Collaborative Filtering Meets Spectrum Shift: Connecting User-Item Interaction with Graph-Structured Side Information

Cong Xu; Jun Wang; Wei Zhang; Yunhang He

REVIEW 3 major objections 9 minor 59 references

Collaborative Filtering Meets Spectrum Shift: Connecting User-Item Interaction with Graph-Structured Side Information

T0 review · 3 major / 9 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that spectrum shift in augmented user-item graphs is why existing spectral GNNs fail on side information, and that an affine correction restores performance with gains up to 23%.

desk verdict A clean spectral diagnosis and a neat plug-in fix, but the experiments don't yet isolate the spectral correction from the side information itself. read the letter →

arxiv 2502.08071 v2 pith:FAP3AWE6 submitted 2025-02-12 cs.IR

Yunhang He , Cong Xu , Jun Wang , Wei Zhang This is my paper · ORCID

classification cs.IR

keywords spectrumshiftcollaborativefilteringgraph-structuredsideinformationspectralgraphneuralnetworksmultimodalrecommendationsocialaugmentedadjacencymatrixaffinecorrection

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 argues that existing spectral graph recommender models fail on augmented graphs because adding graph-structured side information moves the spectrum of the normalized adjacency matrix away from the assumed full interval $[-1,1]$. The maximum eigenvalue stays at $1$, but the minimum eigenvalue drifts rightward, so filters designed for the full spectrum weight frequencies incorrectly. The authors propose Spectrum Shift Correction (SSC), an affine map $\phi(\lambda)=(\lambda-\mu)/\Delta$ that re-centers and rescales the spectrum before applying the original filter, and they show this is equivalent to applying the same operation directly to the augmented adjacency matrix. On five public datasets spanning social and multimodal recommendation, SSC improves LightGCN and JGCF with relative gains up to 23% and no extra per-epoch computational cost. If true, the result means any graph-structured side information can be plugged into existing collaborative-filtering GNNs through a simple preprocessing step rather than task-specific architecture design.

What carries the argument

The central object is the normalized augmented adjacency matrix $\tilde{A}_+$ built by splicing a user-user social graph $S_U(\kappa)$ or item-item $\kappa$NN similarity graph $S_I(\kappa)$ into the bipartite interaction matrix $A$. The load-bearing identity is the affine spectral correction $\phi(\lambda)=(\lambda-\mu)/\Delta$, which maps the shifted spectrum back toward $[-1,1]$; the paper proves $\phi(\tilde{A}_+)=\Delta^{-1}(\tilde{A}_+-\mu I)$, so the correction is applied directly to the matrix. The shift itself is diagnosed with the oracle Rayleigh quotient $R(U_{:,k};B)=|U_{:,k}^T B U_{:,k}|/\|U_{:,k}\|^2$ against the test interaction matrix $B$, and the factors $\mu,\Delta$ are estimated by one power iteration on $I-\tilde{A}_+$.

What would settle it

Compute per-eigenvector ranking utility directly: train LightGCN on augmented graphs, split test interactions by eigenvector participation, and check whether large $R(U_{:,k};B)$ eigenvectors are the ones that improve NDCG. If the correlation is absent, or if SSC produces the same gains when the side-information block is replaced by a random symmetric matrix of the same density, the spectrum-shift explanation is not what drives the reported improvements.

Watch

Extended reading notes

Core claim

The paper's central claim is that the 'spectrum shift' of the augmented adjacency matrix—max eigenvalue fixed at 1, min eigenvalue shifting right as side information is integrated—is the reason previous spectral GNNs underperform on social and multimodal recommendation. It proves in Theorem 3.1 that under $\kappa$-rescaling the minimum eigenvalue tends to $\min(\lambda_{\min}(\tilde{S}),0)$, and under $\kappa$-nearest-neighbor graph construction it tends to approximately $0$. It then shows that composing any existing filter $g$ with the affine correction $\phi(\lambda)=(\lambda-\mu)/\Delta$ restores the intended frequency weighting, and that $\phi(\tilde{A}_+)=(\tilde{A}_+-\mu I)/\Delta$, so the correction is a matrix operation with no eigendecomposition and no added training cost. The experiments report that SSC improves both the low-frequency filter LightGCN and the high-frequency-sensitive Jacobi filter JGCF across five datasets, with the largest relative improvements on multimodal benchmarks.

Load-bearing premise

The load-bearing premise is that the absolute Rayleigh quotient of an eigenvector against the test interaction matrix really tells you which frequencies matter for recommendation; if that oracle metric does not track BPR or NDCG ranking quality, the observed mismatch is an artifact of the metric and the affine correction is just a generic reparameterization of the filter.

Editorial extensions

If this is right

Existing spectral collaborative filtering models can absorb arbitrary graph-structured side information simply by preprocessing the augmented adjacency matrix, removing the need for task-specific modules or auxiliary losses.
SSC applies to multiple backbones—LightGCN, JGCF, AFDGCF, SGL, SimGCL—so the correction is orthogonal to model architecture and can be combined with future filters.
Because $\kappa$-nearest-neighbor item graphs shift the spectrum more than $\kappa$-rescaled social graphs, multimodal settings stand to gain more from the correction, matching the larger reported improvements there.
The unified convolution of the augmented graph plus SSC is more robust to noise in side information than parallel-branch designs, since a corrupted side branch cannot dominate a separate representation stream.

Reading between the lines

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

If the mechanism is as general as stated, the same affine correction should transfer to other graph-structured side information such as knowledge graphs or co-purchase graphs, with $\mu$ and $\Delta$ re-estimated per graph; this is a direct test the paper does not run.
A practical diagnostic suggested by the theorem is to monitor $\lambda_{\min}(\tilde{A}_+)$ as a cheap, training-free proxy for how much side information the augmented graph contributes, potentially guiding the choice of $\kappa$ before training.
The fixed point $\lambda_{\max}\equiv 1$ plus rightward drift of $\lambda_{\min}$ means the effective graph shift operator is being re-centered; this connects SSC to normalized Laplacian regularization and suggests it may also reduce oversmoothing in deep GNNs, but that connection is not explored in the paper.
The sensitivity tables show the estimated factors are close but not always optimal, implying the optimal correction is distributional rather than exact; a learnable but stable variant that tunes $\mu$ and $\Delta$ per layer could extend SSC beyond fixed filters.

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

The paper's core observation is correct and useful: when you add graph-structured side information to a user-item bipartite graph, the spectrum of the normalized adjacency matrix compresses from the right, breaking the full-spectrum assumption that LightGCN and similar spectral filters rely on. The proposed fix, an affine reparametrization of the filter (shifting and scaling), is simple, model-agnostic, and costs nothing at runtime. Theorem 3.1 is basically right given the stated asymptotics, and the proof in Appendix B is standard but clean. The authors also earn credit for shipping code and for testing on several backbones beyond the main two, which suggests the plug-in idea generalizes. That's a real contribution to the graph-CF subfield.

The soft spot is the causal claim. The main tables compare '+SSC' on the augmented graph against LightGCN/JGCF on the original bipartite graph only. There is no row that applies the backbone directly to the augmented graph without SSC. That missing control means most of the reported 23% improvement could simply be the value of adding side information, plus the extra tuning freedom from kappa, mu, and delta. The sensitivity tables undercut the spectral story further: on Baby, moving delta from 1.0 to 0.8 (a mild correction) buys only about 2% relative gain, and Table 7 shows Electronics LightGCN running at mu=0, delta=0.55, which maps lambda=1 to 1.818, not back to 1. So the tuned configuration is not actually restoring the spectrum to [-1,1]. That does not sink the paper, but it means the mechanism is not cleanly supported by the experiments.

Two smaller concerns: the oracle Rayleigh quotient in Figure 2 uses the test interaction matrix, which makes the importance measure dependent on the ground truth; and the main tables lack error bars, even though t-tests are mentioned. The authors are honest about the two extra hyperparameters in the limitations.

I would send this to serious review, but conditional. The authors need to run the augmented-graph no-SSC control and report error bars; if the control shows SSC adds little beyond the side information, the paper should be reframed as a spectral analysis plus a modest trick, not a 23% improvement. As it stands, it deserves referees' time because the diagnosis is novel and the theory holds.

Reading group: maybe, if someone in the group works on graph spectral methods in CF.

Referee Report

3 major / 9 minor

Summary. The paper studies spectral GNNs for collaborative filtering when graph-structured side information (user-user social networks or item-item multimodal similarity graphs) is appended to the user-item interaction graph. It documents a 'spectrum shift': the augmented, symmetrically normalized adjacency matrix retains lambda_max = 1 while its lambda_min moves rightward, so spectral filters designed for bipartite graphs under the assumption that the spectrum spans [-1, 1] assign mismatched frequency weights. The paper proves (Theorem 3.1, Appendix B.2) that lambda_max(kappa) = 1 and that lambda_min tends to min(lambda_min(S-tilde), 0) for kappa-rescaled side information and to approximately 0 for kappa-nearest-neighbor graphs, and it proposes Spectrum Shift Correction (SSC), an affine map phi(lambda) = (lambda - mu)/Delta applied to the augmented adjacency matrix before filtering (Eqs. 7-8). SSC is plug-and-play for LightGCN and JGCF, requires no per-epoch overhead, and is evaluated on three multimodal and two social recommendation datasets with reported relative gains up to 23% over the backbones. The appendices add results on three further backbones (AFDGCF, SGL, SimGCL) and two larger datasets.

Significance. The theoretical core is sound and useful: the proof that lambda_max = 1 for any symmetrically normalized augmented adjacency matrix and the asymptotic characterization of lambda_min for the two graph-integration mechanisms are clean given the stated approximations, and the observation that an affine correction of the spectrum is equivalent to an affine correction of the matrix itself (Eqs. 7-8) makes the proposal genuinely zero-overhead at training time. The paper ships code, reports significance tests over five runs, and the experimental sweep (two main backbones, three additional backbones, two larger datasets, and a noise-robustness study) is considerably broader than the minimum needed for a methods paper. The main gap is attribution: the experiments do not isolate the effect of the affine correction from the effect of simply training the backbone on the augmented graph, and the tuned hyperparameters in Table 7 are inconsistent with the stated spectrum-restoration mechanism.

major comments (3)

[§5.2 (Tables 3-6); §3.2] The headline improvement (up to 23.33%) is computed relative to the backbone trained on the original bipartite graph, but the '+SSC' row differs from that baseline by two simultaneous interventions: the augmented graph A+ (side information plus kappa) and the affine correction (mu, Delta). The manuscript never reports the control that isolates the correction, namely the backbone applied to A+ without SSC. The sensitivity tables provide a partial view: on Baby, the near-identity settings mu=0 (Table 5, with Delta at its tuned value) and Delta=1 (Table 6, with mu at its tuned value) already reach R@10 = 0.0663 and 0.0654, respectively, against the original-graph LightGCN's 0.0543, while the best corrected configuration reaches 0.0667; the marginal gain of the correction is therefore only about 2% relative, not the headline 22.61%. Moreover, Table 7 shows that for LastFM the tuned JGCF configuration is mu=0, Delta=1.0, i.e., phi is the identity map, so the reported 1.1-1.4% gain of JGCF+SSC on LastFM is entirely attributable to the augmented graph and kappa, not to any spectral correction. The paper should add explicit 'augmented graph without SSC' rows (equivalently mu=0, Delta=1) for every backbone/dataset, report the incremental gain of the correction separately, and temper the causal wording in Section 3.2 and the abstract accordingly.
[§3.1, Eq. (4); Figure 2] The entire diagnosis of 'mismatched importance' rests on the oracle score R(U_{:,k}; B), which measures the alignment of the augmented graph's eigenvectors with the test interaction matrix B. The manuscript provides no evidence that this Rayleigh-quotient score tracks the actual training objective (BPR loss) or the evaluation metrics (Recall/NDCG). If R is not predictive of ranking performance, the rightward shift in Figure 2 is an artifact of the chosen metric, and the claimed mismatch does not explain the suboptimality of the baselines. A concrete validation would be to correlate R(U_{:,k}; B) with the change in NDCG when the filter's weight on U_{:,k} is ablated, or to show that spectral filters emphasizing high-R frequencies outperform filters emphasizing low-R frequencies on the same augmented graph. In addition, the oracle nature of the metric, which uses test labels, should be stated explicitly in Section 3.1 so that readers can distinguish the motivational analysis from the model-selection protocol.
[§4.1, §5.3, Table 7] The stated mechanism of SSC is the restoration of the filter domain: Section 4.1 says the full spectrum is 'recovered' when mu = (lambda_max + lambda_min)/2 and Delta = (lambda_max - lambda_min)/2, and Section 5.3 (RQ3) says the corrected spectra are 'more closely aligned within the range of [-1, 1]'. However, none of the ten tuned configurations in Table 7 has (1 - mu)/Delta <= 1, so lambda_max = 1 is mapped outside the intended domain in every case (for example, Ciao LightGCN (mu=0.15, Delta=0.4) gives phi(1) = 2.125; Electronics LightGCN (0, 0.55) gives phi(1) = 1.818; LastFM LightGCN (0, 0.5) gives phi(1) = 2.0; LastFM JGCF (0, 1.0) is the identity map). Evaluated outside [-1, 1], the Jacobi-polynomial construction in Eq. (10) no longer has its stated orthogonality guarantees, and for LightGCN the coefficients simply amplify low frequencies. The reported gains are therefore not evidence for the restoration mechanism: either Figure 4 and Table 7 are inconsistent (Figure 4 apparently shows corrected spectra inside [-1, 1]), or the mechanism is a generic re-scaling reparameterization rather than spectrum restoration. This must be reconciled, and the Conclusion's stated limitation (two extra hyperparameters) understates the issue because the tuned values do not implement the method's stated mechanism.

minor comments (9)

[Abstract; §5.2 (Table 3 Time column)] The claim of 'no additional computational overhead' is only valid relative to running the backbone on the same augmented graph; Table 3 shows +SSC taking 77m12s versus 59m20s for LightGCN and 166m34s versus 140m10s for JGCF on Electronics, which a reader will naturally read as a 19-30% wall-time increase over the tabulated baseline, so the reference point should be stated explicitly.
[Table 4; §5.2] For JGCF+SSC on LastFM, R@10 shows no significance marker (0.2583 versus 0.2555, p > 0.05), yet the bulleted discussion groups all gains as improvements; please state explicitly which differences are not statistically significant.
[Figures 2 and 6] The y-axis tick labels and the kappa annotations are garbled (for example, stray '0.0 2.5 0.0' labels, '0 110', and 'kappa=1.5' placed mid-panel), which makes the oracle-importance curves difficult to read; the figures should be redrawn.
[Eq. (10)] The expression 'P^{a,b}_l (l)' should read 'P^{a,b}_l(lambda)'.
[§3.2] The statement that the kappa-nearest-neighbors mechanism 'results in a more pronounced mismatch' than kappa-rescaling is inferred by comparing Ciao (social, kappa-rescaling) with Baby (multimodal, kappa-NN); because the dataset and side-information type vary together with the construction mechanism, this attribution is not supported and should be qualified or tested within a single dataset.
[Acknowledgments] The acknowledgment of an anonymous reviewer by ID ('Reviewer VXX2') and the reference to the rebuttal period should be removed from the arXiv version, as it reveals review-process details.
[Tables 5-6] The captions do not state what the non-varying hyperparameter is fixed to; please specify (for example, 'mu is varied with Delta fixed at its Table 7 value') so that the sweeps are interpretable.
[Algorithm 1] The pseudocode does not make clear that the input A is the symmetrically normalized augmented matrix A+ and that power_iteration is run on I - A+; a brief comment in the pseudocode would remove this ambiguity.
[Tables 3-4] The main tables report significance stars but no standard deviations or confidence intervals; please add them (at least in the appendix) so that the magnitude of the improvements can be assessed.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-level circularity: Theorem 3.1 is label-free and SSC is an algebraic identity; the empirical attribution is underdetermined but not circular.

full rationale

The paper's central mathematical claims do not reduce to their inputs. Theorem 3.1 (Section 3.2, Appendix B.2) is proved from the limit of the normalized augmented adjacency matrix as kappa tends to infinity; it does not use the test matrix B or any fitted quantity. The SSC transformation phi(lambda)=(lambda-mu)/Delta (Eq. 7) and its matrix form phi(A_tilde_+)=(1/Delta)(A_tilde_+-mu I) (Eq. 8) are a standard functional-calculus identity; no term in the identity is defined by the target it is said to predict. The claimed equivalence therefore holds by algebra, not by construction of the result. The empirical story has two weaknesses that a reader might confuse with circularity: (i) the 'oracle spectrum importance' R(U_{:,k};B) in Eq. (4) is defined using the test interaction graph B, so the observed shift is an analysis artifact of that metric; and (ii) the reported improvements come from jointly tuning kappa, mu, Delta on validation NDCG (Section 5.1), with the paper itself conceding 'it is necessary to retune them' (Section 4.1) and listing tuning as a limitation (Section 7). These are experimental-attribution gaps, including the absence of an augmented-graph-no-SSC control, but they are not circular reductions: no equation in the derivation is equivalent to the empirical outcome by definition. Self-citations (StableGCN for training instability, STAIR as a baseline) are not load-bearing; Table 1 internally documents the instability, and STAIR is an external comparison point. Accordingly, the derivation chain is self-contained and the circularity score is 0.

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

No new physical or model entities are introduced. The spectrum shift is a property of the existing augmented adjacency matrix, not a new postulated object. The main fitted quantities are mu, Delta, and kappa, which are tuned per dataset on validation performance.

free parameters (3)

mu (spectrum shifting factor) = e.g., 0.05 for Baby LightGCN, 0.0 for Electronics LightGCN (Table 7)
Tuned on validation NDCG around the power-iteration estimate; drives the claimed spectral correction.
Delta (spectrum scaling factor) = e.g., 0.75 for Baby LightGCN, 0.55 for Electronics LightGCN (Table 7)
Tuned jointly with mu; selected values do not always correspond to restoring [-1,1] under Eq. (7).
kappa (side-information intensity) = e.g., 5-10 for multimodal, 0.75-1.0 for social (Table 7)
Controls how much side information enters the augmented graph; the spectrum shift is measured as kappa grows and is chosen per dataset.

assumptions (4)

domain assumption The Rayleigh quotient R(x;B)=|x^T B x|/||x||^2 on the test interaction matrix B measures the recommendation importance of frequency x.
Used in Section 3.1 and Figure 2 to identify the mismatch; no proof is given that this metric tracks BPR or NDCG ranking loss.
ad hoc to paper The effect of side information on the spectrum is approximately an affine shift and scale over the range [-1,1].
The authors admit in Section 4.1 that the real transformation is considerably more intricate, yet Eq. (7) assumes a single affine correction is sufficient.
domain assumption The normalized kappa-nearest-neighbors similarity graph converges to the all-ones matrix (1/|U|) 1 1^T as kappa grows.
Used in the proof of Theorem 3.1 in Appendix B.2; depends on kNN neighborhoods covering all nodes and ignores self-loop and degree details.
standard math Spectral decomposition applies to the real symmetric normalized augmented adjacency matrix.
Standard linear algebra background invoked throughout Section 2.2 and Theorem 3.1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Collaborative Filtering Meets Spectrum Shift: Connecting User-Item Interaction with Graph-Structured Side Information." pith.science (2026). https://pith.science/paper/FAP3AWE6

@misc{pith2026250208071,
  author       = {Pith},
  title        = {Pith review of: Collaborative Filtering Meets Spectrum Shift: Connecting User-Item Interaction with Graph-Structured Side Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FAP3AWE6}},
  note         = {Machine review of arXiv:2502.08071}
}

read the original abstract

Graph Neural Networks (GNNs) have demonstrated their superiority in collaborative filtering, where the user-item (U-I) interaction bipartite graph serves as the fundamental data format. However, when graph-structured side information (e.g., multimodal similarity graphs or social networks) is integrated into the U-I bipartite graph, existing graph collaborative filtering methods fall short of achieving satisfactory performance. We quantitatively analyze this problem from a spectral perspective. Recall that a bipartite graph possesses a full spectrum within the range of [-1, 1], with the highest frequency exactly achievable at -1 and the lowest frequency at 1; however, we observe as more side information is incorporated, the highest frequency of the augmented adjacency matrix progressively shifts rightward. This spectrum shift phenomenon has caused previous approaches built for the full spectrum [-1, 1] to assign mismatched importance to different frequencies. To this end, we propose Spectrum Shift Correction (dubbed SSC), incorporating shifting and scaling factors to enable spectral GNNs to adapt to the shifted spectrum. Unlike previous paradigms of leveraging side information, which necessitate tailored designs for diverse data types, SSC directly connects traditional graph collaborative filtering with any graph-structured side information. Experiments on social and multimodal recommendation demonstrate the effectiveness of SSC, achieving relative improvements of up to 23% without incurring any additional computational overhead. Our code is available at https://github.com/yhhe2004/SSC-KDD.

Figures

Figures reproduced from arXiv: 2502.08071 by the authors.

**Figure 2.** Oracle spectrum importance R (𝑈:,𝜆; 𝐵) as more side information is integrated (i.e., 𝜅 increases). or 𝑆𝐼 (𝜅) integrated. Therefore, the full-spectrum assumption no longer holds true for an augmented graph. • In comparison to 𝑆𝑈 , a more significant rightward shift is observed when 𝑆𝐼 is incorporated. The 𝜅-nearest-neighbors mechanism appears to result in a more pronounced mismatch for spectral GNNs, consequently w… view at source ↗

**Figure 3.** Pipeline of SSC. algorithm that returns the largest (in absolute value) eigenvalue of the given matrix. Accordingly, we have: 1 − power_iteration(𝐼 − 𝐴˜ +) ≈ 1 − (1 − 𝜆min) = 𝜆min, and the ‘optimal’ 𝜇 and Δ. Remark that an exact spectral correction may not necessarily be optimal. This is because the spectrum shift induced by the augmented graph is distributionally similar to a shifting and scaling transformation, bu… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Effect of SSC. Implementation details. For a fair comparison, we adopt the Adam optimizer [18] for training and fix embedding dimension to 64, batch size to 2048. We tune learning rates of all methods within {1.e4, 5.e-4, 1.e-3, 5.e-3}. Other involved model/loss hyper…

**Figure 5.** Figure 5: Model performance under progressively increasing [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Oracle spectrum importance R (𝑈:,𝜆; 𝐵) on other 3 datasets [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 51 canonical work pages

[1]

Iván Cantador, Peter Brusilovsky, and Tsvi Kuflik. 2011. 2nd Workshop on Information Heterogeneity and Fusion in Recommender Systems (HetRec 2011). In ACM Conference on Recommender Systems (RecSys) (RecSys 2011)

work page 2011
[2]

Qiwei Chen, Huan Zhao, Wei Li, Pipei Huang, and Wenwu Ou. 2019. Behavior se- quence transformer for e-commerce recommendation in alibaba. In International Workshop on Deep Learning Practice for High-dimensional Sparse Data . 1–4

work page 2019
[3]

Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. 2021. Adaptive univer- sal generalized pagerank graph neural network. In International Conference on Learning Representations (ICLR)

work page 2021
[4]

Jeongwhan Choi, Seoyoung Hong, Noseong Park, and Sung-Bae Cho. 2023. Blurring-sharpening process models for collaborative filtering. In ACM Interna- tional Conference on Research and Development in Information Retrieval (SIGIR) . 1096–1106

work page 2023
[5]

Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. Advances in Neural Information Processing Systems (NeurIPS) 29 (2016)

work page 2016
[6]

Ahmed El-Kishky, Thomas Markovich, Serim Park, Chetan Verma, Baekjin Kim, Ramy Eskander, Yury Malkov, Frank Portman, Sofía Samaniego, Ying Xiao, et al
[7]

Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin
[8]

Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Predict then propagate: Graph neural networks meet personalized pagerank. In International Conference on Learning Representations (ICLR)

work page 2018

Show all 59 references

[9]

Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. 2017. Neural message passing for quantum chemistry. In International Conference on Machine Learning (ICML) . PMLR, 1263–1272

2017
[10]

Jiayan Guo, Lun Du, Xu Chen, Xiaojun Ma, Qiang Fu, Shi Han, Dongmei Zhang, and Yan Zhang. 2023. On Manipulating Signals of User-Item Graph: A Jacobi Polynomial-based Graph Collaborative Filtering. InACM International Conference on Knowledge Discovery & Data Mining (KDD) . 602–613

2023
[11]

Yuhe Guo and Zhewei Wei. 2023. Graph neural networks with learnable and optimal polynomial bases. InInternational Conference on Machine Learning (ICML). 12077–12097

2023
[12]

Mingguo He, Zhewei Wei, and Ji-Rong Wen. 2022. Convolutional neural net- works on graphs with chebyshev approximation, revisited. Advances in Neural Information Processing Systems 35 (2022), 7264–7276

2022
[13]

Ruining He and Julian McAuley. 2015. VBPR: Visual bayesian personalized ranking from implicit feedback

2015
[14]

Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In International Conference on World Wide Web (WWW) . 507–517

2016
[15]

Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In ACM International Conference on Research and Development in Information Retrieval (SIGIR) . 639–648

2020
[16]

Chao Huang, Huance Xu, Yong Xu, Peng Dai, Lianghao Xia, Mengyin Lu, Liefeng Bo, Hao Xing, Xiaoping Lai, and Yanfang Ye. 2021. Knowledge-aware coupled graph neural network for social recommendation. InAAAI conference on Artificial Intelligence (AAAI), Vol. 35. 4115–4122

2021
[17]

Wei Jiang, Xinyi Gao, Guandong Xu, Tong Chen, and Hongzhi Yin. 2024. Chal- lenging Low Homophily in Social Recommendation. In The Web Conference (WWW). 3476–3484

2024
[18]

Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[19]

Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR)

2016
[20]

Jingjing Li, Ke Lu, Zi Huang, and Heng Tao Shen. 2019. On both cold-start and long-tail recommendation with social data. IEEE Transactions on Knowledge and Data Engineering (TKDE) 33, 1 (2019), 194–208

2019
[21]

Jie Liao, Wei Zhou, Fengji Luo, Junhao Wen, Min Gao, Xiuhua Li, and Jun Zeng
[22]

Guojiao Lin, Meng Zhen, Dongjie Wang, Qingqing Long, Yuanchun Zhou, and Meng Xiao. 2024. GUME: Graphs and User Modalities Enhancement for Long-Tail Multimodal Recommendation. In ACM International Conference on Information and Knowledge Management (CIKM) . 1400–1409

2024
[23]

BYYLD Liu and Jiming Liu. 2013. Social collaborative filtering by trust. In Pro- ceedings of the twenty-third international joint conference on artificial intelligence, IEEE

2013
[24]

Information Sciences 589 (2022), 595–607

SocialLGN: Light graph convolution network for social recommendation. Information Sciences 589 (2022), 595–607

2022
[25]

Hao Ma, Haixuan Yang, Michael R Lyu, and Irwin King. 2008. Sorec: social recommendation using probabilistic matrix factorization. In ACM International Conference on Information & Knowledge Management (CIKM) . 931–940

2008
[26]

Hao Ma, Dengyong Zhou, Chao Liu, Michael R Lyu, and Irwin King. 2011. Rec- ommender systems with social regularization. In ACM International Conference on Web Search and Data Mining (WSDM) . 287–296

2011
[27]

Xiaoling Long, Chao Huang, Yong Xu, Huance Xu, Peng Dai, Lianghao Xia, and Liefeng Bo. 2021. Social recommendation with self-supervised metagraph informax network. In ACM International Conference on Information & Knowledge Management (CIKM). 1160–1169

2021
[28]

Rongqing Kenneth Ong and Andy WH Khong. 2025. Spectrum-based modality representation fusion graph convolutional network for multimodal recommen- dation. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining . 773–781

2025
[29]

Shaowen Peng, Kazunari Sugiyama, and Tsunenori Mine. 2022. Less is more: Reweighting important spectral graph features for recommendation. In ACM International Conference on Research and Development in Information Retrieval (SIGIR). 1273–1282

2022
[30]

Kelong Mao, Jieming Zhu, Xi Xiao, Biao Lu, Zhaowei Wang, and Xiuqiang He
[31]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme
[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 ACM International Conference on Information & Knowledge Management (CIKM) . 1619–1629

2021
[33]

Daniel Spielman. 2019. Spectral and algebraic graph theory. Yale lecture notes, draft of December 4 (2019), 47

2019
[34]

Shaowen Peng, Kazunari Sugiyama, and Tsunenori Mine. 2022. SVD-GCN: A simplified graph convolution paradigm for recommendation. In Proceedings of the 31st ACM international conference on information & knowledge management . 1625–1634

2022
[35]

J. Tang, H. Gao, and H. Liu. 2012. mTrust: Discerning multi-faceted trust in a connected world. In ACM International Conference on Web Search and Data Mining (WSDM). ACM, 93–102

2012
[36]

Tianle Wang, Lianghao Xia, and Chao Huang. 2023. Denoised self-augmented learning for social recommendation. arXiv preprint arXiv:2305.12685 (2023)

2023 arXiv
[37]

Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural Graph Collaborative Filtering. InACM International Conference on Research and Development in Information Retrieval (SIGIR) (SIGIR ’19)

2019
[38]

Xiyuan Wang and Muhan Zhang. 2022. How powerful are spectral graph neural networks. In International Conference on Machine Learning (ICML) . PMLR, 23341– 23362

2022
[39]

Tang, J.and Gao, H

H. Tang, J.and Gao, H. Liu, and A. Das Sarma. 2012. eTrust: Understanding trust evolution in an online world. 253–261 pages

2012
[40]

Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. 2019. MMGCN: Multi-modal graph convolution network for personalized recommendation of micro-video. In ACM International Conference on Multimedia (MM). 1437–1445

2019
[41]

Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised graph learning for recommendation. In Proceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval. 726–735

2021
[42]

Le Wu, Junwei Li, Peijie Sun, Richang Hong, Yong Ge, and Meng Wang. 2020. Diffnet++: A neural influence and interest diffusion network for social recom- mendation. IEEE Transactions on Knowledge and Data Engineering (TKDE) 34, 10 (2020), 4753–4766

2020
[43]

Wei Wu, Chao Wang, Dazhong Shen, Chuan Qin, Liyi Chen, and Hui Xiong
[44]

Wei Wei, Chao Huang, Lianghao Xia, and Chuxu Zhang. 2023. Multi-modal self-supervised learning for recommendation. In The Web Conference (WWW). 790–800

2023
[45]

Cong Xu, Jun Wang, and Wei Zhang. 2023. StableGCN: Decoupling and Recon- ciling Information Propagation for Collaborative Filtering. IEEE Transactions on Knowledge and Data Engineering (TKDE) (2023)

2023
[46]

Junliang Yu, Hongzhi Yin, Jundong Li, Qinyong Wang, Nguyen Quoc Viet Hung, and Xiangliang Zhang. 2021. Self-supervised multi-channel hypergraph convo- lutional network for social recommendation. In The Web Conference (WWW) . 413–424

2021
[47]

Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. 2022. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR Collaborative Filtering Meets Spectrum Shift K...

2022
[48]

Penghang Yu, Zhiyi Tan, Guanming Lu, and Bing-Kun Bao. 2023. Multi-view graph convolutional network for multimedia recommendation. In ACM Interna- tional Conference on Multimedia (MM) . 6576–6585

2023
[49]

Jinghao Zhang, Yanqiao Zhu, Qiang Liu, Shu Wu, Shuhui Wang, and Liang Wang. 2021. Mining latent structures for multimedia recommendation. In ACM International Conference on Multimedia (MM) . 3872–3880

2021
[50]

Cong Xu, Yunhang He, Jun Wang, and Wei Zhang. 2024. STAIR: Manipulating Collaborative and Multimodal Information for E-Commerce Recommendation. In AAAI Conference on Artificial Intelligence (AAAI)

2024
[51]

Xin Zhou and Zhiqi Shen. 2023. A tale of two graphs: Freezing and denois- ing graph structures for multimodal recommendation. In ACM International Conference on Multimedia (MM) . 935–943

2023
[52]

Xin Zhou, Hongyu Zhou, Yong Liu, Zhiwei Zeng, Chunyan Miao, Pengwei Wang, Yuan You, and Feijun Jiang. 2023. Bootstrap latent representations for multi- modal recommendation. In Proceedings of the ACM Web Conference (WWW) . ACM

2023
[53]

𝐷1/2 𝑈 1|U| 𝐷1/2 𝐼 1|I| # , 𝑣 − =

Zhangchi Zhu and Wei Zhang. 2024. Exploring Feature-based Knowledge Dis- tillation For Recommender System: A Frequency Perspective. In International Conference on Knowledge Discovery & Data Mining (KDD) . A Oracle Spectra We plot the oracle spectrum importance across different...

2024
[56]

Xin Zhou. 2023. MMRec: Simplifying multimodal recommendation.arXiv preprint arXiv:2302.03497 (2023)

2023 arXiv
[2009]

In Uncertainty in Artificial Intelligence (UAI)

BPR: Bayesian personalized ranking from implicit feedback. In Uncertainty in Artificial Intelligence (UAI)
[2019]

In The World Wide Web Conference (WWW)

Graph neural networks for social recommendation. In The World Wide Web Conference (WWW). 417–426
[2021]

In ACM International Conference on Information & Knowledge Management (CIKM)

UltraGCN: ultra simplification of graph convolutional networks for rec- ommendation. In ACM International Conference on Information & Knowledge Management (CIKM). 1253–1262
[2022]

In International Conference on Knowledge Discovery & Data Mining (KDD)

Twhin: Embedding the twitter heterogeneous information network for per- sonalized recommendation. In International Conference on Knowledge Discovery & Data Mining (KDD) . 2842–2850
[2024]

In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

Afdgcf: Adaptive feature de-correlation graph collaborative filtering for recommendations. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1242–1252

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Iván Cantador, Peter Brusilovsky, and Tsvi Kuflik. 2011. 2nd Workshop on Information Heterogeneity and Fusion in Recommender Systems (HetRec 2011). In ACM Conference on Recommender Systems (RecSys) (RecSys 2011)

work page 2011

[2] [2]

Qiwei Chen, Huan Zhao, Wei Li, Pipei Huang, and Wenwu Ou. 2019. Behavior se- quence transformer for e-commerce recommendation in alibaba. In International Workshop on Deep Learning Practice for High-dimensional Sparse Data . 1–4

work page 2019

[3] [3]

Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. 2021. Adaptive univer- sal generalized pagerank graph neural network. In International Conference on Learning Representations (ICLR)

work page 2021

[4] [4]

Jeongwhan Choi, Seoyoung Hong, Noseong Park, and Sung-Bae Cho. 2023. Blurring-sharpening process models for collaborative filtering. In ACM Interna- tional Conference on Research and Development in Information Retrieval (SIGIR) . 1096–1106

work page 2023

[5] [5]

Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. Advances in Neural Information Processing Systems (NeurIPS) 29 (2016)

work page 2016

[6] [6]

Ahmed El-Kishky, Thomas Markovich, Serim Park, Chetan Verma, Baekjin Kim, Ramy Eskander, Yury Malkov, Frank Portman, Sofía Samaniego, Ying Xiao, et al

[7] [7]

Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin

[8] [8]

Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Predict then propagate: Graph neural networks meet personalized pagerank. In International Conference on Learning Representations (ICLR)

work page 2018