GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering

Ahmad Bin Rabiah; Julian McAuley

REVIEW 4 major objections 5 minor 12 references

GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read GSPRec claims that adding order-derived item-item edges to the collaborative-filtering graph exposes mid-frequency preference signals that a Gaussian bandpass filter amplifies, improving top-k recommendation accuracy.

desk verdict A genuinely new graph construction and an honest ablation story, but the core bandpass equation does not type-check and the protocol has enough inconsistencies that the headline gains are not yet established. read the letter →

arxiv 2505.11552 v3 pith:3U2EE5XO submitted 2025-05-15 cs.IR cs.AI

Ahmad Bin Rabiah , Julian McAuley This is my paper · ORCID

classification cs.IRcs.AI

keywords graphsignalprocessingcollaborativefilteringspectralbandpassfiltermulti-hopdiffusionsequentialrecommendationLaplaciantop-k

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

GSPRec is a graph-signal-processing recommender that claims to improve item representations by folding user interaction order into the graph before any spectral filtering happens. The central claim is that symmetrized item-item transition edges, spread by multi-hop diffusion with exponential decay, expose intermediate-frequency structure in the Laplacian spectrum, and that a Gaussian bandpass filter aimed at those mid-frequencies, combined with a low-pass channel for popularity, beats low-pass-only graph filters and message-passing GCNs on top-k recommendation. The paper reports consistent gains over all compared baselines on ML100K, ML1M, Netflix, and Amazon Beauty, with average NDCG@10 improvements stated as 5.12% in the abstract and 6.77% in the introduction and results. If the claim is right, sequential structure can be captured by graph topology itself, without a separate sequence model, and frequency-selective filtering is the main performance lever. The mechanism claim is that mid-frequency spectral coefficients carry the user-specific preference pattern, a claim the paper asserts with an analysis in Appendix G.

What carries the argument

The central object is the unified symmetric normalized Laplacian $L=I-D^{-1/2}AD^{-1/2}$ over the adjacency $A$ of Eq. 5, whose item-item block is the symmetrized, exponentially decayed multi-hop diffusion $S^{(d)}=\sum_{k=1}^{d}\alpha^{k-1}(S')^k$ of the symmetrized transition matrix $S'$. Symmetry of $A$ guarantees real eigenvalues and orthogonal eigenvectors, which is why the paper symmetrizes directed transitions despite losing arrow information. The Gaussian bandpass $g_{BP}(\lambda)$ selects a mid-frequency window via center $c$ and width $w$, the low-pass channel $F_{LP}$ captures popularity, and the convex fusion $\phi F_{BP}+(1-\phi)F_{LP}$ gives the final item scores. A truncated eigendecomposition with $r\ll m+n$ eigenpairs makes the method much cheaper than training GCN baselines, and the diffusion converges when $\alpha<1/\rho(S')$ by the matrix geometric series.

What would settle it

Re-run every baseline with the paper's own code and evaluation on the exact 8:1:1 split, and re-run GSPRec with the unified graph built strictly from the training split; if the NDCG@10 advantage over the strongest GSP baseline collapses or reverses, the headline claim fails. Separately, plot the spectral coefficients of user-specific versus popularity signals across all four datasets; if mid-frequency coefficients in the range $0.3<\lambda_i<0.8$ do not show higher magnitude for user-specific patterns, the stated mechanism is unsupported.

Watch

Extended reading notes

Core claim

On the paper's own terms, GSPRec consistently outperforms all baselines across metrics and datasets, and the mechanism is spectral: the unified adjacency $A=\begin{pmatrix}0 & X\\ X^\top & \tilde S\end{pmatrix}$ combines the user-item interaction matrix $X$ with a symmetrically normalized item-item block $\tilde S$ built from $d$-hop diffusion of the symmetrized direct-transition matrix, so the normalized Laplacian $L=I-D^{-1/2}AD^{-1/2}$ gains intermediate-frequency structure. A Gaussian bandpass $g_{BP}(\lambda)=\exp(-(\bar\lambda-c)^2/w)$ amplifies that band, a low-pass component retains global trends, and the final scores are $Y=\phi F_{BP}+(1-\phi)F_{LP}$ with $\phi\in[0,1]$. On ML1M the ablations show that removing the bandpass filter causes the largest drop, removing the low-pass component the next largest, and removing sequential diffusion the smallest, which the paper reads as evidence that graph construction and filter design are coupled rather than independent.

Load-bearing premise

The reported gains stand on the untested premise that the baseline numbers were produced under exactly the same data split and preprocessing as GSPRec's, and that the item-transition graph is built only from training interactions rather than from the test window.

Editorial extensions

If this is right

If GSPRec's results hold, adding symmetrized multi-hop item-transition edges to the graph is sufficient to inject sequential signal into a spectral recommender, so a separate sequential model is not needed for these gains.
Tuning the bandpass center and width per dataset becomes a first-class design choice: datasets with different spectral signatures require different mid-frequency windows, and no single fixed filter will dominate.
The model's linear, non-learned filtering means the accuracy gains come without embedding training, so the approach sits on a different accuracy-efficiency curve than message-passing GCNs.
Because ablations show the bandpass filter is the largest single component, future graph-CF work should treat graph construction and filter design as a joint decision rather than independent modules.

Reading between the lines

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

Beyond the paper, the reported per-dataset gains suggest a testable prediction the paper does not run: datasets with stronger community structure should show larger bandpass benefits, so one could rank the four benchmarks by a spectral community measure and check it against the NDCG@10 improvements.
Beyond the paper, if the protocol check passes, the design implies that expensive sequence encoders (RNNs or transformers) may be replaceable by a fixed multi-hop graph operator plus a tuned spectral window in settings where only ranking quality matters.
Beyond the paper, the explicit discarding of directionality opens a natural extension the authors only nod to: a directed-Laplacian version could preserve arrow information and might improve on the symmetric construction, at the cost of a different spectral theory.
Beyond the paper, because the filter is a linear operator on a fixed graph, the same construction could be coupled with polynomial filter approximations or data-driven band learning rather than the fixed Gaussian kernel, which the paper mentions only as a scalability improvement.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: the paper's core construction—a symmetrized, multi-hop-diffused item-transition graph fused with the user-item bipartite graph into one Laplacian, then filtered with a Gaussian bandpass plus low-pass pair—is genuinely new and worth engaging. The ablations are candid and useful. But the bandpass formula in Eq. (8) and Algorithm 1 is dimensionally invalid as written, and the experimental protocol has enough loose ends that the headline numbers are not verifiable in this version.

What's good: encoding sequence order into a symmetric item graph before spectral filtering is a reasonable way to inject temporal information without directed Laplacians. The diffusion convergence lemma is a standard geometric-series argument but correct. The ablation hierarchy (bandpass removal hurts most, sequential graph removal least) is the most informative evidence in the paper and is presented honestly. The writing is clear and the positioning against prior GSP-CF work is solid.

Where it falls down: Eq. (8) defines F_BP = X D_I^{-1/2} U G_BP U^T D_I^{-1/2}. X D_I^{-1/2} is m×n; U G_BP U^T is (m+n)×(m+n); the product is undefined. No zero-padding into the joint node space is described, so the central "bandpass filter" cannot be computed as written. The natural fix is to project onto the item block of the eigenbasis, but that has to be stated explicitly. Appendix G, which is supposed to justify the mid-frequency claim, is asserted without showing the analysis. The reported average NDCG@10 improvement also does not reconcile: the short abstract says 5.12%, the introduction says 6.77%, and Table 4 works out to about 5.4%. The split protocol is 8:1:1 in Section 4.1 but 8:2 in Appendix F, and the paper never says whether the unified graph is built from training interactions only—so test-window leakage into graph construction is an open question. Five of the ten baselines are copied from earlier papers with no reruns or code; the main table has no error bars. None of this is evidence of fabrication, but it means the central empirical claims are not yet reproducible.

The intended audience is the GSP-CF and graph-based recommendation community. The novel graph construction and dual-filter idea deserve serious referee time, but the paper needs major revision: fix the equations, clarify the spectral basis, show the Appendix G analysis, reconcile the splits, report error bars, and release code. Send it to peer review, but expect substantial changes.

Referee Report

4 major / 5 minor

Summary. GSPRec proposes a graph-spectral collaborative filtering method that enriches the user-item graph with symmetrized item-item transition edges derived from the order of user interactions. These edges are diffused over multiple hops with exponential decay, and the resulting unified graph Laplacian is filtered with a Gaussian bandpass filter targeting mid-frequency components, plus a low-pass filter for global popularity, whose outputs are fused with a weight phi. The paper claims consistent improvements over GCN-based and GSP-based baselines on ML100K, ML1M, Netflix, and Beauty, with average NDCG@10 gains reported as 5.12% in one abstract and 6.77% in another. The manuscript provides an explicit algorithm listing, a convergence lemma for the diffusion, and ablations on ML1M.

Significance. If the method is sound, the paper makes a useful contribution by showing how sequential item-proximity information can be encoded into a symmetric Laplacian and exploited through mid-frequency spectral filtering, a band that most prior low-pass graph recommenders discard. The paper also helps map the design space of GSP-based CF by treating graph construction and filter design as coupled choices. Strengths include a self-contained algorithm description, a parameter table for reproducibility, and a stated convergence condition for the diffusion step. However, the central bandpass formula is dimensionally invalid, the evaluation protocol is internally inconsistent, and several baseline numbers are taken from other papers under an asserted protocol equivalence. These issues currently prevent the reader from verifying the headline performance claims.

major comments (4)

[§3.3, Eq. (8); Algorithm 1] Equation (8) is dimensionally invalid as written: X D_I^{-1/2} is m×n, while U G_BP U^T is (m+n)×(m+n), so the product cannot be formed; Algorithm 1 repeats the same expression. The bandpass operation is the central mechanism of the paper, so Table 4 cannot be traced to a well-defined algorithm. The authors must state the intended formula precisely (e.g., embedding the item signal in the full node space as [0, X D_I^{-1/2}] and then extracting the item block, or restricting the filter to an item-only eigenbasis) and, if the corrected formula differs from what was run, re-run the experiments.
[§4.1 vs Appendix F] The evaluation protocol is stated inconsistently: §4.1 reports an 8:1:1 train/validation/test split, while Appendix F reports an 8:2 train/test split with 10% of the training set for validation. It is also never stated whether the unified graph and its Laplacian are built from the training split only or from all interactions. Because the item-item transition edges are derived from user sequences, building the graph from all interactions would let test-window transitions enter the filtering step. The authors should specify the exact protocol and confirm that the baseline numbers, including those copied from Xia et al., use the same split.
[Appendix E] The headline comparison rests on an external premise: results for LR-GCCF, IMP-GCN, SimpleX, UltraGCN, and GF-CF are reported 'as presented in Xia et al. (2024; 2025), which used identical dataset preprocessing and evaluation protocols.' Since the manuscript itself contains contradictory protocol statements, this equivalence is not established. The authors should either run these baselines under their own protocol or provide a detailed equivalence argument, including identical data files, split indices, and evaluation code.
[Appendix G] The empirical basis for the bandpass design is asserted but not shown. The text states that spectral coefficients for intermediate eigenvalues 'have significantly higher magnitude for user-specific preference patterns compared to global trends,' but no figure, table, or procedure supporting this claim appears in Appendix G. This claim is load-bearing for the choice of a mid-frequency filter, and it should either be documented with the actual analysis or removed.

minor comments (5)

[Title and abstracts] The manuscript contains two conflicting abstracts: the initial abstract reports an average NDCG@10 improvement of 5.12% and states that removing item-item proximity still surpasses baselines, while the abstract on page 2 reports 6.77%; the title also appears in two forms. The authors should reconcile these.
[Figures 4 and 5] The captions and axis labels contain 'uni00A0' artifacts; the displayed text should be cleaned up.
[Eq. (4)] The notation \tilde X_I is defined in Eq. (4) but never used elsewhere in the paper; please either use it or remove it.
[Table 4] Table 4 would benefit from standard errors or significance tests, since Tables 5 and 6 report them only for ML1M and the abstract claims consistent gains across all datasets and metrics.
[§3.3 and Appendix C] The statement that mid-frequency components 'correspond to more personalized patterns' is supported only by a qualitative total-variation argument; the argument does not by itself establish that user-specific patterns concentrate in the band 0.3 < \lambda < 0.8, which is an empirical matter that should be made explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's test-set claims are not entailed by its fitted inputs or by self-citation; identified problems are correctness/reproducibility defects, not circular reductions.

full rationale

I walked the derivation chain from graph construction (Eqs. 2, 3, 5, 6), through the dual spectral filters (Eqs. 7-11), to the reported benchmark numbers. The model's output Y is a fixed linear function of the constructed adjacency and the tuned filter parameters; the headline comparisons in Table 4 are held-out test metrics, so they are not forced by construction from the validation-tuned parameters in Table 3. The mid-frequency preference claim in Section 3.3 and Appendix G is an empirical assertion presented without shown evidence, but it is not a circular reduction: the bandpass filter is defined independently by Eq. 7, and its center/width are tuned on validation rather than defined by the test result. The paper cites Xia et al. for baseline numbers, but that is an external citation, not a self-citation; the only self-citations (Kang & McAuley 2018 for SASRec, Ni et al. 2019 for the Beauty dataset) are standard references and are not load-bearing for the central claim. No uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. The serious dimensional inconsistency in Eq. 8 (X D_I^{-1/2} is m×n while U G_BP U^T is (m+n)×(m+n)), the ambiguity about whether the unified graph is built from training interactions only, the 8:1:1 vs 8:2 split discrepancy, and the unshown Appendix G analysis are all reproducibility/correctness risks, but they are not instances of a prediction being equivalent to its input by definition. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The ledger reflects what the paper pulls in without independent proof. Five sets of fitted or hand-picked numbers (c, w, phi, r per dataset; alpha and d universal) carry the results; the diffusion convergence condition is stated but its key quantity rho(S') is never reported. The qualitative story rests on two domain assumptions: symmetrization keeps the useful signal from item transitions, and mid-frequency coefficients encode user-specific preferences, the latter asserted in Appendix G without shown evidence. The mathematical axioms are standard spectral graph theory. No invented entities are introduced; the user-user similarity matrix C_U is computed directly from the interaction matrix and is not a new postulated object.

free parameters (5)

Bandpass center c = 0.2 (ML100K), 0.8 (ML1M), 0.4 (Netflix), 0.8 (Beauty)
Tuned per dataset on validation (Table 3); the paper's mid-frequency personalization narrative is attached to this tuned center.
Bandpass width w = 0.1 (ML100K), 0.1 (ML1M), 0.3 (Netflix), 0.3 (Beauty)
Tuned per dataset (Table 3); controls the width of the boosted frequency band.
Fusion weight phi = 0.5 (ML100K), 0.3 (ML1M), 0.3 (Netflix), 0.5 (Beauty)
Tuned per dataset (Table 3); blends bandpass and low-pass outputs and can absorb scale differences between the two filters.
Truncated eigenvector count r = 32 (ML100K), 128 (ML1M), 256 (Netflix), 512 (Beauty)
Chosen per dataset (Table 3); defines the low/mid-frequency subspace the filters act on.
Diffusion depth d and decay alpha = d=2, alpha=0.4 on all datasets
Set by hand and claimed universal (Section 4.1); Lemma 1 requires alpha < 1/rho(S') but rho(S') is never reported for any dataset.

assumptions (5)

domain assumption Symmetrizing directed item transitions preserves the useful signal (rigid direction is unreliable).
Section 3.2 symmetrizes the directed transition matrix S into S', citing BERT4Rec for direction unreliability; the paper's Limitations concede that 'symmetric construction trades directional information for spectral stability.' If direction carried the signal, this construction discards it.
ad hoc to paper Mid-frequency Laplacian components encode user-specific preferences, low frequencies encode popularity, and high frequencies are noise.
Appendix G asserts that spectral coefficients in 0.3 < lambda_i < 0.8 have higher magnitude for user-specific patterns, but shows no figure, table, or procedure; the bandpass filter design relies on this unshown claim.
domain assumption The truncated diffused symmetric graph (d=2, alpha=0.4) captures the sequentially relevant item proximity.
Section 3.2 fixes d and alpha across all datasets without sensitivity analysis; higher-order structure beyond two hops is dropped by the truncation.
standard math Standard normalized Laplacian properties (symmetric, positive semidefinite, eigenvalues in [0,2]).
Propositions 2 and 3 restate Chung (1997); used to justify the spectral filtering framework and the eigenvalue band interpretation.
domain assumption Baseline results copied from Xia et al. (2024; 2025) are comparable because protocols are identical.
Appendix E reports five baselines 'as presented in' those papers; the paper's own split text is 8:1:1 (Section 4.1) versus 8:2 (Appendix F), so protocol identity is asserted, not demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering." pith.science (2026). https://pith.science/paper/3U2EE5XO

@misc{pith2026250511552,
  author       = {Pith},
  title        = {Pith review of: GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3U2EE5XO}},
  note         = {Machine review of arXiv:2505.11552}
}

read the original abstract

Graph-based collaborative filtering methods act as low-pass filters in the spectral domain and discard the intermediate-frequency components where community-level user preferences reside. Existing GSP-based methods address the loss through sophisticated filter designs, yet derive item representations from the user-item interaction matrix alone. The interaction matrix captures which items each user interacted with, but not which items appear close together in users' interaction sequences. We propose GSPRec, a graph spectral collaborative filtering framework that produces richer item spectral representations by incorporating item-item proximity derived from user interaction ordering before spectral filtering. GSPRec derives item-item edges from user interaction ordering and strengthens the edges through multi-hop diffusion with exponential decay. The unified graph topology incorporates the diffused edges alongside user-item interactions. The resulting Laplacian exposes intermediate-frequency structure that a Gaussian bandpass filter selectively amplifies. A low-pass filter retains broad popularity trends. Experiments on four real-world datasets show that GSPRec outperforms all graph CF baselines, with average improvements of 5.12% in NDCG@10. Ablation studies establish that graph construction and filter design are coupled. GSPRec without the bandpass filter falls below every GSP baseline, whereas GSPRec without item-item proximity still surpasses baselines.

Figures

Figures reproduced from arXiv: 2505.11552 by the authors.

**Figure 3.** Multi-hop diffusion encoding item proximity from sequential patterns. Top: S ′ . Middle: Powers (S ′ ) k . Bottom: S (3) encoding decayed proximity. Symmetric Normalization (S˜): The resulting symmetric, diffused matrix S (d) is then symmetrically normalized to produce the final item-item graph component: S˜ = D −1/2 S S (d)D −1/2 S , where DS = diag(S (d)1n) (3) This standard normalization technique Chung (1997… view at source ↗

**Figure 4.** Effect of bandpass center position (c) on NDCG@10 across datasets. Each line represents a different filter bandwidth (w). 0.1 0.2 0.3 Width (w) 0.61 0.62 0.63 NDCG@10 ML100K 0.1 0.2 0.3 Width (w) 0.50 0.52 ML1M 0.1 0.2 0.3 Width (w) 0.582 0.584 0.586 0.588 Netflix 0.1 0.2 0.3 Width (w) 0.066 0.068 Beauty c=0.2 c=0.4 c=0.6 c=0.8 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Effect of bandpass width (w) on NDCG@10 across datasets. Each line represents a different center position (c). Optimal width depends on center position and dataset characteristics. amplification. This adaptive behavior explains why fixed low-pass or high-pass approache…

**Figure 6.** Figure 6: Frequency response curves of different filter configurations all datasets. The low-pass filter [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[1]

The netflix prize.Proceedings of KDD Cup and Workshop 2007, pp

James Bennett and Stan Lanning. The netflix prize.Proceedings of KDD Cup and Workshop 2007, pp. 3–6,

work page 2007
[2]

The degree matrix DS =diag(S (d)1n) will have non-negative diagonal entries

will also have non-negative entries, i.e., S(d)≥0 . The degree matrix DS =diag(S (d)1n) will have non-negative diagonal entries. For non-isolated nodes in the graph represented by S(d), 14 Preprint. these entries will be positive, making D−1/2 S well-defined with non-negative real entries. Hence, ˜S=D −1/2 S S(d)D−1/2 S ≥0. Therefore,Ais both symmetric an...

work page 1997
[3]

For implementation, we 16 Preprint. process datasets following the procedure described in Section 4.1, which involves splitting datasets in an 8:2 ratio for training/testing, with 10% of the training set used for validation. When implementing GSPRec, the key parameter settings are: • Diffusion depthd= 2and decay factorα= 0.4for sequential graph constructi...

work page 2004
[9]

Frequency- aware graph signal processing for collaborative filtering

Jiafeng Xia, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. Frequency- aware graph signal processing for collaborative filtering. InCompanion Proceedings of the ACM on Web Conference 2025, pp. 1417–1421,

work page 2025
[12]

toO(r2(m+n)) (or better, depending on specific variant, e.g., linear in (m+n) for matrix-vector products), with minimal performance impact. F.2.2 GRAPHSPARSIFICATION For large-scale graphs, we can apply spectral sparsification techniques Spielman & Teng (2011) that preserve the essential spectral properties while reducing edge density, accelerating both g...

work page 2011
[2006]

Justifying recommendations using distantly-labeled reviews and fine-grained aspects

Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pp. 188–197,

2019
[2015]

Retagnn: Relational temporal attentive graph neural networks for holistic sequential recommendation

Cheng Hsu and Cheng-Te Li. Retagnn: Relational temporal attentive graph neural networks for holistic sequential recommendation. InProceedings of the web conference 2021, pp. 2968–2979,

work page 2021
[2019]

Hierarchical graph signal processing for collaborative filtering

Jiafeng Xia, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. Hierarchical graph signal processing for collaborative filtering. InProceedings of the ACM Web Conference 2024, pp. 3229–3240,

work page 2024

Show all 12 references

[2020]

Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939,

Bal´azs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939,

arXiv
[2021]

Personalized graph signal processing for collaborative filtering

Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. Personalized graph signal processing for collaborative filtering. InProceedings of the ACM Web Conference 2023, pp. 1264–1272,

2023
[2024]

Interest-aware message-passing gcn for recommendation

Fan Liu, Zhiyong Cheng, Lei Zhu, Zan Gao, and Liqiang Nie. Interest-aware message-passing gcn for recommendation. InProceedings of the web conference 2021, pp. 1296–1305,

2021
[2025]

Algebraic signal processing theory.arXiv preprint cs/0612077,

Markus P¨uschel and Jos´e MF Moura. Algebraic signal processing theory.arXiv preprint cs/0612077,

arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

The netflix prize.Proceedings of KDD Cup and Workshop 2007, pp

James Bennett and Stan Lanning. The netflix prize.Proceedings of KDD Cup and Workshop 2007, pp. 3–6,

work page 2007

[2] [2]

The degree matrix DS =diag(S (d)1n) will have non-negative diagonal entries

will also have non-negative entries, i.e., S(d)≥0 . The degree matrix DS =diag(S (d)1n) will have non-negative diagonal entries. For non-isolated nodes in the graph represented by S(d), 14 Preprint. these entries will be positive, making D−1/2 S well-defined with non-negative real entries. Hence, ˜S=D −1/2 S S(d)D−1/2 S ≥0. Therefore,Ais both symmetric an...

work page 1997

[3] [3]

For implementation, we 16 Preprint. process datasets following the procedure described in Section 4.1, which involves splitting datasets in an 8:2 ratio for training/testing, with 10% of the training set used for validation. When implementing GSPRec, the key parameter settings are: • Diffusion depthd= 2and decay factorα= 0.4for sequential graph constructi...

work page 2004

[4] [9]

Frequency- aware graph signal processing for collaborative filtering

Jiafeng Xia, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. Frequency- aware graph signal processing for collaborative filtering. InCompanion Proceedings of the ACM on Web Conference 2025, pp. 1417–1421,

work page 2025

[5] [12]

toO(r2(m+n)) (or better, depending on specific variant, e.g., linear in (m+n) for matrix-vector products), with minimal performance impact. F.2.2 GRAPHSPARSIFICATION For large-scale graphs, we can apply spectral sparsification techniques Spielman & Teng (2011) that preserve the essential spectral properties while reducing edge density, accelerating both g...

work page 2011

[6] [2006]

Justifying recommendations using distantly-labeled reviews and fine-grained aspects

Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pp. 188–197,

2019

[7] [2015]

Retagnn: Relational temporal attentive graph neural networks for holistic sequential recommendation

Cheng Hsu and Cheng-Te Li. Retagnn: Relational temporal attentive graph neural networks for holistic sequential recommendation. InProceedings of the web conference 2021, pp. 2968–2979,

work page 2021

[8] [2019]

Hierarchical graph signal processing for collaborative filtering

Jiafeng Xia, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. Hierarchical graph signal processing for collaborative filtering. InProceedings of the ACM Web Conference 2024, pp. 3229–3240,

work page 2024