arxiv: 2604.06928 · v2 · submitted 2026-04-08 · 💻 cs.IR

Recognition: 2 theorem links

· Lean Theorem

Leveraging LLMs and Heterogeneous Knowledge Graphs for Persona-Driven Session-Based Recommendation

Muskan Gupta , Suraj Thapa , Jyotsana Khatri

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:56 UTC · model grok-4.3

classification 💻 cs.IR

keywords session-based recommendationheterogeneous knowledge graphsuser personaslarge language modelspersonalizationgraph infomaxrerankingAmazon datasets

0 comments

The pith

A framework that infers latent user personas from heterogeneous knowledge graphs and LLM embeddings improves session-based recommendations over models using only session history.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that session-based recommenders can gain meaningful personalization by extracting latent user personas in an unsupervised way from a heterogeneous knowledge graph built with time-independent interactions, item relations, features, and DBpedia metadata. Item embeddings from large language models initialize the graph, after which Heterogeneous Deep Graph Infomax learns persona representations that are then fed into a two-stage system: first a modified sequential model produces a candidate set, then the base model reranks to preserve short-term session intent. A sympathetic reader would care because anonymous sessions normally limit personalization in sparse or cold-start settings, and the method replaces reliance on text-only user descriptions or pure sequence modeling with structured relational signals from the graph. Experiments on Amazon Books and Amazon Movies & TV datasets show consistent gains over sequential baselines that use only session-derived user embeddings.

Core claim

The central claim is that latent user personas learned unsupervised via Heterogeneous Deep Graph Infomax over a heterogeneous knowledge graph initialized with LLM-derived item embeddings can be integrated into a data-driven session-based recommender to produce a candidate set that, after reranking by the base sequential model, yields higher accuracy than sequential models equipped only with user embeddings derived from session history.

What carries the argument

The two-stage architecture of personalized information extraction (constructing a heterogeneous KG from user-item interactions, item-item relations, item-feature associations and DBpedia metadata, then learning personas with HDGI) followed by personalized information utilization (incorporating persona representations and LLM item embeddings into a modified SBRS for candidate generation and reranking).

If this is right

The method yields consistent accuracy gains over sequential models that rely solely on session-derived user embeddings.
It supplies personalization even under sparse or cold-start conditions by drawing on time-independent interactions stored in the heterogeneous graph.
Grounding persona modeling in structured relational signals from the knowledge graph outperforms approaches that use only text-based user representations.
The final reranking step preserves short-term session intent while incorporating the longer-term persona information.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same KG-plus-LLM initialization pattern could be tested on other recommendation domains such as music playlists or news feeds where external metadata is available.
If the learned personas form stable clusters, they might support downstream tasks like user segmentation or explanation generation without additional supervision.
Replacing the HDGI step with a supervised persona predictor trained on explicit user attributes would provide a direct test of whether unsupervised relational learning is necessary.

Load-bearing premise

That the unsupervised personas extracted from the heterogeneous knowledge graph supply personalization signals that meaningfully exceed those already available from session history or simple user embeddings and that these signals transfer usefully into the downstream reranking stage.

What would settle it

An ablation on the Amazon Books or Movies & TV datasets in which removing the persona representations or replacing them with random vectors produces no improvement or a performance drop relative to the session-history baseline would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.06928 by Jyotsana Khatri, Muskan Gupta, Suraj Thapa.

**Figure 1.** Figure 1: An overview of our proposed framework: Persona-driven session based recommendation via knowledge graph [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

**Figure 2.** Figure 2: Construction of KG tains multiple node types, including User, Movie, Person, Subject, Genre, Award, OpeningTheme, LiteraryGenre, and Category for the Amazon Movies & TV 2 https://cseweb.ucsd.edu/~jmcauley/datasets/amazon_v2/ 3 https://github.com/WangYuhan-0520/Amazon-KG-v2.0-dataset [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

read the original abstract

Session-based recommendation systems (SBRS) aim to capture user's short-term intent from interaction sequences. However, the common assumption of anonymous sessions limits personalization, particularly under sparse or cold-start conditions. Recent advances in LLM augmented recommendation have shown that LLMs can generate rich item representations, but modeling user personas with LLMs remains challenging due to anonymous sessions. In this work, we propose a persona driven SBRS framework that explicitly models latent user personas inferred from a heterogeneous knowledge graph (KG) and integrates them into a data-driven SBRS. Our framework adopts a two-stage architecture consisting of personalized information extraction and personalized information utilization. In the personalized information extraction stage, we construct a heterogeneous KG that integrates time-independent user-item interactions, item-item relations, item-feature associations, and external metadata from DBpedia. We then learn latent user personas in an unsupervised manner using a Heterogeneous Deep Graph Infomax (HDGI) objective over a KG initialized with LLM-derived item embeddings. In the personalized information utilization stage, the learned persona representations together with LLM-derived item embeddings are incorporated into a modified architecture of data-driven SBRS to generate a candidate set of relevant items, followed by reranking using the base sequential model to emphasize short-term session intent. Unlike prior approaches that rely solely on sequence modeling or text-based user representations, our method grounds user persona modeling in structured relational signals derived from a heterogeneous KG. Experiments on Amazon Books and Amazon Movies & TV demonstrate that our approach consistently improves over sequential models with user embeddings derived using session history.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The two-stage KG-plus-LLM persona extraction is a reasonable architecture but the method does not explain how to obtain a persona for an anonymous session without a user ID.

read the letter

The main pitch is a two-stage setup: first build a heterogeneous KG from user-item edges, item relations, and DBpedia metadata, seed item embeddings with an LLM, run HDGI to extract unsupervised user persona vectors, then feed those vectors plus the item embeddings into a modified session-based model that generates candidates and reranks them with the base sequential model to preserve short-term intent. That combination is the concrete new piece relative to pure sequence models or text-only user representations. The architecture itself is laid out clearly enough and the motivation for bringing in structured relational signals from the KG makes sense on paper. The experiments are claimed to show consistent gains on Amazon Books and Amazon Movies & TV over session-history user embeddings, which at least points to some empirical signal worth checking. The soft spot is exactly the one the stress-test note flags. The KG explicitly includes user-item interactions and HDGI produces user-specific embeddings, yet the target setting is anonymous sessions where no user ID is available. The description gives no inference route for pulling or imputing a persona vector from a fresh session alone, so the reported improvements on datasets that retain user IDs do not actually test the cold-start anonymous case the abstract highlights. Without that link the central claim about better personalization under sparsity does not hold up. The lack of any experimental protocol, baselines, metrics, or ablations in the abstract makes the numbers hard to evaluate either way. This is for people already working on hybrid KG-LLM recommenders who want to see one concrete way to inject long-term persona signals into session models. A reader in that niche could extract the architecture and try to fix the inference gap themselves. I would send it to peer review rather than desk-reject so the authors can address the anonymous-session path and supply the missing experimental details.

Referee Report

2 major / 1 minor

Summary. The paper proposes a two-stage persona-driven SBRS framework. In the first stage, a heterogeneous KG is built from time-independent user-item interactions, item relations, features, and DBpedia metadata; LLM-derived item embeddings initialize the graph, and unsupervised HDGI learns latent user personas. In the second stage, these personas are combined with LLM item embeddings inside a modified data-driven SBRS to produce candidates that are then reranked by a base sequential model to preserve short-term intent. The central claim is that this KG-grounded personalization consistently outperforms sequential models that derive user embeddings only from session history, as demonstrated on Amazon Books and Amazon Movies & TV.

Significance. If the persona inference mechanism works for anonymous sessions and the reported gains are reproducible, the approach would offer a concrete way to inject structured relational signals into SBRS without relying solely on sequence data, which could help cold-start and sparse-session regimes. The two-stage separation of persona extraction (HDGI on LLM-initialized KG) from session reranking is a clean architectural choice that could be adopted more broadly.

major comments (2)

[personalized information extraction stage] The abstract and framework description emphasize applicability to anonymous sessions, yet the KG explicitly incorporates time-independent user-item edges and HDGI produces user-specific persona embeddings. No inference procedure is supplied for obtaining a persona vector from a new session that lacks a known user ID. This omission directly affects the central claim that the method supplies personalization signals beyond what session history alone can provide.
[Experiments] The abstract states that experiments on Amazon Books and Amazon Movies & TV demonstrate consistent improvements, but supplies no experimental protocol, baseline definitions, evaluation metrics, statistical tests, ablation studies, or error analysis. Without these elements the empirical support for the central performance claim cannot be assessed.

minor comments (1)

The abstract is dense; a short sentence clarifying how the persona vector is obtained at inference time for an anonymous session would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of applicability to anonymous sessions and the need for clearer experimental details. We address each major comment below and will revise the manuscript to strengthen these areas.

read point-by-point responses

Referee: [personalized information extraction stage] The abstract and framework description emphasize applicability to anonymous sessions, yet the KG explicitly incorporates time-independent user-item edges and HDGI produces user-specific persona embeddings. No inference procedure is supplied for obtaining a persona vector from a new session that lacks a known user ID. This omission directly affects the central claim that the method supplies personalization signals beyond what session history alone can provide.

Authors: We agree this is a valid point and that the current description does not explicitly detail the inference step for completely anonymous sessions without prior user history. The KG and HDGI are trained on historical data to extract personas, but for new sessions we intend to infer personas by projecting the session items onto the pre-trained KG (using item embeddings and relations) and computing a session-level persona via aggregation or a lightweight forward pass. We will add a dedicated subsection under 'Personalized Information Extraction' describing this inference procedure, including pseudocode and how it avoids reliance on user IDs while still leveraging the KG structure. revision: yes
Referee: [Experiments] The abstract states that experiments on Amazon Books and Amazon Movies & TV demonstrate consistent improvements, but supplies no experimental protocol, baseline definitions, evaluation metrics, statistical tests, ablation studies, or error analysis. Without these elements the empirical support for the central performance claim cannot be assessed.

Authors: The full manuscript contains an Experiments section that defines the datasets, baselines (including session-history-only sequential models), metrics (HR@K, NDCG@K), and reports the gains. However, we acknowledge the abstract is too high-level and that additional ablations, statistical tests, and error analysis would strengthen the paper. We will revise by (1) expanding the abstract with a brief mention of the evaluation protocol and (2) adding a new subsection with ablations on the persona component, significance testing, and qualitative error analysis in the revised version. revision: yes

Circularity Check

0 steps flagged

No circularity: two-stage KG-to-SBRS pipeline is empirically validated and independent of final metrics

full rationale

The derivation proceeds by constructing a heterogeneous KG (user-item edges + item relations + DBpedia metadata), initializing item nodes with external LLM embeddings, learning per-user persona vectors via an unsupervised HDGI objective, then feeding those fixed persona vectors plus LLM item embeddings into a modified sequential SBRS for candidate generation followed by reranking. All performance numbers are obtained from held-out test sessions on Amazon Books and Movies & TV; no equation equates the reported NDCG/HR to the HDGI loss, to the persona dimension, or to any fitted parameter by algebraic identity. The architecture therefore remains falsifiable by external benchmarks and does not reduce to a renaming or self-referential fit.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The framework rests on the ability of the constructed KG plus HDGI to produce useful personas and on the assumption that these personas add value when injected into the SBRS; no explicit free parameters are named, but training of HDGI and the downstream model necessarily involves many.

free parameters (1)

HDGI hyperparameters and embedding dimensions
Unsupervised graph infomax parameters and persona vector size are chosen during training but not reported.

axioms (2)

domain assumption A heterogeneous KG integrating time-independent user-item interactions, item-item relations, item-feature associations, and DBpedia metadata can support unsupervised persona discovery.
Invoked in the personalized information extraction stage.
domain assumption LLM-derived item embeddings provide suitable initial node features for the HDGI objective on the KG.
Used to initialize the graph before persona learning.

invented entities (1)

latent user personas no independent evidence
purpose: To supply personalization signals for anonymous sessions.
Inferred unsupervised from the KG; no ground-truth personas or external validation are mentioned.

pith-pipeline@v0.9.0 · 5576 in / 1564 out tokens · 39737 ms · 2026-05-10T17:56:04.428759+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We then learn latent user personas in an unsupervised manner using a Heterogeneous Deep Graph Infomax (HDGI) objective over a KG initialized with LLM-derived item embeddings.
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The heterogeneous knowledge graph contains multiple node types, including User, Movie, Person, Subject, Genre...

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages · 2 internal anchors

[1]

In: Duh, K., Gomez, H., Bethard, S

Agrawal, G., Kumarage, T., Alghamdi, Z., Liu, H.: Can knowledge graphs re- duce hallucinations in LLMs? : A survey. In: Duh, K., Gomez, H., Bethard, S. (eds.) Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol- ume 1: Long Papers). pp. 3947–3960. Association f...

work page doi:10.18653/v1/2024.naacl-long.219 2024
[2]

Session-based Recommendations with Recurrent Neural Networks

Hidasi, B., Karatzoglou, A., Baltrunas, L., Tikk, D.: Session-based recommenda- tions with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015)

work page internal anchor Pith review arXiv 2015
[3]

ISBN 9781450370233

Hu, Z., Dong, Y., Wang, K., Sun, Y.: Heterogeneous graph transformer. In: Proceedings of The Web Conference 2020 (WWW’20). pp. 2704–2710. ACM, Taipei, Taiwan (2020). https://doi.org/10.1145/3366423.3380027,https://arxiv. org/abs/2003.01332

work page doi:10.1145/3366423.3380027 2020
[4]

Self-attentive sequential recommendation

Kang, W., McAuley, J.: Self-attentive sequential recommendation. In: 2018 IEEE International Conference on Data Mining (ICDM). pp. 197–206. IEEE (2018). https://doi.org/10.1109/ICDM.2018.00035,https://arxiv.org/abs/1808.09781

work page doi:10.1109/icdm.2018.00035 2018
[5]

In: Proceedings of the 39th ACM/SIGAPP Symposium on Applied Computing (SAC ’24)

Kwon, J., Ahn, S., Seo, Y.D.: Reckg: Knowledge graph for recommender systems. In: Proceedings of the 39th ACM/SIGAPP Symposium on Applied Computing (SAC ’24). pp. 489–496. Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3605098.3636009

work page doi:10.1145/3605098.3636009 2024
[6]

In: Proceedings of the 26th ACM International Conference on Information and Knowledge Management (CIKM’17)

Li, J., Ren, P., Chen, Z., Ren, Z., Lian, T., Ma, J.: Neural attentive session-based recommendation. In: Proceedings of the 26th ACM International Conference on Information and Knowledge Management (CIKM’17). pp. 1419–1428. ACM, Sin- gapore (2017). https://doi.org/10.1145/3132847.3132926 14 Gupta et al

work page doi:10.1145/3132847.3132926 2017
[7]

arXiv preprint arXiv:2511.17467 (2025), https://arxiv.org/abs/2511.17467

Liang, S., Zhang, Y., Guo, Y.: Personaagent with graphrag: Community-aware knowledge graphs for personalized llm. arXiv preprint arXiv:2511.17467 (2025), https://arxiv.org/abs/2511.17467

work page arXiv 2025
[8]

ACM Transactions on Information Systems43(2), 1–47 (2025)

Lin, J., Dai, X., Xi, Y., Liu, W., Chen, B., Zhang, H., Liu, Y., Wu, C., Li, X., Zhu, C., et al.: How can recommender systems benefit from large language models: A survey. ACM Transactions on Information Systems43(2), 1–47 (2025)

2025
[9]

In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

Lin, X., Wang, W., Li, Y., Feng, F., Ng, S.K., Chua, T.S.: Bridging items and language: A transition paradigm for large language model-based recommendation. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. p. 1816–1826. KDD ’24, Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/...

work page doi:10.1145/3637528.3671884 2024
[10]

In: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval

Liu, J., Yan, X., Li, D., Zhang, G., Gu, H., Zhang, P., Lu, T., Shang, L., Gu, N.: Improving llm-powered recommendations with personalized information. In: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2560–2565 (2025)

2025
[11]

In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discov- ery and Data Mining (KDD’18)

Liu, Q., Zeng, Y., Mokhosi, R., Zhang, H.: STAMP: Short-term atten- tion/memory priority model for session-based recommendation. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discov- ery and Data Mining (KDD’18). pp. 1831–1839. ACM, London, UK (2018). https://doi.org/10.1145/3219819.3219950

work page doi:10.1145/3219819.3219950 2018
[12]

Advances in Neural Information Processing Systems36, 30821–30839 (2023)

Liu, X., Li, Z., Gao, Y., Yang, J., Cao, T., Wang, Z., Yin, B., Song, Y.: Enhanc- ing user intent capture in session-based recommendation with attribute patterns. Advances in Neural Information Processing Systems36, 30821–30839 (2023)

2023
[13]

In: Duh, K., Gomez, H., Bethard, S

Lyu, H., Jiang, S., Zeng, H., Xia, Y., Wang, Q., Zhang, S., Chen, R., Le- ung, C., Tang, J., Luo, J.: LLM-rec: Personalized recommendation via prompt- ing large language models. In: Duh, K., Gomez, H., Bethard, S. (eds.) Find- ings of the Association for Computational Linguistics: NAACL 2024. pp. 583–

2024
[14]

https://doi.org/10.18653/v1/2024.findings-naacl.39,https://aclanthology.org/ 2024.findings-naacl.39/

Association for Computational Linguistics, Mexico City, Mexico (Jun 2024). https://doi.org/10.18653/v1/2024.findings-naacl.39,https://aclanthology.org/ 2024.findings-naacl.39/

work page doi:10.18653/v1/2024.findings-naacl.39 2024
[15]

arXiv preprint arXiv:2508.20587 (2025)

Narwariya, J., Gupta, P., Gupta, M., Khatri, J., Vig, L.: Semsr: Semantics aware robust session-based recommendations. arXiv preprint arXiv:2508.20587 (2025)

work page arXiv 2025
[16]

https://arxiv.org/abs/1911.08538

Ren, Y., Liu, B., Huang, C., Dai, P., Bo, L., Zhang, J.: Heterogeneous deep graph infomax. arXiv preprint arXiv:1911.08538 (2019)

work page arXiv 1911
[17]

Earthquake shakes Twitter users: real-time event detection by social sensors , booktitle =

Rendle, S., Freudenthaler, C., Schmidt-Thieme, L.: Factorizing personalized markov chains for next-basket recommendation. In: Proceedings of the 19th Inter- national Conference on World Wide Web (WWW’10). pp. 811–820. ACM, Raleigh, NC, USA (2010). https://doi.org/10.1145/1772690.1772773,https://doi.org/ 10.1145/1772690.1772773

work page doi:10.1145/1772690.1772773 2010
[18]

InProceedings of the 28th ACM International Conference on Information and Knowledge Management(Beijing, China)(CIKM ’19)

Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W., Jiang, P.: BERT4Rec: Sequen- tial recommendation with bidirectional encoder representations from transformer. In: Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM’19). pp. 1441–1450. ACM, Beijing, China (2019). https://doi.org/10.1145/3357384.3357895,http...

work page doi:10.1145/3357384.3357895 2019
[19]

In: Proceedings of the 11th ACM International Con- ference on Web Search and Data Mining (WSDM’18)

Tang, J., Wang, K.: Personalized top-N sequential recommendation via convolu- tional sequence embedding. In: Proceedings of the 11th ACM International Con- ference on Web Search and Data Mining (WSDM’18). pp. 565–573. ACM, Ma- rina del Rey, CA, USA (2018). https://doi.org/10.1145/3159652.3159656,https: //arxiv.org/abs/1809.07426 Persona-Driven SBRS with L...

work page doi:10.1145/3159652.3159656 2018
[20]

Team, Q.: Qwen3 technical report (2025),https://arxiv.org/abs/2505.09388

work page internal anchor Pith review Pith/arXiv arXiv 2025
[22]

Deep Graph Infomax

Veličković, P., Fedus, W., Hamilton, W.L., Liò, P., Bengio, Y., Hjelm, R.D.: Deep graph infomax. In: International Conference on Learning Representations (ICLR) — Poster (2019),https://arxiv.org/abs/1809.10341

work page Pith review arXiv 2019
[23]

In: Proceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM’18)

Wang, H., Zhang, F., Wang, J., Zhao, M., Li, W., Xie, X., Guo, M.: Ripplenet: Propagating user preferences on the knowledge graph for recommender systems. In: Proceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM’18). pp. 417–426. ACM, Torino, Italy (2018). https://doi.org/10.1145/3269206.3271739,https://arxiv....

work page doi:10.1145/3269206.3271739 2018
[24]

In: The World Wide Web Con- ference (WWW’19)

Wang, H., Zhao, M., Xie, X., Li, W., Guo, M.: Knowledge graph convolu- tional networks for recommender systems. In: The World Wide Web Con- ference (WWW’19). pp. 3307–3313. ACM, San Francisco, CA, USA (2019). https://doi.org/10.1145/3308558.3313417,https://arxiv.org/abs/1904.12575

work page doi:10.1145/3308558.3313417 2019
[25]

In: Proceed- ings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

Wang, S., Fan, W., Feng, Y., Shanru, L., Ma, X., Wang, S., Yin, D.: Knowledge graph retrieval-augmented generation for llm-based recommendation. In: Proceed- ings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 27152–27168 (2025)

2025
[26]

In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD’19)

Wang, X., He, X., Cao, Y., Liu, M., Chua, T.: KGAT: Knowledge graph attention network for recommendation. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD’19). pp. 950–958. ACM, Anchorage, AK, USA (2019). https://doi.org/10.1145/3292500.3330989,https://arxiv.org/abs/1905.07854

work page doi:10.1145/3292500.3330989 2019
[27]

In: Proceedings of The Web Conference 2019 (WWW’19)

Wang, X., Ji, H., Shi, C., Wang, B., Cui, P., Yu, P.S., Ye, Y.: Hetero- geneous graph attention network. In: Proceedings of The Web Conference 2019 (WWW’19). pp. 2022–2032. ACM, San Francisco, CA, USA (2019). https://doi.org/10.1145/3308558.3313562,https://arxiv.org/abs/1903.07293

work page doi:10.1145/3308558.3313562 2019
[28]

In: Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval

Wang, Y., Xie, Q., Tang, M., Li, L., Yuan, J., Liu, Y.: Amazon-kg: A knowledge graph enhanced cross-domain recommendation dataset. In: Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. pp. 123–130 (2024)

2024
[29]

World Wide Web27(5), 60 (2024)

Wu, L., Zheng, Z., Qiu, Z., Wang, H., Gu, H., Shen, T., Qin, C., Zhu, C., Zhu, H., Liu, Q., et al.: A survey on large language models for recommendation. World Wide Web27(5), 60 (2024)

2024
[30]

IEEE Transactions on Knowledge and Data Engineering (2025)

Zhang, X., Xu, B., Li, C., He, B., Lin, H., Ma, C., Ma, F.: A survey on side information-driven session-based recommendation: From a data-centric perspec- tive. IEEE Transactions on Knowledge and Data Engineering (2025)

2025
[31]

arXiv preprint arXiv:2402.17129 (2024)

Zhang, X., Xu, B., Li, C., Zhou, Y.: Side information-driven session-based recom- mendation: A survey. arXiv preprint arXiv:2402.17129 (2024)

work page arXiv 2024
[32]

Zhang, X., Xu, B., Ma, F., Wang, Z., Yang, L., Lin, H.: Rethinking contrastive learninginsession-basedrecommendation.PatternRecognition169,111924(2026)

2026