Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Criteria-Aware Graph Filtering: Extremely Fast Yet Accurate Multi-Criteria Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A training-free graph filtering method can make accurate multi-criteria recommendations in under 0.2 seconds, outperforming deep-learning baselines.

desk verdict A clean, reproducible training-free MC recommender with a legitimate speed idea, but the headline 0.2s runtime excludes the validation-time filter search and the headline gain is measured against a self-built baseline. read the letter →

arxiv 2502.09046 v1 pith:H3KGYJYJ submitted 2025-02-13 cs.IR cs.AIcs.ITcs.LGcs.SImath.IT

Jin-Duk Park , Jaemin Yoo , Won-Yong Shin This is my paper · ORCID
classification cs.IRcs.AIcs.ITcs.LGcs.SImath.IT
keywords multi-criteriarecommendersystemgraphfilteringpolynomiallow-passfiltertraining-freerecommendationcollaborativecriteriapreferencematrixdecomposition-free
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

This paper introduces CA-GF, a multi-criteria (MC) recommender that requires no training and no matrix decomposition. Its central claim is that by building a single item-item similarity graph from all criteria at once, smoothing each criterion's ratings with its own chosen low-pass polynomial filter, and weighting the smoothed signals by each user's revealed criteria preferences, one can beat all benchmark MC recommenders on accuracy while running in under 0.2 seconds even on the largest dataset tested. If true, this shows that deep learning is not required for accurate multi-criteria recommendation, and that a closed-form, interpretable graph filtering pipeline can be both faster and more accurate. The gains are reported as up to 24% in NDCG@5 over the best competitor.

What carries the argument

The machinery is threefold: (1) an MC user-expansion graph that concatenates the overall and criterion rating matrices into $R_{MC}$ and forms the normalized item-item similarity graph $\tilde{P}_{MC}$; (2) criterion-specific polynomial graph filtering, where each criterion $c$ is convolved with one of three predefined second-order polynomial low-pass filters — linear ($\tilde{P}$), inward ($\tilde{P}^2$), outward ($2\tilde{P} - \tilde{P}^2$) — chosen per criterion on the validation set, with edge weights adjusted via a Hadamard power $s_f$ to control over-smoothing; and (3) criteria preference-infused aggregation, which builds a criterion-criterion similarity graph $T$ from per-criterion sum ratings and uses it to weight each smoothed signal before summing. The polynomial filters avoid eigen-decomposition, so the whole pipeline is closed-form and trivially parallelizable.

What would settle it

Take any MC dataset and run CA-GF alongside a version that searches over a wider family of polynomial low-pass filters (e.g., degree up to 5 with coefficient search on the validation set). If the wider search yields materially higher NDCG@10 than CA-GF's three-filter choice, then the three-filter family does not contain the optimal filters and the central accuracy claim is falsified.

Watch

Extended reading notes

Core claim

The paper's core discovery is that criteria-aware graph filtering yields state-of-the-art multi-criteria recommendation without any learned parameters. CA-GF expands each user into C+1 criterion-user nodes, forms an item-item similarity graph from the concatenated rating matrix, and applies a distinct second-order polynomial low-pass filter — linear ($h(\lambda)=1-\lambda$), inward ($h(\lambda)=(1-\lambda)^2$), or outward ($h(\lambda)=1-\lambda^2$) — to each criterion's rating signal, selecting the best filter per criterion and per dataset on a validation set. The filtered signals are aggregated with per-user weights derived from a criterion-criterion similarity graph, giving an interpretable decomposition of each prediction into criterion contributions. The authors show consistent accuracy gains over DNN-based, matrix-factorization, and graph-convolution MC baselines on TripAdvisor, Yahoo!Movie, and BeerAdvocate, with runtime of 0.2 seconds on the largest dataset.

Load-bearing premise

The central bet is that for every criterion, the best low-pass filter is one of just three fixed second-order polynomial shapes (linear, inward, outward) with a single strength knob per shape; if the true optimal filter for some criterion has a different frequency response, the reported accuracy gains could shrink or vanish.

Editorial extensions

If this is right

  • If CA-GF's claims hold, multi-criteria recommendation can be delivered as a near-instant service that updates when new ratings arrive, since there is no training phase to re-run.
  • The per-criterion filter choice provides a built-in diagnostic: which criteria behave like low-frequency signals and which need less smoothing can be read off the selected filters, aiding dataset understanding.
  • The accuracy gains imply that explicit modeling of inter-criteria relations through a single shared item graph is more effective than treating each criterion independently, a design principle for future MC recommenders.
  • Because CA-GF is matrix decomposition-free, it can scale to graphs that exceed GPU memory by falling back to CPU with graceful degradation, as shown on synthetic data up to 38M ratings.

Reading between the lines

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

  • The speed advantage would likely survive swapping the three predefined filters for higher-order polynomial families, since the bottleneck-avoiding step is the polynomial form itself, not the specific degree.
  • The paper's preference-weighting claim — that users who rate a criterion more often or more highly prefer it — is testable against explicit preference surveys or held-out criteria ratings; a weak correlation would require a different aggregation source.
  • The per-criterion attribution maps could be used as a debugging tool in production recommenders to locate which aspect of an item drives a prediction, a use the paper does not develop.
  • Whether the sub-second runtime survives web-scale data depends on the sparsity of the item-item similarity graph; the paper's linear-in-nonzeros complexity suggests it may, but this is not demonstrated at that scale.
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

3 major / 5 minor

Summary. The paper proposes CA-GF (Criteria-Aware Graph Filtering), a non-parametric, matrix-decomposition-free method for top-K multi-criteria (MC) recommendation. CA-GF concatenates the per-criterion rating matrices into an MC user-expansion matrix, normalizes it, and builds an item–item similarity graph whose edge weights are raised to a per-filter power s_f. For each criterion, one of three second-order polynomial filters (linear, inward, outward) is selected on the validation set; the smoothed criterion signals are then aggregated with weights derived from a smoothed criterion–criterion similarity graph. On TripAdvisor, Yahoo!Movie, and BeerAdvocate, the paper reports that CA-GF outperforms six MC baselines and four single-criterion baselines, runs in 0.2 seconds on BeerAdvocate, scales to 38M synthetic ratings, and provides attribution maps for interpretability. The frequency-response derivations (Corollaries 4.1–4.3) are correct, the algorithmic pipeline is clearly specified, and the code is publicly released.

Significance. CA-GF is a well-specified and easily reproducible extension of the polynomial graph-filtering line (GF-CF, Turbo-CF) to the multi-criteria recommendation setting, and the empirical study is thorough: three benchmark datasets, ten competitors, component-wise ablations, a synthetic scalability study up to 38M ratings, and an interpretability analysis. The theoretical derivations are sound, and the ablation design (Table 4) cleanly isolates the contribution of each module. The central caveats concern the framing of the headline claims rather than the method itself: the sub-0.2-second figure is the cost of a single forward pass and excludes the validation-based model selection that produces the reported accuracy numbers; several accuracy margins over the strongest baselines are at the one-percent level and are presented without error bars or significance tests; and the 'optimal filter' language overstates a search over three fixed filters. If the authors report the end-to-end runtime and add statistical support for the close comparisons, the contribution would be solid.

major comments (3)
  1. [§4.2, Table 2; §3.3.2, Appendix E.3] The headline efficiency figure is the cost of a single forward pass, not the end-to-end cost of producing the reported predictions. Table 2 bills CA-GF at 0.2 seconds with a 'Training: ✗' tag, and the Abstract, Section 1, and Section 4.2 repeat the sub-0.2-second claim, but Algorithm 1 takes the per-criterion filter choices and the adjustment parameters s_f and s as given inputs. Those values are obtained by a validation-based search over three filter types per criterion (Figure 4), over grids of s_f for each filter type (Figure 5), and over s for the criterion–criterion graph (Appendix E.3); the paper itself acknowledges in Section 3.3.2 that the optimal filter 'is easily found using the validation set.' The search time is not reported or bounded, while every training-based baseline in Table 2 is billed with its full 100-epoch training cost. The runtime comparison is therefore not apples-to-apples, and the 'training-free / extremely fast' claim is incomplete as stated. Please report the total wall-clock time from raw MC ratings to final test predictions (including the validation search) or explicitly label the 0.2-second figure as inference-only, and specify the search protocol used to obtain the values in Appendix E.3.
  2. [§4.3, Table 3] The accuracy comparisons lack statistical support. Table 3 reports single values for the deterministic methods (GF-CF, GF-CFMC, CA-GF) and ten-run averages without spread for the stochastic baselines; no standard deviations or significance tests are reported. Because CA-GF's numbers depend on the random 80/20 data split and on validation-based selection, a single run does not support the statement in Section 4.3(i) that CA-GF 'consistently outperforms all the competitors.' The margins on Yahoo!Movie are at the noise level (Recall@5: 0.1224 vs. 0.1223 for GF-CFMC; NDCG@10: 0.1608 vs. 0.1588), and the 'up to 24%' headline refers to the single most favorable cell (TA NDCG@5, 0.0738 vs. 0.0595) measured against GF-CFMC, a baseline introduced in this paper. Please report results over multiple splits with standard deviations and run significance tests (e.g., paired bootstrap or a paired test over splits), and state explicitly against which competitor the 24% figure is computed.
  3. [§3.3.2, §4.5, Appendix E.3] The paper repeatedly refers to 'the optimal filter for each criterion' (Sections 3.3.1–3.3.2, 4.5, and the Conclusion), but the search space is a fixed set of three second-order polynomials. The assertion in Section 3.3.2 that the linear, inward, and outward filters 'essentially embrace a broad set of LPFs' is not substantiated, and the family h(λ) ∈ {1−λ, (1−λ)², 1−λ²} is a small subset of the polynomial-LPF space. If the criterion-specific spectra are not well approximated by this family, the reported gains are conditional on these hand-picked filters rather than on the principle of criterion-aware filtering itself. A concrete test would be to compare against general second-order (or higher-order) polynomial filters with coefficients a_{c,k} in Eq. (12) tuned on the validation set, which would also indicate whether the three-option search overfits the validation data. At minimum, the 'optimal' wording should be qualified as 'best among the three considered filters.'
minor comments (5)
  1. [§1, §2.2, §4.1, §4.6, §6, Appendix C] Several typos remain: 'device' for 'devise' (§1); 'ExtandedSAE' for 'ExtendedSAE' (§4.1 and Table 3); 'relavent' for 'relevant' (§4.6); 'CF-GF' for 'CA-GF' (§6); 'Collorary' for 'Corollary' (Appendix C); and an unbalanced brace in Definition 2 ('h(λ_1),...,h(λ_V}').
  2. [Tables 3, 4, 6, and 7] The BeerAdvocate Recall@10 for CA-GF is reported as 0.1144 in Tables 3 and 7 but as 0.1147 in Tables 4 and 6; please harmonize the value.
  3. [§4.2, Table 2] The claim that CA-GF is 'over 2,160× faster than GF-CFMC on the BA dataset' is inconsistent with the numbers in Table 2, where 274 s / 0.2 s = 1,370×; please correct the factor or the runtime values and re-verify the other speed-up ratios.
  4. [Appendix C, Corollaries 4.1–4.3; Eq. (10)] The proofs are algebraic identities, but the interpretation of the adjusted matrix P̄_f as a low-pass graph filter requires spectral conditions. With non-integer exponents s_f < 1 used in Appendix E.3 (e.g., 0.1 on TA), the Hadamard power in Eq. (10) need not be positive semidefinite, so the frequency response may take negative values; please state the conditions under which P̄_f is a valid LPF or explicitly qualify the 'low-pass' interpretation.
  5. [§4.6, Figure 6] The interpretability claim that 'check-in and rooms contribute less' is based on only two attribution maps; consider reporting an aggregate attribution (e.g., averaged over test interactions) to support the qualitative observation.

Circularity Check

0 steps flagged · score 1.0 of 10

No definitional circularity; validation-tuned per-criterion filters are standard model selection, not fitted predictions, and self-citations are not load-bearing.

full rationale

CA-GF's scoring rule (Eq. 17) is an explicit linear combination of smoothed input signals: s_u = (1/(C+1)) * sum_c C_hat_{u,c} r_{u,c} f(P_bar_{f,c}). The quantities that are selected on data, namely the per-criterion polynomial filter type f, the graph-adjustment parameter s_f, and the criterion-graph parameter s, are all tuned on the validation split (Section 4.1: 'we randomly select 10% of interactions as the validation set for hyperparameter tuning'; Appendix E.3 lists the chosen defaults), while accuracy is reported on the held-out 20% test split. The reported test metrics are therefore not the same quantity that was maximized during selection, so the 'fitted input called prediction' pattern does not apply. The language 'optimal filter for each criterion' (Section 3.3.2, Figure 4) overstates a search over three predefined polynomial LPFs, but this is a scope-of-model-selection concern, not a definitional equivalence. The <0.2s runtime in Table 2 measures a single forward pass of the final CA-GF configuration and does not include the validation search; that affects the fairness of the 'training-free efficiency' narrative but does not make the runtime a circularly constructed output. The self-citations to Turbo-CF [26] and CPA-LGC [25] supply polynomial graph filtering and the MC user-expansion graph as building blocks, yet the paper re-derives the frequency responses in Corollaries 4.1-4.3 and evaluates against external baselines on public datasets, so the central argument does not reduce to an unverified self-citation chain. No circular step can be exhibited from the paper's own equations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's accuracy depends on three groups of fitted values: the per-criterion filter selection, the edge-weight adjustment for the item graph, and the adjustment for the criterion graph, all tuned on the validation set. The underlying axioms are standard graph-signal assumptions plus the paper-specific assumption that the three-filter family is sufficient.

free parameters (3)
  • Per-criterion polynomial LPF choice f(P_f,c) = L/I/O per criterion; see Appendix E.3 (e.g., TA overall=O, service=L)
    Selected from {L, I, O} by validation performance for each criterion and dataset; the reported accuracy is measured after this selection.
  • Adjustment parameter s_f for the item-item graph = TA: {L: 0.1, I: 1, O: 1.2}; YM: {L: 1, I: 1, O: 1.8}; BA: {L: 0.6, I: 0.85, O: 1.5}
    Controls the Hadamard power of edge weights per filter type, tuned on the validation set (Eq. 10, Appendix E.3).
  • Adjustment parameter s_f for the criterion-criterion similarity graph = TA: 2, YM: 4, BA: 2
    Controls the Hadamard power in Eq. (16) for the criterion preference aggregation, tuned on the validation set.
assumptions (4)
  • domain assumption User preferences are low-frequency signals on the item-item similarity graph, so low-pass filtering denoises ratings and improves predictions.
    Inherited from single-criterion graph-filtering CF (GF-CF, [32]); the paper applies it to the MC user-expansion graph without new validation.
  • ad hoc to paper The three second-order polynomial filters L, I, O 'essentially embrace a broad set of LPFs', so per-criterion optimal filters can be found within this family.
    Stated in Section 3.3.2 without proof; the paper only searches this family on validation.
  • domain assumption If a user gives more or higher ratings on a criterion, they prefer that criterion more, so sum-rating per criterion can weight aggregation.
    Section 3.3.3; the ablation CA-GF-p shows only marginal gains, so this assumption is weakly supported.
  • domain assumption Polynomial graph filters with K=2 avoid matrix decomposition and yield linear complexity O(n_P); n_P is 'not significantly large' compared to n_mc.
    Appendix D; the claim about n_P is asserted, not proven, and depends on dataset sparsity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Criteria-Aware Graph Filtering: Extremely Fast Yet Accurate Multi-Criteria Recommendation." pith.science (2026). https://pith.science/paper/H3KGYJYJ

@misc{pith2026250209046,
  author       = {Pith},
  title        = {Pith review of: Criteria-Aware Graph Filtering: Extremely Fast Yet Accurate Multi-Criteria Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H3KGYJYJ}},
  note         = {Machine review of arXiv:2502.09046}
}
read the original abstract

Multi-criteria (MC) recommender systems, which utilize MC rating information for recommendation, are increasingly widespread in various e-commerce domains. However, the MC recommendation using training-based collaborative filtering, requiring consideration of multiple ratings compared to single-criterion counterparts, often poses practical challenges in achieving state-of-the-art performance along with scalable model training. To solve this problem, we propose CA-GF, a training-free MC recommendation method, which is built upon criteria-aware graph filtering for efficient yet accurate MC recommendations. Specifically, first, we construct an item-item similarity graph using an MC user-expansion graph. Next, we design CA-GF composed of the following key components, including 1) criterion-specific graph filtering where the optimal filter for each criterion is found using various types of polynomial low-pass filters and 2) criteria preference-infused aggregation where the smoothed signals from each criterion are aggregated. We demonstrate that CA-GF is (a) efficient: providing the computational efficiency, offering the extremely fast runtime of less than 0.2 seconds even on the largest benchmark dataset, (b) accurate: outperforming benchmark MC recommendation methods, achieving substantial accuracy gains up to 24% compared to the best competitor, and (c) interpretable: providing interpretations for the contribution of each criterion to the model prediction based on visualizations.

Figures

Figures reproduced from arXiv: 2502.09046 by the authors.

Figure 1
Figure 1. An illustration showing (a) four criteria ratings in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The schematic overview of CA-GF. according to the degree of nodes in the graph as in [19, 32, 43]: 𝑅˜𝑀𝐶 = 𝐷 −1/2 𝑈 𝑅𝑀𝐶𝐷 −1/2 𝐼 , (8) where 𝐷𝑈 and 𝐷𝐼 are the diagonal matrices of criterion-user nodes and item nodes, respectively, defined as 𝐷𝑈 = diag(𝑅𝑀𝐶1) and 𝐷𝐼 = diag(1 𝑇 𝑅𝑀𝐶). 3.2.2 Item–item similarity graph with adjustment. To perform graph filtering, we construct the normalized item–item similarity graph 𝑃˜MC ∈… view at source ↗
Figure 3
Figure 3. Log-scaled runtime comparison of CA-GF (with and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The effect of three polynomial LPFs 𝑓 (𝑃¯ 𝑓 , 𝑐) (L, I, and O) for each criterion on the Recall@10. (i) All four modules in CA-GF plays a crucial role in the success of the proposed CA-GF method. (ii) The performance gap between CA-GF and CA-GF-f reveals that using div…
Figure 5
Figure 5. Figure 5: The effect of adjustment parameter 𝑠𝑓 for three polynomial LPFs (L, I, and O) on the Recall@10. (a) (user 1, item 7) (b) (user 1844, item 1) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Attribution maps that visualize the contribution of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: The schematic overview of GF-CFMC. (a) (user 1, item 125) and (user 165, item 1855) on YM (b) (user 10, item 1544) and (user 4895, item 1382) on BA [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Attribution maps that visualize the contribution of [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages

  1. [1]

    Jianxin Chang, Chen Gao, Yu Zheng, Yiqun Hui, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. 2021. Sequential recommendation with graph neural networks. In SIGIR. 378–387

    work page 2021

  2. [2]

    Zhengyu Chen, Ziqing Xu, and Donglin Wang. 2021. Deep transfer tensor decomposition with orthogonal constraint for recommender systems. In AAAI. 4010–4018

    work page 2021

  3. [3]

    Jeongwhan Choi, Seoyoung Hong, Noseong Park, and Sung-Bae Cho. 2023. Blurring-sharpening process models for collaborative filtering. In SIGIR. 1096– 1106

    2023

  4. [4]

    Ge Fan, Chaoyun Zhang, Junyang Chen, Paul Li, Yingjie Li, and Victor CM Leung

    work page

  5. [5]

    Ge Fan, Chaoyun Zhang, Junyang Chen, and Kaishun Wu. 2021. Predicting ratings in multi-criteria recommender systems via a collective factor model. In DeMal@ The Web Conference. 1–6

    work page 2021

  6. [6]

    Ruining He and Julian McAuley. 2016. VBPR: Visual Bayesian personalized ranking from implicit feedback. In AAAI

    work page 2016

  7. [7]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and powering graph convolution network for recommendation. In SIGIR. 639–648

    work page 2020

  8. [8]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In WWW. 173–182

    2017

Show all 51 references
  1. [9]

    Yu Hou, Jin-Duk Park, and Won-Yong Shin. 2024. Collaborative filtering based on diffusion models: Unveiling the potential of high-order connectivity. In SIGIR. 1360–1369

    2024

  2. [10]

    Dietmar Jannach, Zeynep Karakaya, and Fatih Gedikli. 2012. Accuracy improve- ments for multi-criteria recommender systems. In EC. 674–689

    2012

  3. [11]

    Dietmar Jannach, Markus Zanker, and Matthias Fuchs. 2014. Leveraging multi- criteria customer feedback for satisfaction analysis and improved recommenda- tions. Information Technology & Tourism (2014), 119–149

    2014

  4. [12]

    Michel Journée, Yurii Nesterov, Peter Richtárik, and Rodolphe Sepulchre. 2010. Generalized power method for sparse principal component analysis. Journal of Machine Learning Research 11, 2 (2010)

    2010

  5. [13]

    Bin Ju, Yuntao Qian, Minchao Ye, Rong Ni, and Chenxi Zhu. 2015. Using dynamic multi-task non-negative matrix factorization to detect the evolution of user preferences in collaborative filtering. PLoS One 10, 8 (2015), e0135090

    2015

  6. [14]

    Chaehyun Kim, Yoonryung Choi, Jin-Duk Park, and Won-Yong Shin. 2025. Lever- aging member-group relations via multi-view graph filtering for effective group recommendation. In WWW. (to appear)

    2025

  7. [15]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic opti- mization. In ICLR. 1–15

    2015

  8. [16]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In ICLR. 1–14

    2017

  9. [17]

    Pan Li and Alexander Tuzhilin. 2019. Latent multi-criteria ratings for recommen- dations. In RecSys. 428–431

    2019

  10. [18]

    Pan Li and Alexander Tuzhilin. 2020. Learning latent multi-criteria ratings from user reviews for recommendations. IEEE Transactions on Knowledge and Data Engineering 34, 8 (2020), 3854 – 3866

    2020

  11. [19]

    Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. 2023. Personalized graph signal processing for collaborative filtering. In WWW. 1264–1272

    2023

  12. [20]

    Julian McAuley, Jure Leskovec, and Dan Jurafsky. 2012. Learning attitudes and attributes from multi-aspect reviews. In ICDM. 1020–1025

    2012

  13. [21]

    Julian John McAuley and Jure Leskovec. 2013. From amateurs to connoisseurs: Modeling the evolution of user expertise through online reviews. In WWW. 897–908

    2013

  14. [22]

    Nour Nassar, Assef Jafar, and Yasser Rahhal. 2020. A novel deep multi-criteria collaborative filtering model for recommendation system. Knowledge-Based Systems (2020), 104811

    2020

  15. [23]

    Antonio Ortega, Pascal Frossard, Jelena Kovačević, José MF Moura, and Pierre Vandergheynst. 2018. Graph signal processing: Overview, challenges, and appli- cations. Proc. IEEE 106, 5 (2018), 808–828

    2018

  16. [24]

    Jin-Duk Park, Kyung-Min Kim, and Won-Yong Shin. 2024. CF-KAN: Kolmogorov- Arnold network-based collaborative filtering to mitigate catastrophic forgetting in recommender systems. arXiv preprint arXiv:2409.05878 (2024)

    2024 arXiv

  17. [25]

    Jin-Duk Park, Siqing Li, Xin Cao, and Won-Yong Shin. 2023. Criteria tell you more than ratings: Criteria preference-aware light graph convolution for effective multi-criteria recommendation. In KDD. 1356–1365

    2023

  18. [26]

    Jin-Duk Park, Yong-Min Shin, and Won-Yong Shin. 2024. Turbo-CF: Matrix decomposition-free graph filtering for fast recommendation. In SIGIR. 2672– 2676

    2024

  19. [27]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. PyTorch: An imperative style, high-performance deep learning library. (2019)

    2019

  20. [28]

    Fabíola SF Pereira, João Gama, Sandra de Amo, and Gina MB Oliveira. 2018. On analyzing user preference dynamics with temporal social networks. Machine Learning 107 (2018), 1745–1773

    2018

  21. [29]

    Raksha Ramakrishna, Hoi-To Wai, and Anna Scaglione. 2020. A user guide to low-pass graph signal processing and its applications: Tools and applications. IEEE Signal Processing Magazine 37, 6 (2020), 74–85

    2020

  22. [30]

    Jason Sanders and Edward Kandrot. 2010. CUDA by example: an introduction to general-purpose GPU programming . Addison-Wesley Professional

    2010

  23. [31]

    Qusai Shambour. 2021. A deep learning based algorithm for multi-criteria rec- ommender systems. Knowledge-Based Systems (2021), 106545

    2021

  24. [32]

    Yifei Shen, Yongji Wu, Yao Zhang, Caihua Shan, Jun Zhang, B Khaled Letaief, and Dongsheng Li. 2021. How powerful is graph convolution for recommendation?. In CIKM. 1619–1629

    2021

  25. [33]

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. 2013. The emerging field of signal processing on graphs: Ex- tending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine 30, 3 (2013), 83–98

    2013

  26. [34]

    Rama Syamala Sreepada, Bidyut Kr Patra, and Antonio Hernando. 2017. Multi- criteria recommendations through preference learning. In Proceedings of the 4th ACM IKDD Conferences on Data Sciences . 1–11

    2017

  27. [35]

    Dharahas Tallapally, Rama Syamala Sreepada, Bidyut Kr Patra, and Korra Sathya Babu. 2018. User preference learning in multi-criteria recommendations using stacked auto encoders. In RecSys. 475–479

    2018

  28. [36]

    Kipf, and Max Welling

    Rianne van den Berg, Thomas N. Kipf, and Max Welling. 2017. Graph convolu- tional matrix completion. CoRR abs/1706.02263 (2017)

    2017 arXiv

  29. [37]

    Hoi-To Wai, Santiago Segarra, Asuman E Ozdaglar, Anna Scaglione, and Ali Jadbabaie. 2019. Blind community detection from low-rank excitations of a graph filter. IEEE Transactions on signal processing 68 (2019), 436–451

    2019

  30. [38]

    Hongning Wang, Yue Lu, and ChengXiang Zhai. 2011. Latent aspect rating analysis without aspect keyword supervision. In KDD. 618–626

    2011

  31. [39]

    Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua

  32. [40]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In SIGIR. 165–174

    2019

  33. [41]

    In SIGIR

    Diffusion recommender model. In SIGIR. 832–841

  34. [42]

    Di Wu, Xin Luo, Mingsheng Shang, Yi He, Guoyin Wang, and MengChu Zhou

  35. [43]

    Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, and Tat-Seng Chua

  36. [44]

    Leonid Yavits, Amir Morad, and Ran Ginosar. 2014. Sparse matrix multiplication on an associative processor. IEEE Transactions on Parallel and Distributed Systems 26, 11 (2014), 3175–3183

    2014

  37. [45]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In KDD. 974–983

    2018

  38. [46]

    Lei Zheng, Chun-Ta Lu, Fei Jiang, Jiawei Zhang, and Philip S Yu. 2018. Spectral collaborative filtering. In RecSys. 311–319

    2018

  39. [47]

    Jiafeng Xia, Dongsheng Li, Hansu Gu, Jiahao Liu, Tun Lu, and Ning Gu. 2022. FIRE: Fast incremental recommendation with graph signal processing. In WWW. 1808—-1819

    2022

  40. [51]

    Yong Zheng. 2019. Utility-based multi-criteria recommender systems. In SAC. 2529–2531. WWW ’25, April 28-May 2, 2025, Sydney, NSW, Australia Jin-Duk Park, Jaemin Yoo, and Won-Yong Shin A Definition of LPF We formally define the LPF as follows: Definition 5. (LPF) [29, 32, 37]:...

    2019

  41. [2019]

    IEEE Transactions on Systems, Man, and Cybernetics: Systems 51, 7 (2019), 4285–4296

    A deep latent factor model for high-dimensional and sparse matrices in recommender systems. IEEE Transactions on Systems, Man, and Cybernetics: Systems 51, 7 (2019), 4285–4296

    2019

  42. [2020]

    In SIGIR

    Disentangled graph collaborative filtering. In SIGIR. 1001–1010

  43. [2023]

    IEEE Transactions on Network Science and Engineering 10, 6 (2023), 3633 – 3643

    Improving rating prediction in multi-criteria recommender systems Via a collective factor Model. IEEE Transactions on Network Science and Engineering 10, 6 (2023), 3633 – 3643

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.