Pith. sign in

REVIEW 5 major objections 6 minor 31 references

Radial Neighborhood Smoothing Recommender System

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that hidden user-item distances can be recovered from row- and column-wise distances in the observed rating matrix, and that a radial-neighborhood kernel estimator built on these distances consistently recovers true…

desk verdict An empirically promising recommender method whose consistency proof collapses under its own bandwidth and neighbor-set rates. read the letter →

arxiv 2507.09952 v1 pith:2PZJBVJ2 submitted 2025-07-14 cs.LG stat.APstat.ME

Zerui Zhang , Yumou Qiu This is my paper
classification cs.LGstat.APstat.ME MSC 62G0862G2015A83
keywords matrixcompletionlatentfactormodeldistanceestimationkernelsmoothingradialneighborhoodcollaborativefilteringcold-startsingularvaluedecomposition
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

Using only the partially observed, noisy rating matrix, this paper tries to establish that the unobserved latent geometry of users and items is readable from ordinary row and column distances. Concretely, it claims that the squared Euclidean distance between two users or two items in the observed matrix, after subtracting twice the estimated noise variance $2\hat{\sigma}^2$, converges to the squared distance between their latent feature vectors. On that foundation it builds the Radial Neighborhood Estimator: for each missing rating it pools all observed ratings whose row or column shares at least $\beta$ observed entries with the target, then smooths them by a local kernel regression weighted by the estimated latent distances. The paper argues this estimator is consistent for the true scores under a low-rank latent-factor model with missing-completely-at-random data, and that it outperforms collaborative filtering, soft-impute, and blind regression in simulations and on MovieLens, Jester, Last.fm, and Steam data. It also reports that the estimator keeps making predictions for users and items that appear only in the test set, which the other methods cannot do.

What carries the argument

The load-bearing object is the chain of distance identities in Propositions 1-3 and Theorem 1, which converts unobservable latent distances $\|x_u - x_v\|_2$ and $\|y_i - y_j\|_2$ into computable row- and column-wise squared distances in the observed matrix, subtracting $2\hat{\sigma}^2$ to remove the noise-induced non-centrality. On this sits the radial neighborhood set $N^{Rdl}_{ui}$, defined as all observed cells $(v,j)$ such that user $v$ shares at least $\beta$ rated items with $u$ or item $j$ shares at least $\beta$ raters with $i$; this set is shown in Lemma 2 to have size concentrating at $mn\pi^2$, much larger than the directly overlapping neighborhoods used by collaborative filtering or blind regression. The estimator itself is a Nadaraya-Watson-type kernel regression over that neighborhood using the recovered distances as kernel inputs, and the proof machinery is the kernel density estimate $\tilde{f}(x_u,y_i)$ of Lemma 3, whose convergence to its expectation carries the consistency argument for the two bias terms in Theorem 2.

What would settle it

A direct check: take the paper's own rate assumptions $|N^{Rdl}_{ui}| = \Theta_p(mn\pi^2)$ and $h_1 = h_2 = O_p(m^{-1/2}n^{-1/2}\pi^{-1})$, so $h_1 h_2 |N^{Rdl}_{ui}| = O_p(1)$, and compute $\mathrm{Var}(\tilde{f}(x_u,y_i))$ from Lemma 3; the leading term $\sigma^2 \iint K^2 / (h_1 h_2 |N^{Rdl}_{ui}|)$ is then $O_p(1)$ rather than $o_p(1)$. Simulating a low-rank matrix at these bandwidths and checking whether $\tilde{f}/E(\tilde{f})$ concentrates at 1 as $m,n$ grow would settle whether Lemma 3's convergence holds.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is a distance identity. Proposition 1 shows that in the completed score matrix $Z$, the row distance $\|Z_{(u)} - Z_{(v)}\|_2$ equals the latent user distance $\|x_u - x_v\|_2$, and likewise for columns; Proposition 2 shows that in the observed matrix $A$, the squared row distance $m^{-1}\|A_{(u)} - A_{(v)}\|_2^2$ is an estimator of the corresponding latent squared distance once $2\hat{\sigma}^2$ is subtracted; and Proposition 3 shows that the practical version computed from only the observed overlapping entries is close to the fully observed row distance. Theorem 1 assembles these into a convergence statement of the form $\hat{d}_{u,v}^2 - 2\hat{\sigma}^2 \to d_{u,v}^2$ at rate $O_p(m^{-1/2}n^{-1/2}\pi^{-1})$. The authors then define radial neighborhoods that include every observed rating within at least $\beta$ overlapping items of the target user or $\beta$ overlapping users of the target item, and prove in Theorem 2 that the kernel-smoothed estimator built on these neighborhoods is consistent for the true score $z_{u,i}$, with Theorem 3 giving consistency of the fully computable version that uses observed ratings and estimated distances.

Load-bearing premise

The proof assumes the kernel density estimate built from the radial neighborhood concentrates on its expectation at the chosen bandwidths; if, as the stated rates suggest, the leading variance term $\sigma^2/(h_1 h_2 |N^{Rdl}_{ui}|)$ does not vanish, then that concentration is the load-bearing assumption that would fail.

Editorial extensions

If this is right

  • Missing ratings can be imputed without requiring at least two directly overlapping ratings between the target user or item and the neighbor, because the L2 distance needs only one common observation.
  • Neighborhood size grows with the total observation probability: $|N^{Rdl}_{ui}|$ concentrates at $mn\pi^2$, so denser matrices give the kernel regression proportionally more local information.
  • Users or items with no ratings in the training set can still be scored whenever the other dimension has overlapping observations, which is the paper's cold-start mitigation.
  • Consistency of the practical estimator is two-step: the latent-distance-based estimator converges to the true score, and replacing latent distances by observed-distance estimates does not break convergence.
  • Because the distance estimate has rate $O_p(m^{-1/2}n^{-1/2}\pi^{-1})$, the variance correction and the bandwidth choice must be tracked together; the stated orders are what the theory rests on.

Reading between the lines

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

  • Editorial inference: the same distance identity could be used as a diagnostic for latent low-rank structure in any incomplete matrix problem: if estimated latent distances disagree with row and column distances after variance correction, the assumed factor model is doubtful.
  • Editorial inference: the subtraction of a single global $2\hat{\sigma}^2$ assumes homoscedastic noise; with per-item or per-user noise, distances would carry a position-dependent bias, so a local variance estimator is a natural but untested extension.
  • Editorial inference: radial neighborhoods are a graph-expansion construction; treating them as a connectivity rule could connect this estimator to graph-based or iterative collaborative filtering, where second-degree connections already play a role, but the paper does not pursue that link.
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.

Referee Report

5 major / 6 minor

Summary. The paper proposes a recommender-system method, Radial Neighborhood Smoothing (RNE), that estimates user and item latent-space distances from observed matrix entries, corrects for noise via an empirical variance estimator, constructs a "radial" neighborhood set, and applies kernel regression to impute missing ratings. The main theoretical claims are that the latent distance is consistently approximated by observed row/column distances (Theorem 1) and that the resulting estimator is consistent (Theorems 2 and 3). The paper also reports simulation and real-data experiments comparing RNE with collaborative filtering, softImpute, and blind regression, with particular attention to cold-start settings.

Significance. If the consistency results were correct, the paper would offer a useful conceptual bridge between neighborhood-based methods and latent-factor models, and the empirical results suggest that RNE can be competitive, especially for cold-start entries. The paper also provides a concrete relaxation of overlap requirements for neighborhood construction, which is a practically relevant idea. However, the theoretical section is the central contribution, and it contains load-bearing errors: Lemma 3 is inconsistent with Lemma 2, the variance computations do not yield the claimed vanishing rates, and the proof of Theorem 2 relies on these incorrect rates. The paper does not ship machine-checked proofs, code, or parameter-free derivations, and several key components (the first-step estimator used to define σ̂, the kernel bandwidth selection details, and the unspecified weight combination) are left unspecified. Given that the main consistency claim is unsupported, the contribution cannot be accepted in its current form.

major comments (5)
  1. [Lemma 3 (Section 6.6)] Lemma 3 assumes |N^{Rdl}_{ui}| = O_p(m^{-1} n^{-1} π^{-2}), but Lemma 2 (Section 3 and Appendix 6.5) proves that |N^{Rdl}_{ui}| concentrates to mnπ^2, i.e., |N^{Rdl}_{ui}| = Θ_p(mnπ^2). These are reciprocal orders, so the premise of Lemma 3 contradicts the paper's own established neighbor-set size. This inconsistency alone invalidates the proof structure for Lemma 3 and, consequently, for Theorem 2.
  2. [Section 6.6 (proof of Lemma 3)] Even if one adopts the correct |N^{Rdl}_{ui}| = Θ_p(mnπ^2), the leading variance term in the derived expression is 1/(h1 h2 |N^{Rdl}_{ui}|) ∫ K^2(a,b) f^*(ah1, bh2) da db. With h1 = h2 = O_p(m^{-1/2} n^{-1/2} π^{-1}) as stated in Lemma 3 and Theorem 2, this leading term is Θ_p(1), not o_p(1). The proof then claims Var(ef/E[ef]) = O_p(m^{-1} n^{-1} π^{-2}), a rate that would require h1 h2 |N^{Rdl}_{ui}| = Θ_p(mnπ^2), which is incompatible with the stated bandwidths. Therefore the convergence ef/E[ef] →_p 1 is not established.
  3. [Section 6.7, Eq. (24) (proof of Theorem 2)] The proof of bz_{u,i} → z_{u,i} relies on E{bm2^2} = σ^2/(h1 h2 |N^{Rdl}_{ui}|) ∫ K^2 f^* = O_p((h1+h2)mnπ^2). With |N^{Rdl}_{ui}| = Θ_p(mnπ^2) from Lemma 2 and h1 = h2 = O_p(m^{-1/2} n^{-1/2} π^{-1}), the last expression is O_p(1), so Var(bm2/ef) does not vanish. The same O_p(1) problem appears in the proof of Theorem 3, where Var(Δ) is computed as O_p(h1 h2 mnπ^2). Thus the central consistency claim is not supported by the supplied derivation.
  4. [Section 2.2] The variance correction depends on a "first-step estimator" bz^{(1)}_{v,j}, but this estimator is never defined. Proposition 2 and Theorem 1 simply assume |σ̂ − σ| = O_p(m^{-1/2} n^{-1/2} π^{-1}), without specifying any estimator that achieves this rate. Since the distance approximation and the practical RNE estimator both rely on σ̂, the method is not fully specified and the theoretical guarantee is conditional on an unstated construction.
  5. [Section 6.1 (proof of Proposition 3)] The variance calculation for bd^2_{u,v} concludes with Var(bd^2_{u,v} | A(u), A(v)) = O_p(m^{-1} π^{-1}), but then states the convergence rate as O_p(m^{-1/2} π^{-1}). Taking the square root of the variance gives O_p(m^{-1/2} π^{-1/2}), which is slower than the claimed rate for small π. Additionally, the variance of δ_{u,t}δ_{v,t} is written as π^2(1−π)^2, whereas for independent Bernoulli(π) variables it is π^2(1−π^2). These errors propagate into Proposition 3 and Theorem 1.
minor comments (6)
  1. [Lemma 1] The notation {a ∧ b} is defined as max{a,b}, which is nonstandard (∧ usually denotes min), and the proof derives a sum of two terms rather than a bound involving only the maximum of the two scaled distances; the statement and proof need to be reconciled.
  2. [Section 6.7, around Eq. (24)] There is a typographical mismatch: after computing E{bm2^2}, the text says "Then bm1(xu,yi)/ef(xu,yi) = Op((h1+h2)mnπ^2)", which appears to concern bm2, not bm1. The surrounding exposition should be corrected for clarity.
  3. [Section 6.2 (proof of Proposition 2)] The term "(zu,t − zu,t)" should presumably read "(zu,t − zv,t)"; as written, that term is identically zero.
  4. [Section 4] The numerical experiments state that "the bandwidths and the weight combination were optimized by 5-fold cross validation," but the kernel weights' precise form, the bandwidth grid, and the weight-combination formula are not described. This limits reproducibility of the reported empirical comparisons.
  5. [Section 4, Tables 1 and 2] For m = n = 300, k = 3, π = 0.6, RNE has higher RMSE than the collaborative filtering and softImpute baselines; the general claim that RNE "outperformed the other methods" should be qualified where it is not supported by the displayed numbers.
  6. [Section 5] The discussion section refers to the proposed method as "RSE", while the rest of the paper uses "RNE"; this should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the RNE consistency claims are asymptotic arguments against an external true matrix, and the noted Lemma 3 variance issue is a correctness concern, not a circularity.

full rationale

The paper's central target is the unobserved matrix Z and the latent scores z_{u,i}, which are defined independently of the estimator. Theorem 1 is a plug-in asymptotic equivalence: the latent distance is defined from SVD scores, and the estimator bd^2 - 2*sigma_hat^2 is shown, under an assumed rate for sigma_hat, to be close to it. This is a statistical estimation claim, not a definitional identity. Theorem 2 attempts to prove consistency of the kernel estimator bz_{u,i} to z_{u,i} via a Lipschitz expansion and a kernel-density lemma; Theorem 3 compares the practical estimator to the oracle form. Both targets are external quantities, so the derivation is not circular by construction. There are no self-citations by the authors, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The skeptic's concern about Lemma 3 and Theorem 2 is a mathematical gap: Lemma 2 gives |N| = Theta_p(mn pi^2), while Lemma 3 states |N| = O_p(m^{-1} n^{-1} pi^{-2}), and the stated bandwidths make the leading variance term non-vanishing. That is a proof failure or inconsistency, but it does not reduce the conclusion to the inputs by definition. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central results depend on the SVD low-rank representation, the MCAR missingness model, a rate condition on an unspecified variance estimator, and a delicate bandwidth scaling that, as chosen, does not deliver vanishing variance in the proofs.

free parameters (4)
  • σ̂^2 (variance correction) = Not specified; estimated from residuals of an unspecified first-step estimator
    Used in Theorem 1 and equation (17) to debias squared observed distances; its consistency rate |σ̂ - σ| = O_p(...) is assumed, but the estimator itself is never defined.
  • Bandwidths h1, h2 = Chosen by 5-fold cross-validation over a combinatorial grid
    Theory sets h1 = h2 = O_p(m^{-1/2} n^{-1/2} π^{-1}); practice tunes them; the consistency proof depends critically on this scaling.
  • Overlap threshold β = β = 1 in all experiments
    Controls the minimum number of overlapping ratings to include a user or item in the radial neighbor set; fixed, not tuned.
  • Unspecified 'weight combination' = Cross-validated but not defined in the model
    Section 4 says bandwidths and the weight combination were optimized by cross-validation, but equation (17) contains no explicit weight parameter.
assumptions (6)
  • domain assumption Assumption 1: Missingness completely at random, δ_{u,i} i.i.d. Bernoulli(π)
    Invoked in Propositions 2 and 3, Lemma 2, and all theorems; real recommender data often violate MCAR.
  • domain assumption Assumption 2: Z has low rank k and m/n = O(1)
    The entire SVD latent model and Proposition 1 rest on this.
  • domain assumption Assumption 3: Squared singular values bounded away from zero
    Used to define distances and the Lipschitz constant; fails if the matrix is effectively lower rank.
  • ad hoc to paper Assumption 4: Singular vector entries are O(1/√n) and O(1/√m)
    Dictated uniformity of singular vectors for the Lipschitz bound; not generally true for arbitrary low-rank matrices.
  • domain assumption Existence of a smooth joint density f* for latent feature distances and kernel conditions (symmetry, integrable, finite moments)
    Required by Lemmas 3 and Theorem 2 for kernel smoothing asymptotics, but not listed among the paper's formal assumptions.
  • domain assumption Rate assumption |σ̂ - σ| = O_p(m^{-1/2} n^{-1/2} π^{-1})
    Needed for Proposition 2 and Theorem 1; the paper gives no construction of σ̂ achieving this rate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Radial Neighborhood Smoothing Recommender System." pith.science (2026). https://pith.science/paper/2PZJBVJ2

@misc{pith2026250709952,
  author       = {Pith},
  title        = {Pith review of: Radial Neighborhood Smoothing Recommender System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2PZJBVJ2}},
  note         = {Machine review of arXiv:2507.09952}
}
read the original abstract

Recommender systems inherently exhibit a low-rank structure in latent space. A key challenge is to define meaningful and measurable distances in the latent space to capture user-user, item-item, user-item relationships effectively. In this work, we establish that distances in the latent space can be systematically approximated using row-wise and column-wise distances in the observed matrix, providing a novel perspective on distance estimation. To refine the distance estimation, we introduce the correction based on empirical variance estimator to account for noise-induced non-centrality. The novel distance estimation enables a more structured approach to constructing neighborhoods, leading to the Radial Neighborhood Estimator (RNE), which constructs neighborhoods by including both overlapped and partially overlapped user-item pairs and employs neighborhood smoothing via localized kernel regression to improve imputation accuracy. We provide the theoretical asymptotic analysis for the proposed estimator. We perform evaluations on both simulated and real-world datasets, demonstrating that RNE achieves superior performance compared to existing collaborative filtering and matrix factorization methods. While our primary focus is on distance estimation in latent space, we find that RNE also mitigates the ``cold-start'' problem.

Figures

Figures reproduced from arXiv: 2507.09952 by the authors.

Figure 1
Figure 1. Performance on non-cold-start part. NA% referred to the proportions of NA values in the test part excluding the cold-start part. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Performance on cold-start part for the proposed methods. [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [1]

    Fab: content-based, collaborative recommendation

    Marko Balabanovi \'c and Yoav Shoham. Fab: content-based, collaborative recommendation. Communications of the ACM, 40 0 (3): 0 66--72, 1997

    work page 1997

  2. [2]

    Scalable collaborative filtering with jointly derived neighborhood interpolation weights

    Robert M Bell and Yehuda Koren. Scalable collaborative filtering with jointly derived neighborhood interpolation weights. In Seventh IEEE international conference on data mining (ICDM 2007), pages 43--52. IEEE, 2007

    work page 2007

  3. [3]

    Ellis, Brian Whitman, and Paul Lamere

    Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian Whitman, and Paul Lamere. The million song dataset. In Proceedings of the 12th International Conference on Music Information Retrieval ( ISMIR 2011) , 2011

    2011

  4. [4]

    User modeling for adaptive news access

    Daniel Billsus and Michael J Pazzani. User modeling for adaptive news access. User modeling and user-adapted interaction, 10 0 (2): 0 147--180, 2000

    work page 2000

  5. [5]

    Thy friend is my friend: Iterative collaborative filtering for sparse matrix estimation

    Christian Borgs, Jennifer Chayes, Christina E Lee, and Devavrat Shah. Thy friend is my friend: Iterative collaborative filtering for sparse matrix estimation. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 4718--4729, 2017

    work page 2017

  6. [6]

    A singular value thresholding algorithm for matrix completion

    Jian-Feng Cai, Emmanuel J Cand \`e s, and Zuowei Shen. A singular value thresholding algorithm for matrix completion. SIAM Journal on optimization, 20 0 (4): 0 1956--1982, 2010

    1956

  7. [7]

    Exact matrix completion via convex optimization

    Emmanuel J Cand \`e s and Benjamin Recht. Exact matrix completion via convex optimization. Foundations of Computational mathematics, 9 0 (6): 0 717--772, 2009

    2009

  8. [8]

    The power of convex relaxation: Near-optimal matrix completion

    Emmanuel J Cand \`e s and Terence Tao. The power of convex relaxation: Near-optimal matrix completion. IEEE Transactions on Information Theory, 56 0 (5): 0 2053--2080, 2010

    work page 2010

Show all 31 references
  1. [9]

    Bias and debias in recommender system: A survey and future directions

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. Bias and debias in recommender system: A survey and future directions. arXiv preprint arXiv:2010.03240, 2020

    2010 arXiv

  2. [10]

    Smooth neighborhood recommender systems

    Ben Dai, Junhui Wang, Xiaotong Shen, and Annie Qu. Smooth neighborhood recommender systems. Journal of machine learning research, 20, 2019

    2019

  3. [11]

    A comprehensive survey of neighborhood-based recommendation methods

    Christian Desrosiers and George Karypis. A comprehensive survey of neighborhood-based recommendation methods. Recommender systems handbook, pages 107--144, 2011

    2011

  4. [12]

    Trade-offs between machine learning and deep learning for mental illness detection on social media

    Zhanyi Ding, Zhongyan Wang, Yeyubei Zhang, Yuchen Cao, Yunchong Liu, Xiaorui Shen, Yexin Tian, and Jianglai Dai. Trade-offs between machine learning and deep learning for mental illness detection on social media. Scientific Reports, 15 0 (1): 0 14497, 2025

    2025

  5. [13]

    Using collaborative filtering to weave an information tapestry

    David Goldberg, David Nichols, Brian M Oki, and Douglas Terry. Using collaborative filtering to weave an information tapestry. Communications of the ACM, 35 0 (12): 0 61--70, 1992

    1992

  6. [14]

    Eigentaste: A constant time collaborative filtering algorithm

    Ken Goldberg, Theresa Roeder, Dhruv Gupta, and Chris Perkins. Eigentaste: A constant time collaborative filtering algorithm. information retrieval, 4 0 (2): 0 133--151, 2001

    2001

  7. [15]

    The movielens datasets: History and context

    F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis), 5 0 (4): 0 1--19, 2015

    2015

  8. [16]

    Encoder: Entity mining and modification relation binding for composed image retrieval

    Zixu Li, Zhiwei Chen, Haokun Wen, Zhiheng Fu, Yupeng Hu, and Weili Guan. Encoder: Entity mining and modification relation binding for composed image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 5101--5109, 2025 a

    2025

  9. [17]

    Finecir: Explicit parsing of fine-grained modification semantics for composed image retrieval

    Zixu Li, Zhiheng Fu, Yupeng Hu, Zhiwei Chen, Haokun Wen, and Liqiang Nie. Finecir: Explicit parsing of fine-grained modification semantics for composed image retrieval. arXiv preprint arXiv:2503.21309, 2025 b

    2025 arXiv

  10. [18]

    Greg Linden, Brent Smith, and Jeremy York. Amazon. com recommendations: Item-to-item collaborative filtering. IEEE Internet computing, 7 0 (1): 0 76--80, 2003

    2003

  11. [19]

    Matrix completion with covariate information

    Xiaojun Mao, Song Xi Chen, and Raymond KW Wong. Matrix completion with covariate information. Journal of the American Statistical Association, 114 0 (525): 0 198--210, 2019

    2019

  12. [20]

    Collaborative prediction and ranking with non-random missing data

    Benjamin M Marlin and Richard S Zemel. Collaborative prediction and ranking with non-random missing data. In Proceedings of the third ACM conference on Recommender systems, pages 5--12, 2009

    2009

  13. [21]

    Spectral regularization algorithms for learning large incomplete matrices

    Rahul Mazumder, Trevor Hastie, and Robert Tibshirani. Spectral regularization algorithms for learning large incomplete matrices. The Journal of Machine Learning Research, 11: 0 2287--2322, 2010

    2010

  14. [22]

    Time series modeling for heart rate prediction: From arima to transformers

    Haowei Ni, Shuchen Meng, Xieming Geng, Panfeng Li, Zhuoying Li, Xupeng Chen, Xiaotong Wang, and Shiyao Zhang. Time series modeling for heart rate prediction: From arima to transformers. arXiv preprint arXiv:2406.12199, 2024. URL http://arxiv.org/abs/2406.12199

    2024 arXiv

  15. [23]

    A comprehensive survey of neighborhood-based recommendation methods

    Xia Ning, Christian Desrosiers, and George Karypis. A comprehensive survey of neighborhood-based recommendation methods. Recommender systems handbook, pages 37--76, 2015

    2015

  16. [24]

    Fast maximum margin matrix factorization for collaborative prediction

    Jasson DM Rennie and Nathan Srebro. Fast maximum margin matrix factorization for collaborative prediction. In Proceedings of the 22nd international conference on Machine learning, pages 713--719, 2005

    2005

  17. [25]

    Blind regression: Nonparametric regression for latent variable models via collaborative filtering

    Dogyoon Song, Christina E Lee, Yihua Li, and Devavrat Shah. Blind regression: Nonparametric regression for latent variable models via collaborative filtering. Advances in Neural Information Processing Systems, 29: 0 2155--2163, 2016

    2016

  18. [26]

    Unifying user-based and item-based collaborative filtering approaches by similarity fusion

    Jun Wang, Arjen P De Vries, and Marcel JT Reinders. Unifying user-based and item-based collaborative filtering approaches by similarity fusion. In Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 501-...

    2006

  19. [27]

    Enhancing user intent for recommendation systems via large language models, 2025

    Xiaochuan Xu, Zeqiu Xu, Peiyang Yu, and Jiani Wang. Enhancing user intent for recommendation systems via large language models, 2025. URL https://arxiv.org/abs/2501.10871

    2025 arXiv

  20. [28]

    Measuring digitalization capabilities using machine learning

    Jinglan Yang, Jianghuai Liu, Zheng Yao, and Chaoqun Ma. Measuring digitalization capabilities using machine learning. Research in International Business and Finance, 70: 0 102380, 2024. ISSN 0275-5319. doi:https://doi.org/10.1016/j.ribaf.2024.102380. URL https://www.sciencedir...

    2024

  21. [29]

    Leveraging missing ratings to improve online recommendation systems

    Yuanping Ying, Fred Feinberg, and Michel Wedel. Leveraging missing ratings to improve online recommendation systems. Journal of marketing research, 43 0 (3): 0 355--365, 2006

    2006

  22. [30]

    Nonparametric matrix estimation with one-sided covariates

    Christina Lee Yu. Nonparametric matrix estimation with one-sided covariates. In 2022 IEEE International Symposium on Information Theory (ISIT), pages 892--897. IEEE, 2022

    2022

  23. [31]

    A narrative review on large ai models in lung cancer screening, diagnosis, and treatment planning

    Jiachen Zhong, Yiting Wang, Di Zhu, and Ziwei Wang. A narrative review on large ai models in lung cancer screening, diagnosis, and treatment planning. arXiv preprint arXiv:2506.07236, 2025

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.