Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Non-parametric Graph Convolution for Re-ranking in Recommendation Systems

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By aggregating a frozen ranking model's scores on nearby user-item pairs, graph structure can be exploited at test time to improve ranking quality by 8.1% on average with negligible overhead.

desk verdict A clean test-time re-ranking idea with consistent gains, but the paper never states whether the similarity graph is built from training interactions only, leaving the headline 8.1% open to a label-leakage artifact. read the letter →

arxiv 2507.09969 v1 pith:AFFHCHEJ submitted 2025-07-14 cs.IR

Zhongyu Ouyang , Mingxuan Ju , Soroush Vosoughi , Yanfang Ye This is my paper · ORCID
classification cs.IR
keywords recommendersystemsre-rankingtest-timeaugmentationgraphconvolutionnon-parametriccollaborativefilteringCTRprediction
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 graph structure can improve the ranking stage of a recommender system without paying the cost of graph-based training. The proposed non-parametric strategy operates at test time only: given a query user-item pair, it builds user-user and item-item similarity matrices from the interaction graph, retrieves the most similar users and items, forms all pairwise combinations, scores those pairs with the already-trained ranking model, and aggregates the scores with similarity-based weights. Across four datasets and seven ranking models, this lifted average ranking metrics by 8.1% while adding only about 0.5% inference overhead, compared with a naive graph-enhanced encoder that cost roughly 480-600% more. The practical claim is that a frozen deployed model can be improved with minimal engineering and no retraining.

What carries the argument

The mechanism is a non-parametric graph convolution performed at re-ranking time. It consists of four steps: (1) build normalized user-user and item-item similarity matrices from the interaction matrix M (co-occurrence counts normalized by node degrees); (2) retrieve the n_k most similar users and items for the query pair; (3) take the Cartesian product of these retrieved users and items, weighting each constructed pair as the product of its user and item similarity scores; (4) query the frozen ranking model on each constructed pair and aggregate the resulting scores, modifying the weight of the most-similar pair via the pairwise aggregation rule in Equation (12). This replaces parametric message passing during training with a one-time, dataset-level computation plus per-query aggregation.

What would settle it

Run the released code on a dataset whose context features are recorded only for observed user-item interactions, and set n_k > 1 so that the retrieved similar users and items almost certainly include pairs absent from the interaction table; if the re-ranking step errors, falls back to a placeholder, or silently uses zero vectors for those pairs' context features, then the average 8.1% improvement is not produced by the described algorithm without an undeclared imputation step.

Watch

Extended reading notes

Core claim

The central discovery is that the benefit of graph convolution in ranking does not require training with message passing: it can be reproduced at inference time as a non-parametric aggregation. Formally, for a query user i and item j, the method computes degree-normalized co-interaction similarities within users (A_u = $D_u^{{-1/2}}$ M M^T $D_u^{{-1/2}}$) and within items (A_i = $D_i^{{-1/2}}$ M^T M $D_i^{{-1/2}}$), retrieves the top n_k users and items, constructs $n_k^{2}$ cross pairs, feeds each pair into the frozen model (z_{uv} = [f(x_u) || f(x_v) || h(c_uv)]) to get r(z_uv), and computes the re-rank score as a weighted average, with the dominant pair's weight replaced by the sum of complementary weights. This captures most of the ranking improvement of a trained graph encoder at a fraction of the cost.

Load-bearing premise

The method's third step builds $n_k^{2}$ user-item pairs and feeds each into the frozen model, but the context features for pairs that were never observed together are not defined anywhere in the paper; if a deployed system only has context features for real interactions, the procedure cannot be executed as written.

Editorial extensions

If this is right

  • Deployed CTR models can be upgraded without retraining: the module plugs into any trained ranker and only needs the interaction matrix plus a similarity-matrix precomputation.
  • Sparse datasets stand to gain more, because test-time graph signals compensate for weak collaborative signals in the trained ID embeddings.
  • Inference cost scales quadratically with n_k, and larger n_k helps sparse graphs while small n_k suffices for dense graphs; the paper reports the overhead stays near 0.5% for the selected settings.
  • The naive graph-encoder alternative (message passing all through training) is empirically stronger in places but costs roughly 480-600% more training time and over 600% more testing time, making the test-time route the only practical one at industrial scale.

Reading between the lines

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

  • Because the paper never defines the context feature vector c_uv for user-item pairs that do not exist in the interaction data, applying Step 3 as written would require an imputation policy; the reported 8.1% average may therefore depend on an unspecified convention.
  • In a live recommender, user-item interactions accrue over time, so the similarity matrices and retrieved neighborhoods would need periodic recomputation; the 0.5% overhead figure assumes a static snapshot and would grow with refresh frequency.
  • The aggregation rule implicitly assumes that similar users interact with similar items in a way that is informative for the query pair; datasets where context strongly overrides collaborative similarity would likely see smaller or negative gains, a boundary the paper does not test.
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 / 4 minor

Summary. The paper proposes a non-parametric graph-convolution strategy for re-ranking in recommender systems. Given a frozen ranking/CTR model, it constructs user-user and item-item similarity matrices from the interaction matrix, retrieves the top-n_k neighbors for each query user and item, forms n_k^2 user-item pairs, queries the frozen model on each pair, and aggregates the resulting scores with a weighted fusion. The method is intended as a plug-and-play test-time module that avoids graph-based training. Experiments on four datasets (ML-1M, Yelp2018, Amazon-Books, Anime) and seven baselines report an average 8.1% relative improvement in NDCG@10/20 and Recall@10/20, with an average 0.5% additional computational overhead.

Significance. If the reported gains are free of label leakage, the paper offers a simple and broadly applicable test-time augmentation for ranking models, with the practical virtue that no retraining or graph infrastructure is needed during training. The paper provides a code repository, evaluates across seven diverse baselines and four datasets, and includes a proof-of-concept comparison of table- versus graph-based encoders. The central mechanism is method-agnostic and could generalize to other scoring functions. However, the three concerns below—whether the similarity graph is built only from training data, how context features are defined for unobserved pairs, and the inconsistency in the overhead claim—are load-bearing for the validity of the stated 8.1%/0.5% numbers.

major comments (3)
  1. [2.4, Eq. (9); 3.1.1; 3.1.3] The paper does not state that the interaction matrix M used to construct A_u and A_i is restricted to the training split. In Section 3.1.1 the data are randomly split into 0.8/0.1/0.1 train/validation/test, and Section 3.1.3 says only that the similarity matrices are pre-computed on a CPU. If M is the full binarized interaction matrix, then for a held-out test pair (i,j), the test label contributes to both MM^T and M^TM, so the top-n_k neighbor sets in Eq. (9) are partly selected by the very interaction being predicted. The reported 8.1% average improvement would then reflect label leakage rather than a deployable test-time gain. The authors must state explicitly that M_train is used and confirm this in the released code.
  2. [2.4, Eq. (11)] The re-ranking step queries the frozen model on n_k^2 user-item pairs (u,v) in M_ij, but the paper never defines the context vector c_uv for pairs that do not appear in the interaction data. In Section 2.1, c_ij is introduced as the context features of the interaction between user i and item j; for a constructed pair that has no observed interaction, no such vector is well-defined. Without an explicit convention (e.g., zero vector, user/item profile concatenation, or exclusion of context), the procedure in Eq. (11) is not executable. This must be specified and reflected in the experiments.
  3. [3.3, abstract, conclusion] The central efficiency claim of '0.5% on average' additional overhead is inconsistent with the algorithmic complexity described in the paper. Section 3.3 states that the strategy 'quadratically increases the inference time wrt n_k', meaning up to n_k^2 extra model forward passes per test pair. The abstract and conclusion claim 0.5% average overhead while Figure 3 and the text report 'less than 2% of extra time overall'; these numbers are not reconciled. The authors should report the empirical wall-clock overhead per test pair, specify whether the 0.5% is averaged over datasets with different n_k, and clarify why n_k^2 extra queries cost so little relative to the original forward pass.
minor comments (4)
  1. [2.4, Eq. (12)] There is a typo in the equation label ('Wegiht'), and the notation A_u/i is introduced without a compact definition before its first use.
  2. [3.1.3] The setup would benefit from stating which context features are used in each dataset; without this, the reader cannot verify the c_uv convention required by Eq. (11).
  3. [Figure 3] The y-axis labels ('Relative Overall Time', etc.) are not defined; please indicate the baseline and whether lower values are better.
  4. [Table 2] The %Δ column for NFM and xDeepFM on Anime shows small negative deltas; the text acknowledges only 'some rare cases' and should give the exact numbers for transparency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the re-ranking score is an explicit aggregation of the frozen model's outputs, with no fitted parameter or self-citation chain doing the work.

full rationale

The paper's derivation is self-contained. The non-parametric re-ranking procedure (Eqs. 9-12) is an inference-time aggregation: after a ranking model r is trained, the method forms similarity matrices A_u = D_u^{-1/2} M M^T D_u^{-1/2} and A_i = D_i^{-1/2} M^T M D_i^{-1/2}, retrieves top-n_k neighbors, constructs n_k^2 pairs, and computes p'_ij as a normalized weighted sum of r evaluated on those pairs. No parameter of the method is fit to the target ranking labels except n_k, which is selected on the validation split (Sec. 3.1.4); this is standard hyperparameter selection, not fitting the reported test-time prediction. The equation for p'_ij is not defined in terms of itself, and no fitted value is renamed as a prediction. The graph-convolution equations (Eqs. 7-8) are taken from external LightGCN [12], and the similarity matrices are standard normalized co-occurrence; the paper does not lean on a self-citation to justify a uniqueness claim or to smuggle in an ansatz. The self-citations in Related Work (e.g., [25], [27], [28]) are contextual and not load-bearing. The possible concern that M in Eq. (9) may include test interactions is a data-leakage/correctness issue, not a circularity of the derivation, and cannot be substantiated from the text alone: Sec. 2.2 defines M as the full observed interaction matrix and Sec. 3.1.1 describes a random split, but the paper does not state explicitly which split generates M. Because no claimed result reduces by construction to an input or to a self-citation, the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method introduces no new entities. It relies on the training interaction graph being available at serving time, on being able to score non-interaction pairs with the frozen model, and on a hand-designed aggregation rule. The only tuned scalar is n_k.

free parameters (1)
  • n_k (number of similar users/items) = not reported in text, tuned per dataset/model on validation
    The number of neighbors used to construct user-item pairs; controls the extent of graph knowledge injected and the inference overhead (quadratic in n_k).
assumptions (4)
  • domain assumption The full interaction matrix M (training interactions) is available at test time to construct user-user and item-item similarity matrices.
    Required by Eq. (9); industrial systems may not expose the full interaction graph to the serving stack.
  • ad hoc to paper The frozen ranking model can be queried on arbitrary pairs (u,v), including pairs with no observed interaction, and a context vector c_uv exists for each such pair.
    Step 3 forms n_k^2 pairs that are not necessarily in the training data; Eq. (11) uses c_uv without defining it for non-interactions.
  • domain assumption Co-occurrence similarity (normalized M M^T) captures structural knowledge that is complementary to the trained model's predictions.
    This is the standard collaborative filtering assumption; the paper validates it empirically but does not prove it.
  • ad hoc to paper The weighted aggregation rule in Eq. (10)-(12), including the modified max weight, is a sensible combination of neighbor predictions.
    The specific pairwise weight modification is a heuristic chosen to upweight the original pair; no theoretical justification is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Non-parametric Graph Convolution for Re-ranking in Recommendation Systems." pith.science (2026). https://pith.science/paper/AFFHCHEJ

@misc{pith2026250709969,
  author       = {Pith},
  title        = {Pith review of: Non-parametric Graph Convolution for Re-ranking in Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFFHCHEJ}},
  note         = {Machine review of arXiv:2507.09969}
}
read the original abstract

Graph knowledge has been proven effective in enhancing item rankings in recommender systems (RecSys), particularly during the retrieval stage. However, its application in the ranking stage, especially when richer contextual information in user-item interactions is available, remains underexplored. A major challenge lies in the substantial computational cost associated with repeatedly retrieving neighborhood information from billions of items stored in distributed systems. This resource-intensive requirement makes it difficult to scale graph-based methods in practical RecSys. To bridge this gap, we first demonstrate that incorporating graphs in the ranking stage improves ranking qualities. Notably, while the improvement is evident, we show that the substantial computational overheads entailed by graphs are prohibitively expensive for real-world recommendations. In light of this, we propose a non-parametric strategy that utilizes graph convolution for re-ranking only during test time. Our strategy circumvents the notorious computational overheads from graph convolution during training, and utilizes structural knowledge hidden in graphs on-the-fly during testing. It can be used as a plug-and-play module and easily employed to enhance the ranking ability of various ranking layers of a real-world RecSys with significantly reduced computational overhead. Through comprehensive experiments across four benchmark datasets with varying levels of sparsity, we demonstrate that our strategy yields noticeable improvements (i.e., 8.1% on average) during testing time with little to no additional computational overheads (i.e., 0.5 on average). Code: https://github.com/zyouyang/RecSys2025_NonParamGC.git

Figures

Figures reproduced from arXiv: 2507.09969 by the authors.

Figure 1
Figure 1. (a) The table encoder for user/item ID embeddings; (b) The graph-enhanced user/item ID encoder; (c) The calculation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An example of calculating the weights for aggregat [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The relative performance is averaged over the four [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The ranking performance of DCN to varied number [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Generative Recommendation with Semantic IDs: A Practitioner's Handbook

    cs.IR 2025-07 conditional novelty 6.0 of 10

    An open-source framework and ablation study showing which semantic-ID generative recommendation components actually matter, with results that challenge several standard defaults.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages · cited by 1 Pith paper

  1. [1]

    Rianne van den Berg, Thomas N Kipf, and Max Welling. 2018. Graph convolu- tional matrix completion. In ACM SIGKDD Conference

    work page 2018

  2. [2]

    Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. 2022. LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation. In International Conference on Learning Representations

    work page 2022

  3. [3]

    Chong Chen, Min Zhang, Yongfeng Zhang, Yiqun Liu, and Shaoping Ma. 2020. Efficient neural matrix factorization without sampling for recommendation.ACM Transactions on Information Systems (2020)

    work page 2020

  4. [4]

    Cooper Union. 2023. Anime Recommendations Database. https://www.kaggle. com/datasets/CooperUnion/anime-recommendations-database. Accessed on: 2023-11-11

    work page 2023

  5. [5]

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

  6. [6]

    Carlos A Gomez-Uribe and Neil Hunt. 2015. The netflix recommender system: Algorithms, business value, and innovation. ACM Transactions on Management Information Systems (2015)

    work page 2015

  7. [7]

    GroupLens Research. 2023. MovieLens 1M Dataset. https://grouplens.org/ datasets/movielens/1m/. Accessed: 2023-11-11

    work page 2023

  8. [8]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. In International Joint Conference on Artificial Intelligence

    work page 2017

Show all 53 references
  1. [9]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems

    2017

  2. [10]

    Xiaotian Han, Tong Zhao, Yozen Liu, Xia Hu, and Neil Shah. 2023. MLPInit: Embarrassingly Simple GNN Training Acceleration with MLP Initialization. In International Conference on Learning Representations

    2023

  3. [11]

    Xiangnan He and Tat-Seng Chua. 2017. Neural factorization machines for sparse predictive analytics. In ACM SIGIR Conference

    2017

  4. [12]

    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 SIGIR Conference

    2020

  5. [13]

    Mohsen Jamali and Martin Ester. 2010. A matrix factorization technique with trust propagation for recommendation in social networks. In ACM Recommender Systems conference

    2010

  6. [14]

    Clark Mingxuan Ju, Leonardo Neves, Bhuvesh Kumar, Liam Collins, Tong Zhao, Yuwei Qiu, Qing Dou, Sohail Nizam, Sen Yang, and Neil Shah. 2025. Revisiting Self-attention for Cross-domain Sequential Recommendation. In ACM SIGKDD Conference

    2025

  7. [15]

    Clark Mingxuan Ju, Leonardo Neves, Bhuvesh Kumar, Liam Collins, Tong Zhao, Yuwei Qiu, Qing Dou, Yang Zhou, Sohail Nizam, Rengim Ozturk, et al . 2025. Learning Universal User Representations Leveraging Cross-domain User Intent at Snapchat. In ACM SIGIR Conference

    2025

  8. [16]

    Mingxuan Ju, William Shiao, Zhichun Guo, Yanfang Ye, Yozen Liu, Neil Shah, and Tong Zhao. 2024. How Does Message Passing Improve Collaborative Filtering?. In Advances in Neural Information Processing Systems

    2024

  9. [17]

    Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations

    2017

  10. [18]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems. IEEE Computer (2009)

    2009

  11. [19]

    Yehuda Koren, Steffen Rendle, and Robert Bell. 2021. Advances in collaborative filtering. Recommender systems handbook (2021)

    2021

  12. [20]

    Dongha Lee, SeongKu Kang, Hyunjun Ju, Chanyoung Park, and Hwanjo Yu. 2021. Bootstrapping user and item representations for one-class collaborative filtering. In ACM SIGIR Conference

    2021

  13. [21]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature inter- actions for recommender systems. In ACM SIGKDD Conference

    2018

  14. [22]

    Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving graph collaborative filtering with neighborhood-enriched contrastive learning. In ACM Web Conference

    2022

  15. [23]

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

    2008

  16. [24]

    Yunshan Ma, Yingzhi He, An Zhang, Xiang Wang, and Tat-Seng Chua. 2022. CrossCBR: cross-view contrastive learning for bundle recommendation. In ACM SIGKDD Conference

    2022

  17. [25]

    Zhongyu Ouyang, Qianlong Wen, Chunhui Zhang, Yanfang Ye, and Soroush Vosoughi. 2025. Towards Human-like Preference Profiling in Sequential Recom- mendation. arXiv:2506.02261 [cs.IR] https://arxiv.org/abs/2506.02261

    2025 arXiv

  18. [26]

    Zhongyu Ouyang, Chunhui Zhang, Shifu Hou, Shang Ma, Chaoran Chen, Toby Li, Xusheng Xiao, Chuxu Zhang, and Yanfang Ye. 2024. Symbolic Prompt Tuning Completes the App Promotion Graph. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases

    2024

  19. [27]

    Zhongyu Ouyang, Chunhui Zhang, Shifu Hou, Chuxu Zhang, and Yanfang Ye

  20. [28]

    Zhongyu Ouyang, Chunhui Zhang, Yaning Jia, and Soroush Vosoughi. 2025. Scaled supervision is an implicit lipschitz regularizer. In Proceedings of the Inter- national AAAI Conference on Web and Social Media

    2025

  21. [29]

    Steffen Rendle. 2010. Factorization machines. In IEEE ICDM

    2010

  22. [30]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  23. [31]

    J Ben Schafer, Joseph Konstan, and John Riedl. 1999. Recommender systems in e-commerce. In ACM conference on Electronic commerce

    1999

  24. [32]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. In ACM International Conference on Information and Knowledge Management

    2019

  25. [33]

    Yi Tay, Luu Anh Tuan, and Siu Cheung Hui. 2018. Latent relational metric learning via memory-based attention for collaborative ranking. In ACM Web Conference

    2018

  26. [34]

    Zhen Tian, Ting Bai, Wayne Xin Zhao, Ji-Rong Wen, and Zhao Cao. 2023. Euler- Net: Adaptive Feature Interaction Learning via Euler’s Formula for CTR Predic- tion. In ACM SIGIR Conference

    2023

  27. [35]

    Aaron Van den Oord, Sander Dieleman, and Benjamin Schrauwen. 2013. Deep content-based music recommendation. In Advances in Neural Information Pro- cessing Systems

    2013

  28. [36]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. In International Confer- ence on Learning Representations

    2017

  29. [37]

    Chenyang Wang, Yuanqing Yu, Weizhi Ma, Min Zhang, Chong Chen, Yiqun Liu, and Shaoping Ma. 2022. Towards Representation Alignment and Uniformity in Collaborative Filtering. In ACM SIGKDD Conference

    2022

  30. [38]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions

    2017

  31. [39]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. In ACM Web Conference

    2021

  32. [40]

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

    2019

  33. [41]

    Yinwei Wei, Wenqi Liu, Fan Liu, Xiang Wang, Liqiang Nie, and Tat-Seng Chua

  34. [42]

    Jiancan Wu, Xiangnan He, Xiang Wang, Qifan Wang, Weijian Chen, Jianxun Lian, and Xing Xie. 2022. Graph convolution machine for context-aware recommender system. Frontiers of Computer Science (2022)

    2022

  35. [43]

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised graph learning for recommendation. In ACM SIGIR Conference

    2021

  36. [44]

    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 ACM SIGKDD Conference

    2018

  37. [45]

    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 ACM SIGIR Conference

    2022

  38. [46]

    An Zhang, Leheng Sheng, Zhibo Cai, Xiang Wang, and Tat-Seng Chua. 2023. Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss. In Advances in Neural Information Processing Systems

    2023

  39. [47]

    Honglei Zhang, Fangyuan Luo, Jun Wu, Xiangnan He, and Yidong Li. 2023. LightFR: Lightweight federated recommendation with privacy-preserving matrix factorization. ACM Transactions on Information Systems (2023)

    2023

  40. [48]

    Shichang Zhang, Yozen Liu, Yizhou Sun, and Neil Shah. 2022. Graph-less Neural Networks: Teaching Old MLPs New Tricks Via Distillation. In International Conference on Learning Representations

    2022

  41. [49]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, et al. 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. In ACM International Conference on Informati...

    2021

  42. [2009]

    In The Confer- ence on Uncertainty in Artificial Intelligence

    BPR: Bayesian personalized ranking from implicit feedback. In The Confer- ence on Uncertainty in Artificial Intelligence

  43. [2019]

    InACM Web Conference

    Graph neural networks for social recommendation. InACM Web Conference

  44. [2023]

    In ACM SIGIR Conference

    Lightgt: A light graph transformer for multimedia recommendation. In ACM SIGIR Conference

  45. [2024]

    How to improve representation alignment and uniformity in graph-based collaborative filtering?. In Proceedings of the International AAAI Conference on RecSys ’25, September 22–26, 2025, Prague, Czech Republic Zhongyu Ouyang, Mingxuan Ju, Soroush Vosoughi, and Yanfang Ye Web an...

    2025

Pith tools

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