Mutable Low-Rank Sketches for Retrain-Free Recommendation

Hector J. Garcia; Nick Clayton

REVIEW 2 major objections 6 minor 35 references

Mutable Low-Rank Sketches for Retrain-Free Recommendation

T0 review · 2 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A mutable KP-tree sketch lets recommenders update user embeddings on every rating in O(log n) time, with a proven monotonic error bound and no gradient computation.

desk verdict Clean idea and plausible empirical payload, but the headline monotonicity theorem is over-sold: Theorem 1 only bounds projected error by the shrinking raw error, and the 'never degrades' claim is false. read the letter →

arxiv 2607.15242 v1 pith:QLJQP576 submitted 2026-07-16 cs.LG

Hector J. Garcia , Nick Clayton This is my paper

classification cs.LG

keywords mutablesketchesKP-treeretrain-freerecommendationembeddingstalenesslow-rankprojectionincrementalmatrixfactorizationcold-startlatencynorm-proportionalsampling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to eliminate embedding staleness in two-stage recommender systems by decoupling data freshness from model freshness. It stores each user's preference vector in a KP-tree, a mutable sparse segment tree, fits a fixed low-rank item basis once from a small sketch, and recomputes the user's embedding by projection on every new rating. The central theoretical claim is Theorem 1: each newly observed rating strictly reduces the projected error envelope, a guarantee the authors say incremental matrix factorization baselines lack. If correct, the method gives sub-millisecond cold-start personalization, logarithmic-time user-side updates, and competitive accuracy on dense data while reading a tiny fraction of the rating matrix.

What carries the argument

The KP-tree is a sparse segment tree with sum aggregation: leaves store the magnitude of each preference value, internal nodes store child sums, and it supports sample, query, and update all in O(log n). Its mutation capability is the load-bearing mechanism — after a user rates an item, an insert propagates through internal sums, keeping the sampling distribution valid and enabling immediate sketch patching. The fixed low-rank basis V_k, fitted once from a sublinear sketch via truncated SVD, acts as the projection that maps a user's current preference vector to an embedding.

What would settle it

Construct a streaming experiment with a synthetic or real dataset where a sudden, strong distribution shift occurs after V_k is fit (e.g., a new item category becomes dominant). If RMSE degrades sharply and does not recover without a full V_k refit — or if the paper's claimed monotonic improvement fails to translate to any reduction in test RMSE for affected users — the retrain-free accuracy claim is refuted under drift.

Watch

Extended reading notes

Core claim

The paper claims that a KP-tree-based mutable sketch, combined with a one-time low-rank projection, enables retrain-free user embedding updates that improve monotonically as ratings arrive. The proof of Theorem 1 shows that when a user's true preference for an item is revealed, the squared error of the projected embedding decreases by the square of that rating value, because an orthogonal projection cannot increase the norm of the error vector. On KuaiRec, the sketch reaches 0.810 RMSE while reading 1.8% of the matrix, beating full-data ALS at 0.822, and delivers a personalized recommendation to a new user in under a millisecond after their first rating. The authors also show that norm-propo

Load-bearing premise

The fixed item basis V_k, fitted once from a small sketch, must remain a good projection space for all future users and ratings; if the underlying preference distribution drifts, the projection becomes stale and the accuracy advantage erodes until a refit occurs.

Editorial extensions

If this is right

If the monotonic improvement guarantee holds, a serving recommender can promise that a user's embedding only gets closer to its projection of the true preference vector as more ratings arrive, without ever touching the model parameters.
Per-batch update cost drops to O(b log n) with no gradient computation, so data freshness no longer couples to model recomputation in systems where ALS or eALS would require minutes of refitting.
Cold-start latency becomes sub-millisecond for the first personalized recommendation, enabling instant personalization for brand-new users in production flows.
The consistent sampling property allows the sketch itself to be patched or rebuilt from current tree state without scanning the full rating matrix, making drift-triggered refits cheap.
On sparse datasets, norm-proportional sampling gives materially better item coverage than uniform sampling, which could change how production systems choose sketch construction strategies based on density.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The monotonicity guarantee is about the projected-error envelope, not prediction error in raw rating space; a user whose true preferences lie largely outside span(V_k) may see little absolute improvement even as the bound shrinks, so the practical value depends on how well the fixed basis captures the long-term item structure.
The framework likely extends beyond collaborative filtering: any evolving sparse signal (e.g., ad click streams, sensor logs) where rows are users and columns are features could use the same fixed-basis-plus-mutable-sketch pattern for fresh row embeddings without retraining.
The paper's own density sweep suggests a hybrid deployment: use norm-proportional KP-tree sampling for sparse long-tail items and switch to uniform sampling or alias tables for dense head items, since the two regimes have opposite optimal strategies.
The drift-triggered refit controller is essentially a change-point detector on the sampling distribution; its Tier 1 behavior on the evaluated datasets implies that for slowly drifting data, retrain-free operation may hold for long stretches, but a sudden popularity shock would force a full rebuild — a failure mode not stress-tested in the current experiments.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth a look, but the headline theorem is not what it claims. What's actually new is a mutable KP-tree sketch for user-side embedding updates: fit a low-rank basis once, then recompute embeddings by projecting updated preference vectors through the fixed basis. That's a clean idea, and the KuaiRec numbers are plausible: 0.810 RMSE at 1.8% data read vs ALS 0.822 at full data, with sub-millisecond cold-start latency. The density crossover for norm-proportional vs uniform sampling (around 5%) is also a useful empirical result.

The soft spot is Theorem 1. The proof correctly shows ||P e(t+1)|| <= ||e(t+1)|| < ||e(t)||, but that only means the projected error is bounded by the shrinking full error. It does not show the projected error itself decreases. In fact it can increase: the stress-test example with V_k = span((1,1,-1)/sqrt(3)) and a*=(0,1,2) has ||P e(0)|| = 1/sqrt(3) but ||P e(1)|| = 2/sqrt(3) after one true rating. So the "projection never degrades" statement is false, and the claimed "guarantee that FunkSVD and eALS lack" is not established. The inequality is a trivial consequence of observing true values, independent of the KP-tree.

This matters because the monotonic guarantee is a central selling point. The empirical work may still stand, but the theoretical contribution needs a serious rewrite: state the weaker bound honestly or find a different guarantee.

Other issues: the "retrain-free" claim is softened by their own drift-triggered refit controller, which is reasonable operationally but undercuts the title. Code and datasets are "available upon request," which I treat with suspicion. Sparse baselines are limited; ALS at 4.62 on Amazon Music is a red flag about baseline tuning.

The paper is well-written and the system design is thoughtful. The empirical crossover and latency measurements are worth having. But I would not cite the monotonicity claim as stated, and I'd want the theorem fixed before endorsing.

Recommendation: send to peer review with a clear request to fix the theorem interpretation. The operational idea is real and the false claim is fixable. The mismatch between proof and prose is an internal contradiction, so I'd flag the paper as not yet serious about its own theory.

Referee Report

2 major / 6 minor

Summary. The paper proposes 'mutable low-rank sketches' for recommendation, in which each user's preference vector is stored in a KP-tree (a sparse segment tree with sum aggregation) that supports O(log n) point updates and weighted sampling. A fixed low-rank basis V_k is fit once from a sublinear sketch, and user embeddings are recomputed on the fly by projecting the current preference vector through V_k. The authors claim a monotonic improvement guarantee (Theorem 1) that FunkSVD and eALS lack, and report sub-millisecond cold-start latency, competitive RMSE on KuaiRec at low data read, and a density-dependent advantage of norm-proportional sampling over uniform sampling. Evaluation covers six datasets and a drift-triggered refit controller.

Significance. If the claims held, the method would be a practical step toward retrain-free serving: O(log n) per-rating user-side updates, no gradient computation, and sub-millisecond cold-start. The experimental evaluation is broad, spanning dense and ultra-sparse datasets, and the KP-tree's consistent weighted sampling under mutation is a genuinely useful data-structure property. However, the central theoretical guarantee of monotonic improvement is not established by Theorem 1 as stated; the proved inequality is a generic bound on orthogonal projections and does not imply monotonicity of the projected error. The paper's value therefore rests on the empirical results and the data-structure contribution, which need to be separated from the unsupported monotonicity claims. The complexity analysis also needs correction before the scalability claims can be accepted.

major comments (2)

[Section 2.2, Abstract, Table 4] Theorem 1 does not prove that 'the projection never degrades' or that the projected error is monotone. The proof shows only ||P e(t+1)|| <= ||e(t+1)|| < ||e(t)||, but never compares ||P e(t+1)|| with ||P e(t)||. Since P is an orthogonal projection, ||P e(t)|| can be much smaller than ||e(t)||, and the projected error can increase. Counterexample: take V_k = span((1,1,-1)/sqrt(3)), a*=(0,1,2), a(0)=(0,0,0), a(1)=(0,1,0). Then ||P e(0)|| = 1/sqrt(3) but ||P e(1)|| = 2/sqrt(3). Thus the claims in the Abstract and Section 2 are false as stated. Moreover, the inequality is a generic property of any orthogonal projection of observed ratings and does not depend on the KP-tree, so it does not establish a guarantee that FunkSVD/eALS lack. Table 4's 'Monotonic? Yes' row is unsupported. Please either prove a genuine monotonicity statement under additional assumptions or remove these claims.
[Section 2.1, Table 1] The claimed construction cost O(m + rc log n) 'independent of nnz' is not justified. After sampling r rows and c columns per row, the text says 'Extract values at the union of sampled columns to form an r x |C| sketch S', which requires reading r*|C| entries, not rc entries. On dense data, |C| can be nearly as large as the number of items, so the data read is much larger than rc. Furthermore, constructing the KP-trees from a raw rating matrix requires processing all nnz entries. Thus the total fit/retrain cost is at least O(nnz + r|C| log n) unless a KP-tree feature store is already available. The '1.8% data read' in the abstract corresponds to r|C|/total matrix size, not rc/total. Please correct the complexity model and qualify the 'independent of nnz' claim.

minor comments (6)

[Section 2.2] The noise extension says the bound is 'monotone whenever the signal exceeds the noise variance', but this refers to the bound, not to the actual RMSE for a single user. Clarify to avoid implying a stronger guarantee.
[Tables 7-8 placement] Table 8 appears after Section 5 (Conclusion) but is referenced in Section 4. Move it to the section discussing serving latency.
[Sections 2.3 and 3.4] The density crossover is stated as 'below ~5%' in Section 2.3 and the Abstract, but Section 3.4 says 'roughly 5–10%'. Make these consistent.
[Section 3.1] The text says Amazon Video Games yields 1.562 RMSE 'with norm-proportional sampling', but Table 6 shows 1.485 without bias and 1.562 with bias. Clarify which configuration is being cited.
[Figure 2] The log-scale values printed next to the bars (−0.71, −0.29, 5.3, 6.59) are unexplained. Add axis labels and a legend or caption describing what these numbers denote.
[Reproducibility] The paper states 'Code, datasets, and experiment scripts are available upon request' but provides no public repository. Consider releasing the code to support reproducibility of the reported latency and RMSE numbers.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain; the monotonicity proof is a genuine projection bound, while the stronger 'never degrades' sentence is an unsupported overclaim rather than a circular step.

full rationale

The paper's derivation is self-contained. Theorem 1's proof starts from the update rule e(t+1) = e(t) with the newly observed component zeroed, then applies the orthogonal-projection inequality ||P x|| <= ||x||; it does not assume its conclusion. The displayed bound ||P e(t+1)|| <= ||e(t+1)|| < ||e(t)|| is a direct consequence of the stated update and of the fixed-basis projection, so it is not an input-output circularity. The separate sentence in Section 2 that 'the projection never degrades as observations accumulate' is stronger than the theorem and is not proved -- the theorem only bounds the new projected error by the old full error, not by the old projected error -- but this is a correctness/proof gap, not a circular reduction. The experimental claims are also not circular: all experiments fix k=10 and sketch size 200x100 (with one 800x200 ablation) and use held-out splits, so no fitted parameter is renamed as a prediction. The only overlapping-author citation is [8] (H. Garcia's technical report), used for implementation details of the drift-triggered refit controller; it is not load-bearing for the theorem, accuracy, latency, or sampling results. No uniqueness theorem is imported from the authors, and no ansatz is smuggled in via self-citation. The central empirical and algorithmic content stands independently of the cited prior work.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the low-rank assumption and the representative-ness of the one-time sketch. The only hand-chosen numbers are the rank and sketch size. No entirely new entities (particles, forces, dimensions) are introduced; the KP-tree is an existing data structure reused with a new purpose.

free parameters (2)

Rank k = 10
All experiments use rank k=10. This is a hyperparameter chosen by hand, not fit to optimize a specific dataset, but it is a free parameter that affects the projection quality.
Sketch dimensions (r, c) = 200, 100
The sketch samples 200 rows and 100 columns per row. This is a fixed experimental configuration, chosen by hand, and determines the data-read fraction and accuracy.

assumptions (4)

domain assumption The rating matrix is approximately low-rank, so a fixed low-rank basis V_k captures the latent structure.
Invoked in Section 2.1 when fitting a low-rank projection from a sketch; if the matrix is not low-rank, the projection cannot represent the data well.
domain assumption A basis V_k fitted once from a sublinear sketch generalizes to unseen users and items.
The method holds V_k fixed for all future updates; Section 4 concedes that this basis accumulates staleness under distribution shift.
ad hoc to paper Observed ratings equal true preferences (for the monotonicity proof).
Theorem 1's proof sets the observed entry to its true value. The paper appended a noise note with an additive σ² term, which is not part of the formal theorem and weakens the guarantee.
standard math KP-tree operations (update, query, weighted sample) run in O(log n) and maintain correct sums.
Standard segment-tree properties, cited from prior data-structure work; used throughout Section 2.1 without proof.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mutable Low-Rank Sketches for Retrain-Free Recommendation." pith.science (2026). https://pith.science/paper/QLJQP576

@misc{pith2026260715242,
  author       = {Pith},
  title        = {Pith review of: Mutable Low-Rank Sketches for Retrain-Free Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QLJQP576}},
  note         = {Machine review of arXiv:2607.15242}
}

read the original abstract

A common bottleneck in two-stage recommendation is embedding staleness: when a user rates a new item, their embedding remains fixed until the next retrain cycle. We propose mutable sketches, which store each user's preferences in a KP-tree (a sparse segment tree with sum aggregation), fit a low-rank projection once, and recompute embeddings on-the-fly as ratings arrive. We prove that each new observation monotonically tightens the prediction error envelope (Theorem 1), a guarantee that FunkSVD and eALS lack. On KuaiRec, the mutable sketch achieves 0.810 RMSE at 1.8% data read vs. ALS 0.822 at 100%, with 8x faster per-batch updates. A new user receives personalized recommendations in <1 ms after their first rating, with no model retraining required. A comparison of sampling strategies across density regimes shows that the KP-tree's norm-proportional sampling provides 40-130% better item coverage on sparse data (<1% density), while uniform sampling suffices on dense matrices.

Figures

Figures reproduced from arXiv: 2607.15242 by the authors.

**Figure 2.** Cold-start latency vs. per-batch update cost of in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Density sweep on KuaiRec subsamples (bias [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 3 linked inside Pith

[1]

Juan Miguel Arrazola, Alain Delgado, Bhaskar Roy Bardhan, and Seth Lloyd
[2]

Matthew Brand. 2006. Fast low-rank modifications of the thin singular value decomposition.Linear Algebra Appl.415, 1 (2006), 20–30

2006
[3]

Clarkson, Lior Horesh, Hongyuan Lin, and David P

Natalie Chepurko, Kenneth L. Clarkson, Lior Horesh, Hongyuan Lin, and David P. Woodruff. 2022. Quantum-inspired algorithms from randomized numerical linear algebra. InProc. 39th International Conference on Machine Learning (ICML). 3879– 3900

2022
[4]

Nai-Hui Chia, András Gilyén, Tongyang Li, Han-Hsuan Lin, Ewin Tang, and Chunhao Wang. 2020. Sampling-based sublinear low-rank matrix arithmetic framework for dequantizing quantum machine learning. InProc. 52nd ACM Symposium on Theory of Computing (STOC). 387–400

2020
[5]

Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for YouTube recommendations. InProc. 10th ACM Conference on Recommender Systems (RecSys). 191–198

2016
[6]

Ernesto Diaz-Aviles, Lucas Drumond, Lars Schmidt-Thieme, and Wolfgang Nejdl
[7]

Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. 2022. KuaiRec: A fully-observed dataset and insights for evaluating recommender systems. InProc. 31st ACM CIKM. 540–550

2022
[8]

H. Garcia. 2026.Multi-backend storage for KP-tree recommendation: Spark, Flink, and JVM implementations. Technical Report. University of Michigan

2026

Show all 35 references

[9]

Nathan Halko, Per-Gunnar Martinsson, and Joel A. Tropp. 2011. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions.SIAM Rev.53, 2 (2011), 217–288

2011
[10]

Maxwell Harper and Joseph A

F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens datasets: History and context.ACM Transactions on Interactive Intelligent Systems5, 4 (2015), 19:1– 19:19

2015
[11]

Bowei He, Xu He, Renrui Zhang, Yingxue Zhang, Ruiming Tang, and Chen Ma. 2023. Dynamic embedding size search with minimum regret for streaming recommender system. InProc. 32nd ACM International Conference on Information and Knowledge Management (CIKM)

2023
[12]

Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProc. 26th International Conference on World Wide Web. 173–182

2017
[13]

Xiangnan He, Hanwang Zhang, Min-Yen Kan, and Tat-Seng Chua. 2016. Fast matrix factorization for online recommendation with implicit feedback. InProc. 39th International ACM SIGIR Conference. 549–558

2016
[14]

Yupeng Hou, Jiacheng Zhang, Zhankui Lin, Hongzhi Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Bridging language and items for retrieval and recommendation.arXiv preprint arXiv:2403.03952(2024)

2024 arXiv
[15]

Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative filtering for implicit feedback datasets. InProc. IEEE International Conference on Data Mining (ICDM). 263–272

2008
[16]

Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs.IEEE Transactions on Big Data7, 3 (2019), 535–547

2019
[17]

Iordanis Kerenidis and Anupam Prakash. 2016. Recommendation systems.arXiv preprint arXiv:1603.08675(2016)

2016 arXiv
[18]

Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems.Computer42, 8 (2009), 30–37

2009
[19]

Edo Liberty. 2013. Simple and deterministic matrix sketching. InProc. 19th ACM SIGKDD. 581–588

2013
[20]

Zhuoran Liu, Leqi Zou, Xuan Zou, Caihua Wang, Biao Zhang, Da Tang, Bolin Zhu, Yijie Zhu, Peng Wu, Ke Wang, and Youlong He. 2022. Monolith: Real time recommendation system with collisionless embedding table.arXiv preprint arXiv:2209.07663(2022)

2022 arXiv
[21]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt- Thieme. 2009. BPR: Bayesian personalized ranking from implicit feedback. In Proc. 25th Conference on Uncertainty in Artificial Intelligence (UAI). 452–461

2009
[22]

Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. 2002. Incremen- tal singular value decomposition algorithms for highly scalable recommender systems. InProc. 5th International Conference on Computer and Information Sci- ence

2002
[23]

Chijun Sima, Yao Fu, Man-Kit Sit, Liyi Guo, Xuri Gong, Feng Lin, Junyu Wu, Yongsheng Li, Haidong Rong, Pierre-Louis Aublin, and Luo Mai. 2022. Ekko: A Large-Scale Deep Learning Recommender System with Low-Latency Model Up- date. In16th USENIX Symposium on Operating Systems Des...

2022
[24]

Ewin Tang. 2019. A quantum-inspired classical algorithm for recommendation systems. InProc. 51st ACM Symposium on Theory of Computing (STOC)

2019
[25]

Alastair J. Walker. 1977. An efficient method for generating discrete random variables with general distributions.ACM Trans. Math. Software3, 3 (1977), 253–256. Hector J. Garcia and Nick Clayton

1977
[26]

Mengting Wan and Julian McAuley. 2018. Item recommendation on monotonic behavior chains. InProc. 12th ACM Conference on Recommender Systems (RecSys). 86–94

2018
[27]

Rui Yin, Wei Wen, Ke Li, Hanchen Wei, Tao Zhang, Yin Huang, and Dingming Li
[28]

Wenjun Yu, Sitian Chen, Cheng Chen, and Amelie Chi Zhou. 2025. Near-Zero- Overhead Freshness for Recommendation Systems via Inference-Side Model Updates.arXiv preprint arXiv:2512.12295(2025). To appear in IEEE HPCA 2026

2025
[29]

Xiangyu Zhao, Haochen Liu, Wenqi Fan, Hui Liu, Jiliang Tang, and Chong Wang
[30]

McNee, Joseph A

Cai-Nicolas Ziegler, Sean M. McNee, Joseph A. Konstan, and Georg Lausen
[2005]

Improving recommendation lists through topic diversification. InProc. 14th International Conference on World Wide Web (WWW). 22–32
[2012]

Real-time top-n recommendation in social streams. InProc. 6th ACM Conference on Recommender Systems (RecSys). 59–66
[2020]

Quantum-inspired algorithms in practice.Quantum4 (2020), 307

2020
[2021]

AutoEmb: Automated embedding dimensionality search in streaming recommendations. InProc. 21st IEEE International Conference on Data Mining (ICDM). 896–905
[2024]

VLDB Endowment 17, 9 (2024), 2149–2161

Optimal matrix sketching over sliding windows.Proc. VLDB Endowment 17, 9 (2024), 2149–2161

2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Juan Miguel Arrazola, Alain Delgado, Bhaskar Roy Bardhan, and Seth Lloyd

[2] [2]

Matthew Brand. 2006. Fast low-rank modifications of the thin singular value decomposition.Linear Algebra Appl.415, 1 (2006), 20–30

2006

[3] [3]

Clarkson, Lior Horesh, Hongyuan Lin, and David P

Natalie Chepurko, Kenneth L. Clarkson, Lior Horesh, Hongyuan Lin, and David P. Woodruff. 2022. Quantum-inspired algorithms from randomized numerical linear algebra. InProc. 39th International Conference on Machine Learning (ICML). 3879– 3900

2022

[4] [4]

Nai-Hui Chia, András Gilyén, Tongyang Li, Han-Hsuan Lin, Ewin Tang, and Chunhao Wang. 2020. Sampling-based sublinear low-rank matrix arithmetic framework for dequantizing quantum machine learning. InProc. 52nd ACM Symposium on Theory of Computing (STOC). 387–400

2020

[5] [5]

Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for YouTube recommendations. InProc. 10th ACM Conference on Recommender Systems (RecSys). 191–198

2016

[6] [6]

Ernesto Diaz-Aviles, Lucas Drumond, Lars Schmidt-Thieme, and Wolfgang Nejdl

[7] [7]

Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. 2022. KuaiRec: A fully-observed dataset and insights for evaluating recommender systems. InProc. 31st ACM CIKM. 540–550

2022

[8] [8]

H. Garcia. 2026.Multi-backend storage for KP-tree recommendation: Spark, Flink, and JVM implementations. Technical Report. University of Michigan

2026