Pith. sign in

REVIEW 5 major objections 6 minor 42 references

Revisiting Graph Projections for Effective Complementary Product Recommendation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that complementary product recommendations can be obtained by ranking neighbors in a directed, temporally weighted item-item graph projected from user purchase histories, and that this training-free method outperforms…

desk verdict A simple, training-free graph-projection method with genuinely informative ablations, but the headline +43%/+38% margins rest on an evaluation protocol that likely inflates the numbers. read the letter →

arxiv 2506.09209 v1 pith:Z33MISNC submitted 2025-06-10 cs.IR cs.LG

Leandro Anghinoni , Pablo Zivic , Jorge Adrian Sanchez This is my paper
classification cs.IRcs.LG
keywords complementaryproductrecommendationbipartitegraphprojectionitem-itemtemporaldirectionalitycollaborativefilteringsequentialgraph-basedexplainable
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

The paper argues that the noisy, sparse signal in purchase histories is best exploited not by learning embeddings but by explicitly constructing a directed item graph from co-purchase and purchase-order statistics. The method projects the user-item bipartite graph onto items, weights edges by a λ-step random-walk co-occurrence probability, then multiplies in a temporal direction term that records how often and how quickly one item is bought after another, at both item and category level. Recommendation is a local lookup that ranks the weighted neighbors of the query item, with no training and no latent representations. On four Amazon subsets the paper reports average gains of +43% over sequential recommenders and +38% over graph-based recommenders. If the comparisons hold, a nonparametric, explainable baseline beats state-of-the-art neural models on standard complementary-recommendation benchmarks.

What carries the argument

The central object is the directed weighted item graph with adjacency matrix $W = W_V \odot ((1-\alpha)\tilde{C}_I + \alpha\tilde{C}_C)$. $W_V$ is the symmetric one-mode projection of the user-item bipartite graph: the $\lambda$-th power of $P^{V\leftrightarrow V} = Q + Q^T$, where $Q = (1/|V|) P^{V\to U} P^{U\to V}$ encodes two-step random-walk transition probabilities between items through users. The factor $((1-\alpha)\tilde{C}_I + \alpha\tilde{C}_C)$ breaks that symmetry: $\tilde{C}_I$ counts, within $\kappa$ purchase steps, how often item $j$ is bought after item $i$ with weight $1/\delta^{i\to j}$, while $\tilde{C}_C$ does the same at the category level and acts as a sparsity regularizer. The Hadamard product fuses co-purchase strength with directed temporal evidence, and inference is the ranking of the query row of $W$. The only tunable parameters, $(\alpha, \lambda, \kappa)$, are selected on a validation split.

What would settle it

Re-run the baseline models (for instance TIGER, SASRec, BERT4Rec, and UltraGCN) on the exact preprocessed Amazon subsets and split procedures described in the paper, and compare the margins; if the improvements shrink or disappear, the central claim fails. A cheaper check is to re-evaluate the sequential baselines on Amazon Beauty with same-timestamp purchases randomly ordered, as the paper does for its own model, and observe whether the reported gap to the proposed method persists.

Watch

Extended reading notes

Core claim

The central claim is that the projected item-item graph with temporal directionality, defined by $W = W_V \odot ((1-\alpha)\tilde{C}_I + \alpha\tilde{C}_C)$, captures complementarity so well that simple neighborhood search over its weighted edges outperforms trained sequential and graph neural recommenders. The symmetric part $W_V$ is the $\lambda$-th power of the two-step random-walk transition matrix on the bipartite graph, encoding co-purchase strength through shared users. The direction matrices $\tilde{C}_I$ and $\tilde{C}_C$ are row-normalized counts, within a window of $\kappa$ purchase steps, of how often an item (or a category) is bought after the query item, weighted by the inverse step distance, and blended by $\alpha$. The paper reports average improvements of +43% over sequential baselines and +38% over graph-based baselines, with the largest single gain being +54.6% Recall@20 and +26.3% NDCG@20 over UltraGCN on Books.

Load-bearing premise

The claimed margins over baselines rest on the assumption that the published baseline numbers were produced on the same dataset versions, n-core filters, timestamp ordering, and evaluation splits as the new results, since the paper does not re-run the baselines and the default ASIN ordering of same-timestamp purchases is shown to bias a large share of Beauty interactions.

Editorial extensions

If this is right

  • Complementary recommendation can be served by a local graph lookup with no trained embeddings, negative sampling, or GPUs, making it cheap to deploy and update.
  • Restricting temporal evidence to a short window ($\kappa = 2$ to $4$ in the tuned configurations) helps accuracy over models that attend to the entire purchase history.
  • The model is more robust to timestamp-order ambiguity than positional-embedding sequential models: with same-timestamp purchases randomly ordered on Beauty, NDCG@5 drops about 35% for this method versus about 72% for SASRec.
  • Pure cold-start items receive zero recommendations from the graph alone; a title-embedding affinity heuristic restores only partial performance, so cold-start remains an open limitation.
  • The method's explainability is a direct byproduct of using weighted edges: each recommended item is justified by co-purchase strength and by observed 'bought after' counts.

Reading between the lines

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

  • The same projection-plus-direction construction could transfer to other asymmetric item relations, such as substitute detection or bundle recommendation, by redefining what the direction matrix counts.
  • Because the method reduces training to counting and matrix powers, it could serve as a much stronger baseline than random or popularity for future neural CPR papers, potentially changing how gains over baselines are reported.
  • On datasets where same-timestamp purchases dominate, the direction term may be less informative; a follow-up ablation that sets the temporal term to uniform would test how much of the gain is truly temporal rather than co-occurrence.
  • If the reported margins survive a controlled re-run on identical preprocessing, the practical implication is that learned embeddings are not necessary for competitive complementary recommendation on these benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes a nonparametric, training-free method for complementary product recommendation. A directed weighted item-item graph is built by projecting the user-item bipartite graph: co-purchase counts define an undirected affinity matrix (W_V via a random-walk projection and exponentiation), and temporal purchase-order statistics define directionality (item-level co-occurrence C_I and category-level co-occurrence C_C). The final adjacency matrix is the Hadamard product W = W_V ⊙ ((1-α)~C_I + α~C_C). Recommendations are the neighbors of the query item in this graph, ranked by edge weight. The authors evaluate on Amazon datasets (Beauty, Sports, Toys, Books) and report large improvements over sequential recommenders (+43% average, Table I) and graph-based recommenders (+54.6% Recall@20 on Books, Table III). The paper includes ablations on temporal weighting, the random-walk exponent λ, k-core density, and a cold-start extension using title embeddings.

Significance. If the reported results are valid, the paper is significant: it shows that a simple, explainable, hyperparameter-light graph projection can outperform sophisticated neural sequential and GNN baselines by large margins, while requiring no training and no latent embeddings. The method is appealing for industrial settings where explainability and computational efficiency matter. The paper also provides useful ablations (temporal aggregation, high-order walks, k-core sensitivity, same-timestamp noise) that are informative for the CPR community. However, the significance is currently conditional on resolving several evaluation-protocol and comparability concerns; the headline margins are not yet credible as stated.

major comments (5)
  1. [§IV-B, Tables I and III] The baseline numbers are taken from prior papers without re-running, and the paper does not establish that the data versions, n-core filters, timestamp ordering, and split procedures are identical to those used for the proposed method. The paper explicitly says it 'adopt[s] different split procedures' for sequential and graph-based models, and it uses 5-core for sequential and 10-core for graph-based baselines. If the baselines were evaluated on different dataset versions (e.g., the 2014 vs. 2018 Amazon snapshots) or different tie-ordering conventions, the reported +43% and +54.6% margins are not valid. The authors should either re-run all baselines under a single unified protocol or provide concrete evidence (e.g., reproduce published baseline numbers on their exact splits and preprocessing) that the comparison is apples-to-apples.
  2. [§III-B, §III-C, Eq. (8)] The method recommends items by ranking only the neighbors of the query in the projected graph, i.e., items with non-zero entries in the row of W. It is not stated whether non-neighbor items (zero weight) are included in the ranked list during evaluation. If Recall@k and NDCG@k are computed over the restricted neighbor set rather than over the full item catalogue, the reported numbers are inflated relative to baselines that rank all items, because the candidate set is far smaller (typically a few hundred vs. tens of thousands). The authors must specify the exact evaluation procedure, confirm that zero-weight items are included at the bottom of the ranked list, and ideally release code to verify this. Without this clarification, the headline improvements over sequential and graph baselines are not interpretable.
  3. [§IV-D, Eq. (8)] For the graph-based comparison (Books, Table III), the reported configuration is (α, λ, κ) = (1, 1, 4), i.e., α=1. This makes the directionality term purely category-level (~C_C) and discards item-level co-purchase directionality (~C_I). The graph-based baselines (UltraGCN, LightGCN, NGCF, etc.) do not use category information, so the proposed method is given access to an additional signal that the baselines cannot exploit. This is not an invalid modeling choice, but it undermines the claim of a fair comparison and should be analyzed: the authors should include an ablation over α in the graph-based setup and discuss what fraction of the improvement comes solely from category labels rather than from the graph projection itself.
  4. [§IV-C, §IV-F4] The paper finds that 39% of Beauty interactions share the same timestamp with the consecutive purchase and that randomizing the tie order changes NDCG@5 by -35% for the proposed method and -72% for SASRec. It is unclear whether the main experiments in Tables I and III were run on the original ASIN-ordered data or on the shuffled data. If the authors used the shuffled (bias-removed) data for their own method while the baseline numbers in Table I come from papers that used the original order, the comparison is unfair (or at least biased in the authors' favor). The paper must state exactly which preprocessing was used for all reported numbers and should ideally re-run baselines under both settings.
  5. [§IV-C] The hyperparameters (α, λ, κ) are tuned on a validation split, which is standard, but the paper does not report sensitivity to these choices or the variance of the results. For instance, on Beauty the optimal α=1, meaning the directionality is entirely category-based; small changes in α might substantially change the results. Reporting sensitivity curves (or at least the validation performance across the grid) would strengthen the claim that the method is robust and not overfit to a particular hyperparameter setting.
minor comments (6)
  1. [§IV-A] The paper should state which version of the Amazon Product Review dataset (e.g., the 2014 or 2018 snapshot) is used and whether the dataset statistics in Table II correspond to the 5-core or 10-core filtering (or to some other filtering). This is essential context for reproducibility.
  2. [§III-D, footnote 1] There is a notation conflict: the symbol k is used both for the top-k evaluation metric and for the absolute time difference (k=|t_j - t_i|) in the cold-start footnote. Please use a different variable (e.g., Δt) for the time difference.
  3. [Abstract and Section I] The word 'manyfold' should be 'manifold', and 'an one-mode projection' should be 'a one-mode projection'. Minor grammar issues throughout (e.g., 'S˜ao Paulo').
  4. [§IV-F1] The seven aggregation functions are mentioned but not listed; the reader cannot see which functions were compared. Please include a table or a list defining them.
  5. [§IV-F4] The paper reports degradation percentages for the shuffled data but does not report whether the differences are statistically significant across the 5 runs. A paired significance test or confidence intervals would help.
  6. [§IV-G] The qualitative examples are interesting, but the paper does not state whether these examples are representative or selected. A caveat sentence would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the graph-projection ranking is computed directly from co-purchase and temporal co-occurrence counts, with only standard validation-set hyperparameter tuning.

full rationale

No load-bearing circular step is present. The proposed score W = W_V ⊙ ((1−α) C̃_I + α C̃_C) (Eq. 8) is a deterministic function of co-purchase counts, category co-counts, and powers of a two-step transition matrix; it contains no fitted parameter that is defined in terms of the evaluation target, and α, λ, κ are selected by grid search on a validation split and then applied to a held-out test split. The paper's ground-truth definition ('a complementary item is an item bought together or after the query item', Sec. IV-A) intentionally matches the signal used to build the graph, but this is the standard formulation of next-item/co-purchase evaluation rather than a derivation that reduces to its own input. All cited prior work is external; there are no self-citations or imported uniqueness theorems carrying the argument. The principal concerns with the paper are external-validity issues in the benchmark comparison (e.g., whether the neighborhood-search candidate set is comparable to full-corpus baselines), which are correctness/benchmarking risks, not circular reasoning.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The model introduces no new entities. It relies on several domain assumptions about co-purchase meaning complementarity, temporal order indicating direction, and category co-occurrence as a valid regularizer. It has three hyperparameters (α, λ, κ) tuned on validation. The method is nonparametric, so the circularity burden is low.

free parameters (3)
  • alpha (α) = 0.85 (Sports), 1.0 (Beauty), 0.9 (Toys), 1.0 (Books graph-based)
    Convex combination weight between item-level (C_I) and category-level (C_C) temporal co-occurrence, tuned by grid search on validation split (Sec. IV-C).
  • lambda (λ) = 4 (sequential datasets), 1 (Books graph-based)
    Exponent for the random-walk projection matrix W_V, controlling high-order path aggregation; tuned on validation (Sec. IV-F2).
  • kappa (κ) = 2 (Sports), 4 (Beauty, Toys, Books)
    Maximum number of interaction steps between purchases counted as co-purchase; tuned on validation (Sec. IV-C).
assumptions (5)
  • domain assumption Items co-purchased by the same users are complementary
    The entire projection is built on the premise that shared users imply product complementarity rather than substitution or noise (Sec. I, III-C).
  • domain assumption Temporal purchase order reveals the direction of complementarity (A→B)
    The directionality matrix C_I encodes how often B is bought after A within κ steps; this assumes order correlates with complementarity direction (Sec. III-B, III-D).
  • domain assumption Category-level co-occurrence (C_C) is a valid regularizer for item-level sparsity
    The method replaces item-level counts with category-level counts to combat sparsity; on Beauty, α=1 makes this the only directional signal. This assumes categories are a meaningful proxy for complementarity (Sec. III-D).
  • domain assumption The λ-th power of P^{V↔V} gives meaningful item-affinity weights
    W_V = (P^{V↔V})^λ treats the unnormalized symmetric matrix as a weighted adjacency and uses path counts as affinity; this is a heuristic not derived from a probabilistic model (Sec. III-C, Eq. 6).
  • domain assumption Reported baseline numbers are comparable to the authors' results
    The paper relies on prior publications for baseline metrics, assuming identical dataset versions, preprocessing, and splits (Sec. IV-B).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Graph Projections for Effective Complementary Product Recommendation." pith.science (2026). https://pith.science/paper/Z33MISNC

@misc{pith2026250609209,
  author       = {Pith},
  title        = {Pith review of: Revisiting Graph Projections for Effective Complementary Product Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z33MISNC}},
  note         = {Machine review of arXiv:2506.09209}
}
read the original abstract

Complementary product recommendation is a powerful strategy to improve customer experience and retail sales. However, recommending the right product is not a simple task because of the noisy and sparse nature of user-item interactions. In this work, we propose a simple yet effective method to predict a list of complementary products given a query item, based on the structure of a directed weighted graph projected from the user-item bipartite graph. We revisit bipartite graph projections for recommender systems and propose a novel approach for inferring complementarity relationships from historical user-item interactions. We compare our model with recent methods from the literature and show, despite the simplicity of our approach, an average improvement of +43% and +38% over sequential and graph-based recommenders, respectively, over different benchmarks.

Figures

Figures reproduced from arXiv: 2506.09209 by the authors.

Figure 1
Figure 1. Overview of the proposed model. In (a), items [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Ablation studies on Amazon Beauty. In all studies we evaluate NDCG@5. In (a) we show the effect of different [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Performance under permutation of items with the same [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative examples for Our model (on top) vs. SASRec (below). Notice that our model not only ranks higher than the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 41 canonical work pages

  1. [1]

    Complementary recommendation in e-commerce: Definition, approaches, and future directions,

    L. Li and Z. Du, “Complementary recommendation in e-commerce: Definition, approaches, and future directions,” arXiv preprint, 2024

    work page 2024

  2. [2]

    Inferring networks of substitutable and complementary products,

    J. McAuley, R. Pandey, and J. Leskovec, “Inferring networks of substitutable and complementary products,” in Proc. Intl. Conf. on knowledge discovery and data mining, 2015, pp. 785–794

    work page 2015

  3. [3]

    Large scale product graph construction for recommendation in e-commerce,

    X. Yang, Y . Zhu, Y . Zhang, X. Wang, and Q. Yuan, “Large scale product graph construction for recommendation in e-commerce,” arXiv preprint, 2020

    work page 2020

  4. [4]

    Com- plete the look: Scene-based complementary product recommendation,

    W.-C. Kang, E. Kim, J. Leskovec, C. Rosenberg, and J. McAuley, “Com- plete the look: Scene-based complementary product recommendation,” in Proc. the IEEE/CVF Conf. on computer vision and pattern recognition, 2019, pp. 10 532–10 541

    work page 2019

  5. [5]

    Uniform sequence better: Time interval aware data augmentation for sequential recommendation,

    Y . Dang, E. Yang, G. Guo, L. Jiang, X. Wang, X. Xu, Q. Sun, and H. Liu, “Uniform sequence better: Time interval aware data augmentation for sequential recommendation,” in Proc. the AAAI Conf. on artificial intelligence, vol. 37, no. 4, 2023, pp. 4225–4232

    work page 2023

  6. [6]

    Personalized complementary product recommendation,

    A. Yan, C. Dong, Y . Gao, J. Fu, T. Zhao, Y . Sun, and J. McAuley, “Personalized complementary product recommendation,” in Companion Proc. the Web Conf. 2022, 2022, pp. 146–151

    work page 2022

  7. [7]

    Item relationship graph neural networks for e-commerce,

    W. Liu, Y . Zhang, J. Wang, Y . He, J. Caverlee, P. P. Chan, D. S. Yeung, and P.-A. Heng, “Item relationship graph neural networks for e-commerce,” IEEE Tr. on Neural Networks and Learning Systems, vol. 33, no. 9, pp. 4785–4799, 2021

    work page 2021

  8. [8]

    Learning from substitutable and complementary relations for graph-based sequential product recom- mendation,

    W. Zhang, Z. Chen, H. Zha, and J. Wang, “Learning from substitutable and complementary relations for graph-based sequential product recom- mendation,” ACM Tr. on Information Systems (TOIS), vol. 40, no. 2, pp. 1–28, 2021

    work page 2021

Show all 42 references
  1. [9]

    Shopping queries dataset: A large-scale esci benchmark for improving product search,

    C. K. Reddy, L. M `arquez, F. Valero, N. Rao, H. Zaragoza, S. Bandyopad- hyay, A. Biswas, A. Xing, and K. Subbian, “Shopping queries dataset: A large-scale esci benchmark for improving product search,” arXiv preprint, 2022

    2022

  2. [10]

    Bipartite network pro- jection and personal recommendation,

    T. Zhou, J. Ren, M. Medo, and Y .-C. Zhang, “Bipartite network pro- jection and personal recommendation,” Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, vol. 76, no. 4, p. 046115, 2007

    2007

  3. [11]

    Personal recommendation using weighted bipartite graph projection,

    M.-S. Shang, Y . Fu, and D.-B. Chen, “Personal recommendation using weighted bipartite graph projection,” in 2008 Intl. Conf. on Apperceiving Computing and Intelligence Analysis. IEEE, 2008, pp. 198–202

    2008

  4. [12]

    Collaborative filtering using weighted bipartite graph projection: a recommendation system for yelp,

    S. Sawant, “Collaborative filtering using weighted bipartite graph projection: a recommendation system for yelp,” in Proc. the CS224W: Social and information network analysis Conf., vol. 33, 2013

    2013

  5. [13]

    Image-based recommendations on styles and substitutes,

    J. McAuley, C. Targett, Q. Shi, and A. Van Den Hengel, “Image-based recommendations on styles and substitutes,” in Proc. Conf. on research and development in information retrieval, 2015, pp. 43–52

    2015

  6. [14]

    A path-constrained framework for discriminating substitutable and complementary products in e-commerce,

    Z. Wang, Z. Jiang, Z. Ren, J. Tang, and D. Yin, “A path-constrained framework for discriminating substitutable and complementary products in e-commerce,” in Proc. Intl. Conf. on Web Search and Data Mining, 2018, pp. 619–627

    2018

  7. [15]

    Two is better than one: Dual embeddings for complementary product recommendations,

    G. Kvernadze, P. A. G. Sudyanti, N. Subedi, and M. Hajiaghayi, “Two is better than one: Dual embeddings for complementary product recommendations,” in 2022 IEEE Intl. Conf. on Knowledge Graph (ICKG). IEEE, 2022, pp. 131–140

    2022

  8. [16]

    Item2vec: neural item embedding for collaborative filtering,

    O. Barkan and N. Koenigstein, “Item2vec: neural item embedding for collaborative filtering,” in 2016 IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2016, pp. 1–6

    2016

  9. [17]

    E-commerce in your inbox: Product recommendations at scale,

    M. Grbovic, V . Radosavljevic, N. Djuric, N. Bhamidipati, J. Savla, V . Bhagwan, and D. Sharp, “E-commerce in your inbox: Product recommendations at scale,” in Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, 2015, pp. 1809–1818

    2015

  10. [18]

    Self-attentive sequential recommendation,

    W.-C. Kang and J. McAuley, “Self-attentive sequential recommendation,” in 2018 IEEE Intl. Conf. on data mining (ICDM). IEEE, 2018, pp. 197–206

    2018

  11. [19]

    Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer,

    F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang, “Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer,” in Proc. the 28th ACM Intl. Conf. on information and knowledge management, 2019, pp. 1441–1450

    2019

  12. [20]

    P-companion: A principled framework for diversified complementary product recommendation,

    J. Hao, T. Zhao, J. Li, X. L. Dong, C. Faloutsos, Y . Sun, and W. Wang, “P-companion: A principled framework for diversified complementary product recommendation,” in Proc. the 29th ACM Intl. Conf. on Information & Knowledge Management, 2020, pp. 2517–2524

    2020

  13. [21]

    Neural graph collaborative filtering,

    X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph collaborative filtering,” in Proc. Conf. on Research and development in Information Retrieval, 2019, pp. 165–174

    2019

  14. [22]

    Lightgcn: Sim- plifying and powering graph convolution network for recommendation,

    X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “Lightgcn: Sim- plifying and powering graph convolution network for recommendation,” in Proc. Conf. on research and development in Information Retrieval, 2020, pp. 639–648

    2020

  15. [23]

    Ccgen: Explainable complementary concept generation in e-commerce,

    J. Huang, Y . Gao, Z. Li, J. Yang, Y . Song, C. Zhang, Z. Zhu, H. Jiang, K. C.-C. Chang, and B. Yin, “Ccgen: Explainable complementary concept generation in e-commerce,” arXiv preprint, 2023

    2023

  16. [24]

    Efficient high-quality clustering for large bipartite graphs,

    R. Yang and J. Shi, “Efficient high-quality clustering for large bipartite graphs,” Proc. the ACM on Management of Data, vol. 2, no. 1, pp. 1–27, 2024

    2024

  17. [25]

    Recommender systems with generative retrieval,

    S. Rajput, N. Mehta, A. Singh, R. Hulikal Keshavan, T. Vu, L. Heldt, L. Hong, Y . Tay, V . Tran, J. Samost et al., “Recommender systems with generative retrieval,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  18. [26]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5),

    S. Geng, S. Liu, Z. Fu, Y . Ge, and Y . Zhang, “Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5),” in Proc. the 16th ACM Conf. on Recommender Systems, 2022, pp. 299–315

    2022

  19. [27]

    Personalized top-n sequential recommendation via convolutional sequence embedding,

    J. Tang and K. Wang, “Personalized top-n sequential recommendation via convolutional sequence embedding,” in Proc. the eleventh ACM Intl. Conf. on web search and data mining, 2018, pp. 565–573

    2018

  20. [28]

    Hierarchical gating networks for sequential recommendation,

    C. Ma, P. Kang, and X. Liu, “Hierarchical gating networks for sequential recommendation,” in Proc. the 25th ACM SIGKDD Intl. Conf. on knowledge discovery & data mining, 2019, pp. 825–833

    2019

  21. [29]

    Session-based recommendations with recurrent neural networks,

    B. Hidasi, “Session-based recommendations with recurrent neural networks,” arXiv preprint, 2015

    2015

  22. [30]

    Feature-level deeper self-attention network for sequential recommendation

    T. Zhang, P. Zhao, Y . Liu, V . S. Sheng, J. Xu, D. Wang, G. Liu, X. Zhou et al., “Feature-level deeper self-attention network for sequential recommendation.” in IJCAI, 2019, pp. 4320–4326

    2019

  23. [31]

    S3-rec: Self-supervised learning for sequential recom- mendation with mutual information maximization,

    K. Zhou, H. Wang, W. X. Zhao, Y . Zhu, S. Wang, F. Zhang, Z. Wang, and J.-R. Wen, “S3-rec: Self-supervised learning for sequential recom- mendation with mutual information maximization,” in Proc. the 29th ACM Intl. Conf. on information & knowledge management, 2020, pp. 1893–1902

    2020

  24. [32]

    Ultragcn: ultra simplification of graph convolutional networks for recommendation,

    K. Mao, J. Zhu, X. Xiao, B. Lu, Z. Wang, and X. He, “Ultragcn: ultra simplification of graph convolutional networks for recommendation,” in Proc. the 30th ACM Intl. Conf. on information & knowledge management, 2021, pp. 1253–1262

    2021

  25. [33]

    Matrix factorization techniques for recommender systems,

    Y . Koren, R. Bell, and C. V olinsky, “Matrix factorization techniques for recommender systems,” Computer, vol. 42, no. 8, pp. 30–37, 2009

    2009

  26. [34]

    Collaborative metric learning,

    C.-K. Hsieh, L. Yang, Y . Cui, T.-Y . Lin, S. Belongie, and D. Estrin, “Collaborative metric learning,” in Proc. the 26th Intl. Conf. on world wide web, 2017, pp. 193–201

    2017

  27. [35]

    Efficient neural matrix factorization without sampling for recommendation,

    C. Chen, M. Zhang, Y . Zhang, Y . Liu, and S. Ma, “Efficient neural matrix factorization without sampling for recommendation,” ACM Tr. on Information Systems (TOIS), vol. 38, no. 2, pp. 1–28, 2020

    2020

  28. [36]

    Deepwalk: Online learning of social representations,

    B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk: Online learning of social representations,” in Proc. the 20th ACM SIGKDD Intl. Conf. on Knowledge discovery and data mining, 2014, pp. 701–710

    2014

  29. [37]

    Line: Large- scale information network embedding,

    J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and Q. Mei, “Line: Large- scale information network embedding,” in Proc. the 24th Intl. Conf. on world wide web, 2015, pp. 1067–1077

    2015

  30. [38]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” in Proc. the 22nd ACM SIGKDD Intl. Conf. on Knowledge discovery and data mining, 2016, pp. 855–864

    2016

  31. [39]

    Neighbor interaction aware graph convolution networks for recommendation,

    J. Sun, Y . Zhang, W. Guo, H. Guo, R. Tang, X. He, C. Ma, and M. Coates, “Neighbor interaction aware graph convolution networks for recommendation,” in Proc. the 43rd Intl. ACM SIGIR Conf. on research and development in information retrieval, 2020, pp. 1289–1298

    2020

  32. [40]

    Revisiting graph based collaborative filtering: A linear residual graph convolutional network approach,

    L. Chen, L. Wu, R. Hong, K. Zhang, and M. Wang, “Revisiting graph based collaborative filtering: A linear residual graph convolutional network approach,” in Proc. the AAAI Conf. on artificial intelligence, vol. 34, no. 01, 2020, pp. 27–34

    2020

  33. [41]

    Disentangled graph collaborative filtering,

    X. Wang, H. Jin, A. Zhang, X. He, T. Xu, and T.-S. Chua, “Disentangled graph collaborative filtering,” in Proc. the 43rd Intl. ACM SIGIR Conf. on research and development in information retrieval, 2020, pp. 1001– 1010

    2020

  34. [42]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” in Proc. the 2019 Conf. on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019

    2019

Pith tools

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