Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Boosting Knowledge Graph-based Recommendations through Confidence-Aware Augmentation with Large Language Models

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read This paper claims that a confidence-gated pipeline, in which an LLM suggests knowledge-graph edits and a learned score decides which triples to keep, improves ranking accuracy and explanation quality in knowledge-graph-based recommendation.

desk verdict A coherent LLM+KG recommendation framework with a clean architecture, but the evaluation pipeline has a potential label-leakage problem that must be fixed before the empirical claims can be trusted. read the letter →

arxiv 2502.03715 v1 pith:GONSWJLC submitted 2025-02-06 cs.IR cs.AI

classification cs.IRcs.AI
keywords knowledgegraphrecommendationlargelanguagemodelsconfidence-awareaugmentationcontrastivelearningmessagepropagationexplanationgenerationneuralnetworkscollaborativefiltering
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

Recommender systems that use knowledge graphs often inherit noise and stale relations from the graph. This paper proposes CKG-LLMA, a pipeline that asks a large language model to suggest additions and deletions to item knowledge subgraphs, then learns a confidence score for every triplet so that untrustworthy LLM suggestions can be gated before they reach the recommender. The framework couples this confidence-aware message propagation with dual-view contrastive learning that aligns user and item representations across an item-view graph and a user-view graph. The paper reports that the resulting model outperforms both graph-based and LLM-based baselines on AmazonBook, Steam, and Anime, and that the learned confidence scores also help an LLM choose more plausible explanation paths for recommendations. The claim is that treating LLM knowledge edits as noisy, confidence-scored supervision is a better way to combine LLMs with knowledge-graph recommenders than using the LLM output at face value.

What carries the argument

The load-bearing mechanism is the confidence-aware mixture-of-experts gate on item-attribute triples. For each triple $(i, r, a)$, the model computes a score $\varphi(i, r, a)$ from an attention-updated item embedding, the relation embedding, and the attribute embedding via a set of expert networks, then uses the sigmoid of that score as the probability of keeping the link during differentiable Gumbel-softmax sampling. This gate is applied both to the original graph and to the LLM's added triples, so that hallucinated or irrelevant knowledge can be pruned before message propagation. A second mechanism, the dual-view two-step contrastive learning, builds two augmented graphs from the user-view and item-view advice pools and aligns their user and item representations through a contrastive loss, with a stability score deciding which interaction edges to mask. Together these components turn the LLM's free-form add/delete advice into soft, learnable graph structure.

What would settle it

Split each user's interactions into train/validation/test before querying the LLM, so that only training interactions appear in the subgraphs and advice pools, then re-run the full CKG-LLMA pipeline; if Recall@10 and NDCG@10 on AmazonBook, Steam, and Anime drop toward or below the best baselines, the reported gains depend on test-set leakage. As a cheaper check, count how many test-set user-item pairs appear among the LLM's 'add' or retained 'delete' triples in the augmented graphs $G_U$ and $G_I$.

Watch

Extended reading notes

Core claim

The central discovery is that LLM-generated knowledge graph edits, when screened by a learnable triplet confidence score instead of being trusted wholesale, yield consistent ranking gains in knowledge-graph-based recommendation. Concretely, an LLM is prompted in two views—one containing user-item and item-attribute triples, the other item-attribute and item-item triples—to return 'add' and 'delete' advice; these four advice pools are sampled and merged into the original graph. A mixture-of-experts layer scores each item-attribute triple, a Gumbel-softmax routine drops low-confidence links in a differentiable way, and two augmented views of the interaction graph are contrasted to learn stable user and item embeddings. On AmazonBook, Steam, and Anime, the paper reports Recall@10 of 0.1883, 0.0672, and 0.1069, exceeding KGRec and the LLM-based baselines LLMRec and RLMRec. The same confidence scores, fed to an LLM in a chain-of-thought prompt, steer it toward a plausible reasoning path for why a user might like an item.

Load-bearing premise

The load-bearing premise is that the LLM subgraph augmenter, which runs before the train/validation/test split, does not expose the model to test labels; if the collaborative signals used to build subgraphs include test interactions, the LLM's delete advice and the augmented knowledge graph leak future behavior into training, inflating the reported Recall/NDCG gains.

Editorial extensions

If this is right

  • If the reported results hold, LLM-augmented knowledge graph recommenders can improve top-10 recall and NDCG on book, game, and anime benchmarks without retraining the LLM during model training.
  • The confidence score becomes a reusable filter: the same mechanism can denoise any source of KG triplets, not only LLM output, by scoring and dropping low-confidence item-attribute links.
  • Ablation results imply that the contrastive loss and the confidence dropout each contribute substantially; removing either degrades Recall@10 by roughly 0.05 on AmazonBook.
  • The explanation module suggests that learned triplet confidence can be exported as structured evidence, making LLM-generated recommendation explanations more grounded in graph paths.

Reading between the lines

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

  • A natural extension is to run the LLM augmentation exclusively on training interactions before the split; if the reported gains persist under that restriction, the leakage concern is resolved and the method becomes safer to deploy.
  • The confidence gate could be transferred to other GNN recommenders that consume external knowledge, treating it as a drop-in denoising layer rather than as part of a full pipeline.
  • The two-view prompt design suggests a general recipe: ask the LLM for both explicit fact verification and implicit preference reasoning, then let a learned gate arbitrate between the two kinds of advice.
  • The reported low preprocessing cost suggests the augmentation step could be re-run with an open-source LLM, and the confidence gate would be expected to handle the different noise profile of a weaker model.
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 CKG-LLMA, a framework that augments knowledge graphs for recommendation by using LLM-generated subgraph modifications (add/delete advice from user-view and item-view prompts), learning a confidence score for each triplet via a mixture-of-experts layer, and combining KG propagation with dual-view contrastive learning. The framework also uses the learned confidence scores to guide LLM-generated explanations. Experiments on AmazonBook, Steam, and Anime report consistent improvements over CF, KG-based, self-supervised, and LLM-based baselines, with ablations supporting the confidence and contrastive components.

Significance. If the results are valid, CKG-LLMA offers a practical template for using LLMs to refresh static KGs in recommender systems, with a learned confidence mechanism to mitigate LLM hallucinations and a contrastive objective that ties KG augmentation to interaction modeling. Strengths of the paper include the clearly described pipeline, the use of offline LLM augmentation with reported API costs, extensive baseline comparisons, and ablations. The central risk is whether the offline LLM augmentation inadvertently uses held-out interaction labels, which would invalidate the empirical claims.

major comments (3)
  1. [Section 4.1, Section 3.2, Eq. (11), Figure 3] The LLM subgraph augmenter is run before the train/validation/test split and samples collaborative signals from the full interaction graph, which includes test interactions. The user-view prompt in Figure 3 asks the LLM to delete user-item triples if 'user may not like this item,' and the resulting delete set is applied in Eq. (11) to construct the user-view augmented interaction graph. Because the LLM can condition its augmentation advice on the very user-item labels later used for evaluation, the reported improvements (e.g., AmazonBook Recall@10 0.1883 vs. 0.1821 for KGRec) could be inflated by label leakage. The paper should either explicitly state that augmentation is restricted to the training split, or re-run the evaluation with augmentation performed only on training interactions, and report whether the conclusions change.
  2. [Table 2, Section 4.4] The paper reports no error bars or significance tests for the main results. The margins over the best baselines are modest (roughly 3–6% relative in Recall@10), so without multiple runs it is impossible to judge whether the differences are statistically reliable. The text in Section 4.4 mentions 'significance hypothesis testing' but provides no test details or p-values. I request mean and standard deviation over at least three random seeds, and a paired significance test (e.g., t-test or Wilcoxon) for the comparisons against the strongest baselines.
  3. [Section 4.3] The 'w/o Contrastive Loss' ablation is described as nullifying the contrastive loss by setting its weight to zero and training 'solely on lbpr without LLM-enhanced information or the confidence-aware denoising mechanism.' This conflates several components: removing contrastive loss alone is different from also removing LLM enhancement and confidence dropout. The observed large performance drop (e.g., AmazonBook Recall@10 from 0.1883 to 0.1312) cannot be attributed specifically to the contrastive loss under this description. Please provide separate ablations for each component, or clearly state which components are removed in each variant.
minor comments (4)
  1. [Section 4.4, Table 5] The optimal delete ratio for AmazonBook is reported inconsistently: Table 5 lists μd = 0.08, while Section 4.4 states 'μd = 4% for AmazonBook.' The figure appears to be consistent with 0.08. Also, Table 5 uses the symbol λθ for both the contrastive loss weight and the regularization loss weight; one of these should be λc.
  2. [Figure 3, Section 8] The system prompts in Figure 3 say 'You are GPT-4,' but Section 8 states that GPT-3.5-turbo was used for subgraph augmentation. Please align the described LLM with the reported experimental setup.
  3. [Section 3.4, Eq. (11)] The formula defining the keep probability pi has ambiguous parenthesization: 'pi = 1 − pdrop / meani(exp(si)) ∗ exp(si) − maxi(exp(si)) / maxi(exp(si)) − mini(exp(si))' is hard to parse. Please rewrite with clear parentheses and define all symbols in the equation.
  4. [General] The manuscript contains several typos and grammatical slips (for example, 'and and formally defining' in Section 3.1, 'recomendation' in Section 2, and 'propogation' in Section 3). A careful proofreading pass is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the recommendation scores come from a BPR objective on held-out split interactions, and the confidence/contrastive components are auxiliary learned modules rather than renamed inputs or fitted predictions.

full rationale

The paper's central predictive claim is the ranking score x_u^T x_i, optimized by the BPR loss in Eq. (13) after the 8:1:1 train/validation/test split described in Section 4.1. The confidence scores in Eq. (2) are learned latent scalars used to filter triplets and construct contrastive views; they are not the evaluation labels, and no parameter is fitted directly to Recall@10/NDCG@10. The LLM augmentation pools are produced by prompting ChatGPT on extracted subgraphs and are used to build the augmented graphs in Eqs. (6)-(11), but the paper does not present a formal derivation that reduces the final prediction to these pools. Self-citations by the authors appear only as related-work background (e.g., [Wang et al., 2023], [Wu et al., 2024], [Wang et al., 2024]) and are not load-bearing for the proposed mechanism. The explanation module reuses the same learned confidence scores, but it is a qualitative downstream demonstration, not the main predictive result, so it does not constitute circularity. A possible pre-split leakage from running the LLM augmenter before the data split is a correctness/validity concern rather than a circular derivation, and it is not an instance where an output equation equals an input by construction.

Assumptions & free parameters 12 free parameters · 7 assumptions · 2 invented entities

The central claim rests on standard neural optimization plus several domain assumptions. The main components are learnable model parameters and validation-tuned hyperparameters; no new physical entity is introduced. The most consequential assumption is that the pre-split LLM augmentation does not leak test information into the graph.

free parameters (12)
  • Hidden dimension d = 64
    Model dimension; set by validation.
  • Add ratio mu_a = 0.60 (AmazonBook), 0.70 (Steam best); table shows 0.60 for AmazonBook
    Controls how many LLM 'add' triplets enter the augmented KG; tuned per dataset.
  • Delete ratio mu_d = 0.08 in Table 5 for AmazonBook, but 4 percent stated optimal in Section 4.4
    Controls how many LLM 'delete' triplets are removed; report is internally inconsistent.
  • Number of experts Ne = 8
    Mixture-of-experts width for confidence scoring.
  • Scale K in selection probability = 5
    Controls sharpness of Sigmoid(phi * K) for triplet keeping.
  • Gumbel temperature tau_g = 0.9
    Relaxes discrete triplet sampling.
  • Drop probability p_drop = 0.01
    Base probability for dropping item nodes in stability-based IG augmentation.
  • Message propagation layers L = 3
    Depth of LightGCN propagation.
  • Contrastive temperature tau = 0.2
    Temperature in node-contrastive loss.
  • Loss weights lambda_c and lambda_theta = 1e-3 and 1e-4
    Weights for contrastive and regularization losses.
  • Subgraph size N = 32 (AmazonBook), 40 (Steam), 64 (Anime)
    Number of collaborative signals sampled per LLM subgraph.
  • Learnable confidence network parameters = trained end-to-end
    The confidence score phi(i,r,a) used to filter triplets and guide explanations is not derived; it is learned jointly with BPR and contrastive losses.
assumptions (7)
  • domain assumption LLM augmentation advice from GPT-3.5-turbo is factually grounded and useful for the recommendation task.
    The whole pipeline depends on the LLM's add/delete suggestions being at least partly correct; hallucination is acknowledged but not quantified.
  • domain assumption The constructed TKG, built from item attributes and item-item links inferred from shared attributes, adequately represents side information for these datasets.
    Section 4.1 describes KG construction; if the KG is noisy or misaligned, results are not comparable to benchmarks using standard KG data.
  • standard math Gumbel-softmax provides a differentiable approximation to discrete triplet sampling.
    Invoked from Jang et al. in Section 3.4; no new proof needed.
  • domain assumption BPR ranking loss plus contrastive loss is an appropriate surrogate for Recall@10/NDCG@10.
    Standard practice, but no guarantee that optimizing this objective yields ranking improvements.
  • domain assumption The held-out split and full-ranking protocol fairly estimate generalization.
    Requires that no test labels enter the LLM augmentation or the KG construction; the paper does not explicitly ensure this.
  • domain assumption The user-item interactions in the dataset reflect stable preferences that an LLM can reason about.
    User-view prompt deletes items the LLM believes the user may not like, which assumes LLM preference inference is valid.
  • domain assumption The learned confidence scores generalize from recommendation training to the explanation task.
    Section 3.6 uses confidence scores to direct LLM path selection; no independent validation of these scores as explanations.
invented entities (2)
  • Learned triplet confidence score phi(i,r,a)
    purpose: Filters noisy or LLM-generated triplets in message propagation and ranks reason paths in explanation generation.
    The score is a learnable scalar trained on the recommendation objective; the paper gives no external calibration or falsifiable prediction that would validate the score outside the model.
  • Cross-view stability score s_i
    purpose: Used to drop low-stability item nodes when augmenting interaction graph views.
    Derived from cosine similarity of view-specific item embeddings; no external handle. It is a model-internal latent quantity, not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting Knowledge Graph-based Recommendations through Confidence-Aware Augmentation with Large Language Models." pith.science (2026). https://pith.science/paper/GONSWJLC

@misc{pith2026250203715,
  author       = {Pith},
  title        = {Pith review of: Boosting Knowledge Graph-based Recommendations through Confidence-Aware Augmentation with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GONSWJLC}},
  note         = {Machine review of arXiv:2502.03715}
}
read the original abstract

Knowledge Graph-based recommendations have gained significant attention due to their ability to leverage rich semantic relationships. However, constructing and maintaining Knowledge Graphs (KGs) is resource-intensive, and the accuracy of KGs can suffer from noisy, outdated, or irrelevant triplets. Recent advancements in Large Language Models (LLMs) offer a promising way to improve the quality and relevance of KGs for recommendation tasks. Despite this, integrating LLMs into KG-based systems presents challenges, such as efficiently augmenting KGs, addressing hallucinations, and developing effective joint learning methods. In this paper, we propose the Confidence-aware KG-based Recommendation Framework with LLM Augmentation (CKG-LLMA), a novel framework that combines KGs and LLMs for recommendation task. The framework includes: (1) an LLM-based subgraph augmenter for enriching KGs with high-quality information, (2) a confidence-aware message propagation mechanism to filter noisy triplets, and (3) a dual-view contrastive learning method to integrate user-item interactions and KG data. Additionally, we employ a confidence-aware explanation generation process to guide LLMs in producing realistic explanations for recommendations. Finally, extensive experiments demonstrate the effectiveness of CKG-LLMA across multiple public datasets.

Figures

Figures reproduced from arXiv: 2502.03715 by the authors.

Figure 1
Figure 1. The overview of the proposed CKG-LLMA framework for knowledge graph based recommendations. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Parameter Analysis of add ratio and delete ratio. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Prompts for LLM-based Subgraph Augmenter. Upper left is Item-view prompt and lower left is its related example. Upper right is [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The overall pipeline for Confidence-aware Explanation Generation. The left is the framework of generation process, and the right [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 21 canonical work pages

  1. [1]

    Can knowledge graphs reduce hallucinations in llms?: A survey

    [Agrawal et al., 2023] Garima Agrawal, Tharindu Ku- marage, Zeyad Alghami, and Huan Liu. Can knowledge graphs reduce hallucinations in llms?: A survey. arXiv preprint arXiv:2311.07914,

  2. [3]

    Long short-term memory

    [Graves and Graves, 2012] Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37–45,

  3. [4]

    Then we identify items with an item-item relation- ship to target item i and derive all corresponding I-I triplets

    For certain user-item interaction pair ( u, i), the module initially retrieves interacted items of user u to search for relevant U-I triplets. Then we identify items with an item-item relation- ship to target item i and derive all corresponding I-I triplets. In the next step, find the co-occupied attribute in each I-I triplet and obtain two item-entity tr...

  4. [5]

    Table 5: Hyperparameter settings of CKG-LLMA in AmazonBook dataset Hyperparameter Settings Learning rate lr 1e-4 Hidden dimension d 64 Subgraph size N 32 Delete ratio µd 0.08 Add ratio µa 0.60 Num of Experts Ne 8 Scalable parameter K 5 Temperature in gumbel softmax τg 0.9 Drop probability pdrop 0.01 Message propogation layers L 3 Temperature in contrastiv...

  5. [7]

    Masked au- toencoders are scalable vision learners

    [He et al., 2022] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked au- toencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009,

  6. [9]

    Self-attentive sequential recommendation

    [Kang and McAuley, 2018] Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), pages 197–206. IEEE,

  7. [12]

    [Ren et al., 2024] Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang

    Accessed: 2020-01-23. [Ren et al., 2024] Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Representation learning with large language mod- els for recommendation. In Proceedings of the ACM on Web Conference 2024, pages 3464–3475,

  8. [14]

    Outrageously large neural net- works: The sparsely-gated mixture-of-experts layer

    [Shazeer et al., 2017] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hin- ton, and Jeff Dean. Outrageously large neural net- works: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538,

Show all 33 references
  1. [16]

    Graph attention networks

    [Veliˇckovi´c et al., 2017] Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903,

  2. [19]

    Unleashing the power of knowl- edge graph for recommendation via invariant learning

    [Wang et al., 2024] Shuyao Wang, Yongduo Sui, Chao Wang, and Hui Xiong. Unleashing the power of knowl- edge graph for recommendation via invariant learning. In Proceedings of the ACM on Web Conference 2024 , pages 3745–3755,

  3. [20]

    Chain-of-thought prompting elicits reasoning in large language models

    [Wei et al., 2022] Jason Wei, Xuezhi Wang, Dale Schuur- mans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837,

  4. [21]

    Llmrec: Large language models with graph augmentation for recommendation

    [Wei et al., 2024] Wei Wei, Xubin Ren, Jiabin Tang, Qiny- ong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Llmrec: Large language models with graph augmentation for recommendation. In Pro- ceedings of the 17th ACM International Conference on Web Search ...

  5. [22]

    Self-supervised graph learning for recommendation

    [Wu et al., 2021] Jiancan Wu, Xiang Wang, Fuli Feng, Xi- angnan He, Liang Chen, Jianxun Lian, and Xing Xie. Self-supervised graph learning for recommendation. In Proceedings of the 44th international ACM SIGIR con- ference on research and development in information re- trieval...

  6. [23]

    A survey on large language models for recommendation

    [Wu et al., 2023] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. A survey on large language models for recommendation. arXiv preprint arXiv:2305.19860,

  7. [24]

    Afdgcf: Adaptive feature de-correlation graph collaborative filtering for rec- ommendations

    [Wu et al., 2024] Wei Wu, Chao Wang, Dazhong Shen, Chuan Qin, Liyi Chen, and Hui Xiong. Afdgcf: Adaptive feature de-correlation graph collaborative filtering for rec- ommendations. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Inf...

  8. [25]

    Multi-perspective improvement of knowledge graph completion with large language mod- els

    [Xu et al., 2024] Derong Xu, Ziheng Zhang, Zhenxi Lin, Xian Wu, Zhihong Zhu, Tong Xu, Xiangyu Zhao, Yefeng Zheng, and Enhong Chen. Multi-perspective improvement of knowledge graph completion with large language mod- els. arXiv preprint arXiv:2403.01972,

  9. [26]

    Knowledge graph contrastive learning for recommendation

    [Yang et al., 2022] Yuhao Yang, Chao Huang, Lianghao Xia, and Chenliang Li. Knowledge graph contrastive learning for recommendation. In Proceedings of the 45th interna- tional ACM SIGIR conference on research and develop- ment in information retrieval, pages 1434–1443,

  10. [27]

    Knowledge graph self-supervised rationalization for recommendation

    [Yang et al., 2023] Yuhao Yang, Chao Huang, Lianghao Xia, and Chunzhen Huang. Knowledge graph self-supervised rationalization for recommendation. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, pages 3046–3056,

  11. [28]

    Are graph augmentations necessary? simple graph contrastive learning for recommendation

    [Yu et al., 2022] Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and develop...

  12. [29]

    Xsimgcl: Towards extremely simple graph contrastive learning for recommendation

    [Yu et al., 2023] Junliang Yu, Xin Xia, Tong Chen, Lizhen Cui, Nguyen Quoc Viet Hung, and Hongzhi Yin. Xsimgcl: Towards extremely simple graph contrastive learning for recommendation. IEEE Transactions on Knowledge and Data Engineering, 36(2):913–926,

  13. [30]

    Collaborative knowledge base embedding for recommender systems

    [Zhang et al., 2016] Fuzheng Zhang, Nicholas Jing Yuan, Defu Lian, Xing Xie, and Wei-Ying Ma. Collaborative knowledge base embedding for recommender systems. In Proceedings of the 22nd ACM SIGKDD international con- ference on knowledge discovery and data mining , pages 353–362,

  14. [31]

    Knowledge graph enhanced large language model editing

    [Zhang et al., 2024] Mengqi Zhang, Xiaotian Ye, Qiang Liu, Pengjie Ren, Shu Wu, and Zhumin Chen. Knowledge graph enhanced large language model editing. arXiv preprint arXiv:2402.13593,

  15. [1985]

    Myanimelist - anime and manga database and community,

    [MyAnimeList, 2020] MyAnimeList. Myanimelist - anime and manga database and community,

  16. [2012]

    A survey on knowledge graph-based recommender systems

    [Guo et al., 2020] Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu Zhu, Xing Xie, Hui Xiong, and Qing He. A survey on knowledge graph-based recommender systems. IEEE Transactions on Knowledge and Data Engineering , 34(8):3549–3568,

  17. [2016]

    Lightgcn: Simplifying and powering graph convolution network for recommendation

    [He et al., 2020] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Informatio...

  18. [2017]

    Think-on-graph: Deep and re- sponsible reasoning of large language model with knowl- edge graph

    [Sun et al., 2023] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Heung- Yeung Shum, and Jian Guo. Think-on-graph: Deep and re- sponsible reasoning of large language model with knowl- edge graph. arXiv preprint arXiv:2307.07697,

  19. [2018]

    A family of bivariate distributions generated by the bivariate bernoulli distribution

    [Marshall and Olkin, 1985] Albert W Marshall and Ingram Olkin. A family of bivariate distributions generated by the bivariate bernoulli distribution. Journal of the American Statistical Association, 80(390):332–338,

  20. [2019]

    Learning intents behind interactions with knowledge graph for recommendation

    [Wang et al., 2021] Xiang Wang, Tinglin Huang, Dingxian Wang, Yancheng Yuan, Zhenguang Liu, Xiangnan He, and Tat-Seng Chua. Learning intents behind interactions with knowledge graph for recommendation. In Proceedings of the web conference 2021, pages 878–887,

  21. [2020]

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

    [He and McAuley, 2016] Ruining He and Julian McAuley. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In proceed- ings of the 25th international conference on world wide web, pages 507–517,

  22. [2021]

    Setrank: A setwise bayesian approach for collaborative ranking in rec- ommender system

    [Wang et al., 2023] Chao Wang, Hengshu Zhu, Chen Zhu, Chuan Qin, Enhong Chen, and Hui Xiong. Setrank: A setwise bayesian approach for collaborative ranking in rec- ommender system. ACM Transactions on Information Sys- tems, 42(2):1–32,

  23. [2022]

    Categorical reparameterization with gumbel-softmax

    [Jang et al., 2016] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144,

  24. [2023]

    Tele-knowledge pre-training for fault analysis

    [Chen et al., 2023] Zhuo Chen, Wen Zhang, Yufeng Huang, Mingyang Chen, Yuxia Geng, Hongtao Yu, Zhen Bi, Yichi Zhang, Zhen Yao, Wenting Song, et al. Tele-knowledge pre-training for fault analysis. In 2023 IEEE 39th Inter- national Conference on Data Engineering (ICDE) , pages 3...

  25. [2024]

    Lkpnr: Llm and kg for personalized news recommendation frame- work

    [Runfeng et al., 2023] Xie Runfeng, Cui Xiangyang, Yan Zhou, Wang Xin, Xuan Zhanwei, Zhang Kai, et al. Lkpnr: Llm and kg for personalized news recommendation frame- work. arXiv preprint arXiv:2308.12028,

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.