weak.

Kohavi, Tang, and Xu's "Trustworthy Online Controlled Experiments" (Cambridge University Press, 2020), synthesizing years of experimentation at Microsoft and other large-scale properties, documents that the majority of seemingly compelling offline improvements fail to replicate in online A/B tests. Zhao et al.'s "Revisiting Is Offline Evaluation for Recommender Systems Meaningful?" showed that in their production dataset, NDCG improvement had a Pearson correlation of roughly 0.3 with online click improvement — meaning offline NDCG explains approximately 9% of variance in online CTR.

Three mechanisms drive the gap:

Position bias is absent from offline evaluation. Items at rank 1 receive 3–10× the clicks of items at rank 5 regardless of relevance. A model that improves NDCG by surfacing a different item at the top position may be capturing position effects, not relevance gains.

Exposure effects are unmodeled. What a user saw before interacting with your recommendations changes their behavior. Repeat exposure increases CTR; novelty effects suppress it. None of this is present in offline datasets collected under the previous model.

Feedback loops are invisible. Your test set was collected under the previous model's ranking decisions. Items your model would surface that the previous model buried were never shown — so your model's genuine discoveries are invisible to offline evaluation.

When a paper reports offline NDCG improvements to justify an architectural choice, that justification is suggestive, not conclusive. It is evidence, not proof. Cite accordingly.

Cold Start and Day-Zero: The Papers Most Products Actually Need

Day-zero personalization — producing useful recommendations for a new user with zero interaction history — is the problem that almost no benchmark paper evaluates. It is also the most important problem for most products: new user retention is largely determined in the first session, before any interaction history exists.

The research literature approaches cold start three ways worth understanding.

Meta-learning cold start

MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation (KDD 2020) uses a task memory and a feature memory to enable fast adaptation to new users from 1–5 preference signals. It is one of the more directly applicable papers for products that run an onboarding flow and collect a small number of explicit preference signals before entering the main feed. The meta-learning framing — treat each new user as a new few-shot task — maps cleanly onto a product onboarding sequence.

LLM-augmented cold start

In 2024-2026, the most production-relevant cold-start research has been LLM-based preference elicitation and synthetic clone initialization. Rather than defaulting to popularity-weighted random recommendations, you can use a language model to generate a preference prior from contextual signals, map natural language preference statements from an onboarding flow to item-space embeddings, and bootstrap a preference vector before any interaction is logged.

TALLRec (Bao et al., 2023, arXiv 2305.00447) fine-tunes LLaMA on recommendation tasks framed as instruction-following and shows cross-domain transfer with few-shot prompting. Its movie-to-book transfer experiments are the closest the published literature gets to cross-modal cold start — a user whose preferences are known in one domain being initialized in another.

P5 (RecSys 2022) unified five recommendation tasks into a single seq2seq framework. It is worth citing when justifying a unified LLM-backed recommendation layer, but its inference cost is significantly higher than embedding-based approaches — a constraint the paper does not surface clearly.

The 2024-2026 Papers Worth Watching

The field has largely settled on hybrid architectures: a fast traditional model (collaborative filtering, ANN retrieval) for the recall stage, combined with an LLM or reranker for scoring and explanation. The LLM-as-full-ranker proposals from 2023 hit a wall when deployed at catalog scale — ranking latency and cost make them impractical for retrieval over millions of items.

LightGCN (He et al., SIGIR 2020, arXiv 2002.02126) stripped NGCF of its non-linear transformations and feature matrices, showing the essential operation in GNN-based collaborative filtering is neighborhood aggregation alone. LightGCN is the current standard citation for graph-based CF in production — it is faster, simpler, and achieves comparable or better results than heavier GNN variants.

Contrastive learning for de-biasing addresses a practical problem that BPR-trained models have: popularity bias. Collaborative filtering embeddings overfit to popular items because popular items appear in more training pairs. Contrastive objectives applied to user representations partially correct this without requiring an explicit debiasing loss. Cite when justifying contrastive pre-training or hard negative mining in your embedding pipeline.

"Is ChatGPT a Good Recommender?" (Liu et al., 2023) and related zero-shot LLM evaluation work established a consistent finding: LLMs are competitive for explanation quality and cross-domain cold start, but systematically overweight well-known items when ranking. If you are using an LLM in your reranking or explanation layer, this is the citation for why you still need a popularity-aware recall stage upstream.

What to Build Your Citation Stack On

If you are writing a design doc, funding pitch, or technical blog post on personalization, here is a practical citation map:

For collaborative filtering foundations: Koren et al. 2009 (matrix factorization decomposition), Rendle 2009 (BPR for implicit feedback). Include Dacrema et al. 2019 if you are being honest about baseline comparisons — it is the citation that separates practitioners from paper-chasers.

For sequential recommendation: SASRec (Kang & McAuley 2018) for causal/autoregressive architectures; BERT4Rec (Sun et al. 2019) for bidirectional, with the Petrov & Macdonald caveat on evaluation inconsistency.

For knowledge graph-backed personalization: RippleNet for preference propagation from seed entities; KGAT for attention-weighted graph traversal; Guo et al. 2020 survey for the full taxonomy.

For cold start and day-zero: MAMO for meta-learning from 1–5 explicit signals; TALLRec for LLM-based few-shot adaptation; any 2025-2026 work on synthetic clone initialization.

For GNN-based CF in production: LightGCN as the simplified standard.

For explaining why offline metrics do not predict A/B results: Dacrema et al. 2019, Kohavi et al. 2020. These are the papers that explain why your NDCG number does not close the argument.

The underlying discipline is the same for every paper: read the evaluation section before you cite. Check whether cold users are included (they almost never are). Check whether the baselines were tuned with equivalent effort (they often were not). Check whether the dataset has the side information your catalog actually has (it usually does not). A citation that does not survive this three-question test is not supporting your claim — it is dressing it up in academic authority it did not earn. In 2026, the personalization research frontier is genuinely interesting and genuinely worth reading. Just read it critically.