Pith. sign in

REVIEW 3 major objections 6 minor 40 references

KGIF: Optimizing Relation-Aware Recommendations with Knowledge Graph Information Fusion

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read This paper claims that explicitly fusing relation embeddings into entity embeddings via dynamic projection vectors improves top-20 recommendation accuracy on three collaborative knowledge graph benchmarks.

desk verdict A small, clearly explained fusion trick in a KGAT-style pipeline, undermined by a tuning-unfair comparison and margins too thin for seed noise. read the letter →

arxiv 2501.04161 v1 pith:RRHZMJRC submitted 2025-01-07 cs.LG cs.IR

classification cs.LGcs.IR
keywords recommendersystemsknowledgegraphinformationfusionattentionnetworkcollaborativefilteringrelationembeddingTransDtop-Krecommendation
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 proposes KGIF, a recommender system built on a collaborative knowledge graph, and claims that explicitly fusing relation information into entity embeddings improves top-20 recommendation accuracy. The core idea is to replace the implicit message passing used by prior graph-attention recommenders with an element-wise fusion of TransD-style projected entity embeddings and relation embeddings, followed by a nonlinear reparameterization and an attentive propagation layer. On Amazon-book, Last-FM, and Yelp2018, KGIF reports Recall@20 and NDCG@20 numbers above eight baselines, including strongly tuned graph and self-supervised models. If the claim holds, it says that relation-aware fusion, not just more propagation or contrastive regularization, is the missing ingredient in knowledge-graph recommenders. The reported gains are small, with the largest NDCG margin 0.0029 over KGRec on Last-FM and the smallest Recall margin 0.0002 over KGCL on Yelp2018.

What carries the argument

At the center is the fusion step $h^* = \mathrm{ReLU}((M^h_r h \odot r) W_1 + b)$ and its tail analogue, where $M^h_r$ is the relation-dependent projection matrix constructed as $r_p h_p^\top + I$ in the TransD style and $\odot$ is the Hadamard product. This operation is what makes the fusion explicit: the relation vector $r$ directly modulates every coordinate of the projected entity embedding before any attention or propagation occurs. The same projection matrices are reused in the attention scoring function $\delta(h,r,t) = (M^t_r t^*)^\top \tanh(M^h_r h^* + r)$, which ties the propagation layer back to the same relation geometry. The rest of the architecture, including ego-network softmax attention, stacked propagation, and Bi-interaction aggregation, follows the knowledge-graph attention line of models, so the fusion operation is the component that carries the paper's claimed improvement.

What would settle it

Re-run the three benchmarks giving KGRec, KGIN, and KGCL the same grid-search budget and multiple random seeds as KGIF, and check whether KGIF's reported leads, up to 0.0029 NDCG@20 on Last-FM and as small as 0.0002 Recall@20 on Yelp2018, persist; if any dataset flips, the claim that KGIF consistently outperforms all baselines is false.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the way side information enters a collaborative knowledge graph matters: implicitly attending over item attributes is not the same as deliberately aligning each entity embedding with the relation that connects it. KGIF first embeds the CKG with TransD, which builds relation-dependent projection matrices from head and tail projection vectors; it then fuses each projected entity embedding with the relation vector by Hadamard product and pushes the result through a ReLU reparameterization. The fused embeddings are propagated through stacked self-attention layers and aggregated with a Bi-interaction aggregator, and the final user-item score is the inner product of concatenated layer embeddings. The paper reports that this pipeline beats FM, NFM, RippleNet, GC-MC, KGNN-LS, KGAT, CKAN, KGCL, KGIN, and KGRec on both Recall@20 and NDCG@20 across all three datasets, with ablations attributing the gain to multiplicative fusion under dynamic projection vectors.

Load-bearing premise

The load-bearing premise is that the comparison in Table II is tuning-fair: baseline models keep their original published hyperparameters while KGIF's hyperparameters are chosen by grid search on the validation set, and the reported advantages are small enough that a differently tuned baseline or a different random seed could erase them.

Editorial extensions

If this is right

  • If KGIF is right, knowledge-graph recommenders should treat relation embeddings as first-class signal at the representation level, not just as edge labels for attention.
  • The ablation results imply that multiplicative fusion without shared weights adds consistent value over additive, concatenative, and shared-weight variants on all three datasets.
  • Depth of four propagation layers remains beneficial where competing models degrade, suggesting that explicit fusion partially counteracts oversmoothing.
  • The attention-score case study offers a concrete path to explainable recommendations: a recommendation for Stephen King's IT is justified by aggregated attention along author and co-item paths.
  • On sparse graphs like Yelp2018 the model still reports the best Recall and NDCG, which the paper reads as robustness for sparse collaborative signals.

Reading between the lines

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

  • A fair-tuning test would be the natural next experiment: giving KGRec and KGIN the same grid-search budget and multiple seeds could narrow the reported margins, since the largest lead is 0.0029 NDCG and the smallest is 0.0002 Recall.
  • The fusion trick is orthogonal to contrastive self-supervision, so combining explicit relation fusion with KGCL-style augmentation may yield further gains.
  • Because the projection vectors are relation-specific, the same mechanism could transfer to other CKG tasks, such as knowledge-graph link prediction, where distinguishing relation contexts is also central.
  • One could test whether simpler bilinear or complex multiplication also recovers the gains, isolating what the ReLU reparameterization actually adds.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes KGIF, a knowledge-graph-aware recommender that (a) initializes collaborative knowledge graph embeddings with TransD, (b) explicitly fuses entity and relation embeddings through dynamic projection vectors followed by a ReLU nonlinearity, (c) propagates the fused embeddings through self-attention and bi-interaction aggregation, and (d) trains with a combined TransD and BPR objective. The evaluation compares KGIF with eight baselines on Amazon-book, Last-FM, and Yelp2018, and includes ablations over fusion type, projection-vector choice, and propagation depth. The central claim is that KGIF consistently outperforms all baselines in Recall@20 and NDCG@20.

Significance. If the positive empirical result can be established, KGIF offers a simple, reusable fusion mechanism that is a plausible improvement over implicit side-information integration in CKG recommenders. The paper is transparent about implementation and datasets, and the ablation study directly interrogates the proposed fusion operation. However, the current evidence for the central claim is not convincing: the reported margins over the best baseline are at most 0.0029 NDCG (Last-FM) and as small as 0.0002 Recall (Yelp2018), and the comparison is single-run with an asymmetric tuning protocol. Since the value of the paper rests primarily on this empirical claim, the evaluation must be hardened before the result can be considered established.

major comments (3)
  1. [Section IV-E, Table II] The central claim that KGIF consistently outperforms all baseline models is supported only by single-run numbers, with no standard deviations, confidence intervals, or significance tests. The margins are extremely small: on Yelp2018 Recall@20 KGIF is 0.0697 versus KGCL's 0.0695 (difference 0.0002), and on Last-FM NDCG@20 KGIF is 0.1424 versus KGIN's 0.1395 (difference 0.0029). Such differences are plausibly within seed-level noise for recommender models. Please report mean and standard deviation over at least five random seeds for each model and dataset, and include a paired significance test (e.g., Wilcoxon signed-rank) for the comparison of KGIF against the best baseline. The same variance reporting is needed for Tables IV and V if those ablations are used to justify the design choices.
  2. [Section IV-D] The tuning protocol is not described consistently and is not tuning-fair as stated. The text first says 'All models were configured with an embedding size of 64, initialized using the Xavier method. Optimization was performed with the Adam optimizer, using a learning rate of 0.0001 and a batch size of 1024,' and then says 'For baseline models, the original hyperparameter settings were retained.' It is unclear which settings the baselines actually received. Since KGIF's L2 regularization and dropout are selected by grid search on the validation set, while baseline hyperparameters are taken from their original papers, the comparison is asymmetric. With margins below 0.003, a baseline re-tuned on the same 7:2:1 splits with the same validation procedure could plausibly outperform KGIF. Please re-run all baselines under the same data splits with per-model validation-based hyperparameter selection, or provide a justification that the retained settings are optimal for these splits.
  3. [Section III-B and III-D, Eq. (8)] The dimensional notation is inconsistent. The paper defines entity embeddings h,t in R^m and relation embeddings r in R^n, with projection vectors h_p,t_p in R^m and r_p in R^n, so the projection matrices M_h^r and M_t^r in Eq. (2) are n by m. After the fusion layer in Eqs. (6)-(7), the vectors h* and t* are n-dimensional. Eq. (8) then applies the n by m matrices M_h^r and M_t^r to h* and t*, which is only well-defined if m = n. The experiments use embedding size 64, so m = n = 64 may be intended, but this equality should be stated explicitly in the formulation and the identity matrix should be written as I_m rather than the nonstandard I_{m×n}.
minor comments (6)
  1. [Section I, Contributions] The bullet 'Outperforming the state-of-the-art (SOTA) by a substantial margin' is difficult to reconcile with the margins in Table II, which are at most 0.0029 NDCG and 0.0005 Recall above the best baseline; please temper the wording or tie it to the multi-seed results requested above.
  2. [Section IV-H, Table V] On Yelp2018, Recall@20 at five layers (0.0698) is higher than at four layers (0.0697), so the statement that performance degrades beyond four layers is not fully supported by the table.
  3. [Figures 5 and 6] Figures 5 and 6 lack axis labels and legends, making it difficult to interpret the loss/recall curves and the comparison of embedding methods; please add explicit captions and error bars if available.
  4. [Table IV] The meaning of '#' in the first column and the role of the 'SW' column are not defined; please add a caption explaining shared weights and the marker used for the without-shared-weights condition.
  5. [Section II-A] The text cites 'KGCL [18]' when discussing similar approaches, but reference [18] is a session-based KGCL paper that is distinct from the KGCL baseline used in the experiments (reference [6]); please correct the citation.
  6. [Section III-E and Table III] Eq. (16) defines a single L2 regularization coefficient lambda for the parameter norm, but Table III reports separate lambda_user2 and lambda_item2 values and Section IV-D mentions a grid-searched L2 value of 10^-4; please clarify how these relate to lambda in the objective.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KGIF's central claims are empirical benchmarks against external datasets and baselines.

full rationale

The paper's load-bearing claim, stated in Section IV-E ('KGIF consistently outperforms all baseline models across the three benchmark datasets in terms of Recall@20 and NDCG@20'), is supported by Table II, which compares KGIF against eight external baseline models on three standard benchmarks (Amazon-book, Last-FM, Yelp2018). The prediction scores are produced by Eq. (14), an inner product of user and item embeddings trained with the BPR loss in Eq. (15), and evaluated on held-out test interactions that are not used to fit the model. No target quantity is defined in terms of the model's own output: the test labels come from a random 7:2:1 split of user-item interactions, and the baselines are independently published methods with their own reported hyperparameters. The hyperparameter grid search over L2 regularization and dropout is performed on a validation set, and these values are not re-labeled as predictions; they are standard training choices. The fusion design h' = h⊥ ⊙ r (Eq. 6) and the propagation/aggregation operations are architectural proposals, not quantities fitted to the evaluation targets. There is no self-citation of the authors' prior work carrying an unverified premise, and no uniqueness theorem is imported to force a choice. The tuning-fairness concern noted in Section IV-D ('For baseline models, the original hyperparameter settings were retained to ensure optimal performance as reported in their respective papers') is a threat to the strength of the empirical comparison, but it is not circularity: it does not make the reported test metrics equivalent by construction to the fitted parameters. Accordingly, the paper is self-contained with respect to circularity, and the score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or ontological entities. Its dynamic projection vectors and weight matrices are learnable parameters, not independently evidenced entities. The ledger therefore records free hyperparameters and modeling assumptions only.

free parameters (6)
  • Embedding size d = 64
    All models use embedding dimension 64; KGIF also uses hidden dimensions 64, 32, 16, 8 across propagation layers.
  • Learning rate rho = 1e-4
    Selected by grid search for KGIF and listed in Table III.
  • L2 regularization lambda_user2 and lambda_item2 = 1e-5 for Amazon-book and Last-FM; 1e-4 for Yelp2018
    Tuned per dataset by grid search and enters the objective in Eq. (16).
  • Dropout p_drop = 0.1
    Grid-searched and applied at each layer.
  • Number of propagation layers L = 4
    Chosen after ablation AB-03 (Table V), where 4 layers gives the best Recall and NDCG on all datasets.
  • N-core threshold = 10
    Data preprocessing retains only users and items with 10 or more interactions; this changes dataset density and all downstream results.
assumptions (5)
  • domain assumption CKG formulation treats user-item interactions and item-attribute relations as edges in one heterogeneous graph with a shared embedding space.
    Section III-A defines GCKG; Section III-B embeds all entities in R^m. The paper claims to address homogeneity but still uses a single embedding space for users, items, and attributes.
  • domain assumption TransD scoring g(h,r,t) = -||h_perp + r - t_perp||^2 measures triplet plausibility.
    Eq. (4) and the embedding loss Eq. (5) rely on this scoring; no independent justification is given beyond citing TransD [24].
  • domain assumption Self-attention scores in Eq. (8)-(9) encode the importance of triplets for propagation.
    The attentive propagation layer assumes attention weights reflect relational importance; this is inherited from KGAT [8] without proof.
  • domain assumption BPR pairwise loss with randomly sampled negative items is appropriate for implicit-feedback evaluation.
    Eq. (15) uses BPR loss from [30]; it is a standard recommender assumption about user preference data.
  • domain assumption Random 7:2:1 split after N-core filtering produces a representative train/validation/test evaluation.
    Section IV-A does a single random split with no seeds or repeated runs; evaluation robustness is assumed rather than demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KGIF: Optimizing Relation-Aware Recommendations with Knowledge Graph Information Fusion." pith.science (2026). https://pith.science/paper/RRHZMJRC

@misc{pith2026250104161,
  author       = {Pith},
  title        = {Pith review of: KGIF: Optimizing Relation-Aware Recommendations with Knowledge Graph Information Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRHZMJRC}},
  note         = {Machine review of arXiv:2501.04161}
}
read the original abstract

While deep-learning-enabled recommender systems demonstrate strong performance benchmarks, many struggle to adapt effectively in real-world environments due to limited use of user-item relationship data and insufficient transparency in recommendation generation. Traditional collaborative filtering approaches fail to integrate multifaceted item attributes, and although Factorization Machines account for item-specific details, they overlook broader relational patterns. Collaborative knowledge graph-based models have progressed by embedding user-item interactions with item-attribute relationships, offering a holistic perspective on interconnected entities. However, these models frequently aggregate attribute and interaction data in an implicit manner, leaving valuable relational nuances underutilized. This study introduces the Knowledge Graph Attention Network with Information Fusion (KGIF), a specialized framework designed to merge entity and relation embeddings explicitly through a tailored self-attention mechanism. The KGIF framework integrates reparameterization via dynamic projection vectors, enabling embeddings to adaptively represent intricate relationships within knowledge graphs. This explicit fusion enhances the interplay between user-item interactions and item-attribute relationships, providing a nuanced balance between user-centric and item-centric representations. An attentive propagation mechanism further optimizes knowledge graph embeddings, capturing multi-layered interaction patterns. The contributions of this work include an innovative method for explicit information fusion, improved robustness for sparse knowledge graphs, and the ability to generate explainable recommendations through interpretable path visualization.

Figures

Figures reproduced from arXiv: 2501.04161 by the authors.

Figure 1
Figure 1. The SOTA recommender models work without fusion. KGIF utilizes not only attribute but also relation information to update both user and item representation for fused CKG embedding. has sparse user-item interactions and complex item-attribute interactions. Intuitively, incorporating such closely related information could enhance recommendations. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall process of KGIF. (a) Learn a set of initial CKG embedding with TransD. (b) Information fusion by using projection vectors from the previous step. (c) Attentive embedding propagation with self-attention of triplets for updating CKG embedding. (d) Recommendation￾making based on user-item information aggregation. attention layer, enabling the attentive learning process to con￾sider side information in the final… view at source ↗
Figure 3
Figure 3. Comparison between CF and CKG. CF recommends items solely based on user-item interactions, while CKG incorporates item attributes in addition to user-item interactions when making recommendations to users. B. CKG Embedding with TransD The raw CKG must first be converted into a latent semantic space to facilitate easier manipulation. Fusing the side infor￾mation with relationships at the beginning can introduce un￾ne… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: First-order ego-network of a node h. The head entity is in the middle denoted as hEgo and tails are denoted as t0−4. hEgo = X (h,r,t)∈Egoh ˆδ(h, r, t)t ∗ . (10) Furthermore, the propagation layers can be stacked to explore a higher-order connectivity: h (l−1) Ego = X (…
Figure 5
Figure 5. Figure 5: Comparative Analysis of Loss and Recall: Fusion vs. Without Fusion (Original) Approach [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance of Fusion with Various Embedding Methods. gesting that explicit information fusion further enhances the model’s ability to manage noisy data and extract meaningful patterns. KGIF’s ability to dynamically adapt to both user￾item and item-attribute relationsh…
Figure 7
Figure 7. Figure 7: Visual Interpretation of Recommendation-Making with Atten￾tion Scores additional contributions come from the interaction with The Throttle (0.2378) and another author-related score from The Shining (0.5495). These scores are aggregated, resulting in a strong cumulative…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages

  1. [1]

    Data scarcity in recommendation systems: A survey,

    Z. Chen, W. Gan, J. Wu, K. Hu, and H. Lin, “Data scarcity in recommendation systems: A survey,” ACM Trans. Recomm. Syst. , jan 2024, just Accepted. [Online]. Available: https://doi.org/10.1145/ 3639063

  2. [2]

    Grou- plens: An open architecture for collaborative filtering of netnews,

    P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl, “Grou- plens: An open architecture for collaborative filtering of netnews,” in Proceedings of the 1994 ACM conference on Computer supported cooperative work, 1994, pp. 175–186

  3. [3]

    Translating embeddings for modeling multi-relational data,

    A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, “Translating embeddings for modeling multi-relational data,” Advances in neural information processing systems , vol. 26, 2013

  4. [4]

    Knowledge graph embedding by translating on hyperplanes,

    Z. Wang, J. Zhang, J. Feng, and Z. Chen, “Knowledge graph embedding by translating on hyperplanes,” in Proceedings of the AAAI conference on artificial intelligence , vol. 28, no. 1, 2014

  5. [5]

    Learning entity and relation embeddings for knowledge graph completion,

    Y . Lin, Z. Liu, M. Sun, Y . Liu, and X. Zhu, “Learning entity and relation embeddings for knowledge graph completion,” in Proceedings of the AAAI conference on artificial intelligence , vol. 29, no. 1, 2015

  6. [6]

    Knowledge graph contrastive learning for recommendation,

    Y . Yang, C. Huang, L. Xia, and C. Li, “Knowledge graph contrastive learning for recommendation,” in Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, 2022, pp. 1434–1443

  7. [7]

    Knowledge graph self- supervised rationalization for recommendation,

    Y . Yang, C. Huang, L. Xia, and C. Huang, “Knowledge graph self- supervised rationalization for recommendation,” in Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, 2023, pp. 3046–3056

  8. [8]

    Kgat: Knowledge graph attention network for recommendation,

    X. Wang, X. He, Y . Cao, M. Liu, and T.-S. Chua, “Kgat: Knowledge graph attention network for recommendation,” inProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 950–958

Show all 40 references
  1. [9]

    Graph neural networks for rec- ommender system,

    C. Gao, X. Wang, X. He, and Y . Li, “Graph neural networks for rec- ommender system,” in Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining , 2022, pp. 1623–1625

  2. [10]

    Higher-order attribute-enhancing heterogeneous graph neural networks,

    J. Li, H. Peng, Y . Cao, Y . Dou, H. Zhang, S. Y . Philip, and L. He, “Higher-order attribute-enhancing heterogeneous graph neural networks,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 1, pp. 560–574, 2021

  3. [11]

    Towards robust graph neural networks for noisy graphs with sparse labels,

    E. Dai, W. Jin, H. Liu, and S. Wang, “Towards robust graph neural networks for noisy graphs with sparse labels,” in Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, 2022, pp. 181–191

  4. [12]

    CKAN: Collaborative knowledge-aware attentive network for recommender systems,

    Z. Wang, G. Lin, H. Tan, Q. Chen, and X. Liu, “CKAN: Collaborative knowledge-aware attentive network for recommender systems,” in Pro- ceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval , 2020, pp. 219–228

  5. [13]

    Learning intents behind interactions with knowledge graph for recommendation,

    X. Wang, T. Huang, D. Wang, Y . Yuan, Z. Liu, X. He, and T.-S. Chua, “Learning intents behind interactions with knowledge graph for recommendation,” in Proceedings of the web conference 2021, 2021, pp. 878–887

  6. [14]

    Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,

    R. He and J. McAuley, “Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,” in proceedings of the 25th international conference on world wide web , 2016, pp. 507– 517

  7. [15]

    Celma, Music Recommendation and Discovery in the Long Tail

    O. Celma, Music Recommendation and Discovery in the Long Tail . Springer, 2010. [Online]. Available: https://doi.org/10.1145/3639063

  8. [16]

    Neural graph collaborative filtering,

    X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph collaborative filtering,” in Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Re- trieval, 2019, pp. 165–174

  9. [17]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

  10. [18]

    Kgcl: A knowledge- enhanced graph contrastive learning framework for session-based rec- ommendation,

    X. Zhang, H. Ma, F. Yang, Z. Li, and L. Chang, “Kgcl: A knowledge- enhanced graph contrastive learning framework for session-based rec- ommendation,” Engineering Applications of Artificial Intelligence , vol. 124, p. 106512, 2023

  11. [19]

    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, and others, “Feature-level deeper self-attention network for sequential recommendation.” in IJCAI, 2019, pp. 4320–4326

  12. [20]

    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 Proceedings of the 29th ACM international conference on information & knowledge managem...

  13. [21]

    Parallel recurrent neural network architectures for feature-rich session-based recommen- dations,

    B. Hidasi, M. Quadrana, A. Karatzoglou, and D. Tikk, “Parallel recurrent neural network architectures for feature-rich session-based recommen- dations,” in Proceedings of the 10th ACM conference on recommender systems, 2016, pp. 241–248

  14. [22]

    Icai-sr: Item categorical attribute integrated sequential recommendation,

    X. Yuan, D. Duan, L. Tong, L. Shi, and C. Zhang, “Icai-sr: Item categorical attribute integrated sequential recommendation,” in Proceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval , 2021, pp. 1687–1691

  15. [23]

    Decoupled side information fusion for sequential recommendation,

    Y . Xie, P. Zhou, and S. Kim, “Decoupled side information fusion for sequential recommendation,” in Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, 2022, pp. 1611–1621

  16. [24]

    Knowledge graph embedding via dynamic mapping matrix,

    G. Ji, S. He, L. Xu, K. Liu, and J. Zhao, “Knowledge graph embedding via dynamic mapping matrix,” in Proceedings of the 53rd annual meeting of the association for computational linguistics and the 7th international joint conference on natural language processing (volume 1: Lon...

  17. [25]

    Prefix-tuning: Optimizing continuous prompts for generation,

    X. L. Li and P. Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” arXiv preprint arXiv:2101.00190 , 2021

  18. [26]

    P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,

    X. Liu, K. Ji, Y . Fu, W. L. Tam, Z. Du, Z. Yang, and J. Tang, “P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,” arXiv preprint arXiv:2110.07602 , 2021

  19. [27]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio et al., “Graph attention networks,” stat, vol. 1050, no. 20, pp. 10–48 550, 2017

  20. [28]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

  21. [29]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

  22. [30]

    BPR: Bayesian personalized ranking from implicit feedback,

    S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “BPR: Bayesian personalized ranking from implicit feedback,” arXiv preprint arXiv:1205.2618, 2012

  23. [31]

    Fast context-aware recommendations with factorization machines,

    S. Rendle, Z. Gantner, C. Freudenthaler, and L. Schmidt-Thieme, “Fast context-aware recommendations with factorization machines,” in Pro- ceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval , 2011, pp. 635–644

  24. [32]

    Neural factorization machines for sparse predictive analytics,

    X. He and T.-S. Chua, “Neural factorization machines for sparse predictive analytics,” in Proceedings of the 40th international ACM SIGIR conference on research and development in information retrieval, 2017, pp. 355–364

  25. [33]

    Ripplenet: Propagating user preferences on the knowledge graph for recommender systems,

    H. Wang, F. Zhang, J. Wang, M. Zhao, W. Li, X. Xie, and M. Guo, “Ripplenet: Propagating user preferences on the knowledge graph for recommender systems,” in Proceedings of the 27th ACM international conference on information and knowledge management , 2018, pp. 417– 426

  26. [34]

    Graph convolutional matrix completion,

    R. van den Berg, T. N. Kipf, and M. Welling, “Graph convolutional matrix completion,” 2018

  27. [35]

    Knowledge-aware graph neural networks with label smooth- ness regularization for recommender systems,

    H. Wang, F. Zhang, M. Zhang, J. Leskovec, M. Zhao, W. Li, and Z. Wang, “Knowledge-aware graph neural networks with label smooth- ness regularization for recommender systems,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , ...

  28. [36]

    Understanding the difficulty of training deep feedforward neural networks,

    X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proceedings of the thirteenth international conference on artificial intelligence and statistics . JMLR Workshop and Conference Proceedings, 2010, pp. 249–256

  29. [37]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

  30. [38]

    Rein- forcement knowledge graph reasoning for explainable recommendation,

    Y . Xian, Z. Fu, S. Muthukrishnan, G. De Melo, and Y . Zhang, “Rein- forcement knowledge graph reasoning for explainable recommendation,” in Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval , 2019, pp. 285–294

  31. [39]

    Jointly learning explainable rules for recommendation with knowledge graph,

    W. Ma, M. Zhang, Y . Cao, W. Jin, C. Wang, Y . Liu, S. Ma, and X. Ren, “Jointly learning explainable rules for recommendation with knowledge graph,” in The world wide web conference , 2019, pp. 1210–1221

  32. [40]

    Measuring and relieving the over-smoothing problem for graph neural networks from the topological view,

    D. Chen, Y . Lin, W. Li, P. Li, J. Zhou, and X. Sun, “Measuring and relieving the over-smoothing problem for graph neural networks from the topological view,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, 2020, pp. 3438–3445, number: 04

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.