Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Integrating LLM-Derived Multi-Semantic Intent into Graph Model for Session-based Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Adding LLM-inferred multi-semantic intents to graph-based session recommenders lifts ranking quality, with top-five rank gains up to 65.33% on ML-1M.

desk verdict Useful LLM+GNN integration for session-based recommendation, but the central attribution to semantic understanding is confounded by a candidate-set loop and needs an ablation before the gains can be trusted. read the letter →

arxiv 2507.20147 v1 pith:3HLTO6OU submitted 2025-07-27 cs.IR

classification cs.IR
keywords session-basedrecommendationgraphneuralnetworkslargelanguagemodelmulti-semanticintentalignmentKLdivergencecandidateitemsetexplicitandlatent
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

Session-based recommenders that use graph neural networks usually read only item IDs, ignoring what the items actually are. This paper argues that a large language model can recover that missing meaning: given a short list of the graph's own top candidate items, the LLM names the user's explicit and latent intents, and those intents are then aligned back into the graph via a KL-divergence loss. The authors show that adding this module to five GNN backbones improves ranking quality on the Beauty and ML-1M datasets; on the GCE-GNN backbone, the true next item's rank in the top five improves by 42.56% on Beauty and 65.33% on ML-1M. If the claim holds, existing graph recommenders can gain semantic understanding without changing their architecture, simply by borrowing an LLM as an intent interpreter.

What carries the argument

The machinery is a two-sided alignment loop. On the structural side, the GNN produces a session embedding $\bar{G}$ from the item-ID graph. On the semantic side, the LLM infers intents from the candidate item titles, BERT encodes them into explicit and latent features $\bar{E}^{e}$ and $\bar{E}^{l}$, and the KL divergence $$L_{\mathrm{info}} = \$\alpha$\,\mathrm{KL}(\bar{G};\bar{E}^{e}) + \$\beta$\,\mathrm{KL}(\bar{G};\bar{E}^{l})$$ pulls the semantic embeddings toward the structural one during joint training. The final session representation is $S = W[\bar{G}, \bar{E}^{e}, \bar{E}^{l}]$, a linear compression of the concatenated views. The detail that makes the loop informative is the candidate-set constraint: the prompt requires every inferred intent to be selected from the GNN's own top-50 items, so the semantic channel can only name items the structural channel already ranked high.

What would settle it

Run the identical pipeline but replace the LLM's semantic inference with a deterministic rule that re-ranks the same top-50 candidates using co-occurrence or popularity scores, keeping BERT and the KL loss intact; if MRR@5 still improves by roughly the reported margin, the causal load is carried by the candidate-selection mechanism rather than by the LLM's semantic understanding.

Watch

Extended reading notes

Core claim

The central claim is that a session's next-item prediction improves when the model is given, alongside the graph's structural representation of item transitions, a semantic representation of the user's intents inferred by an LLM from item titles. The pipeline has three stages: a pre-trained GNN selects the top-50 items for the session and maps them to titles; the Qwen2.5-7B-Instruct model infers the user's intents from the session text and that candidate list, and each intent is marked explicit if it names an item already in the session and latent otherwise; a pre-trained BERT model encodes both intent groups, and a KL-divergence term aligns the semantic embeddings with the GNN's session embedding during joint training. On Beauty and ML-1M the module improves P@5 and MRR@5 for each of the five tested backbones in almost every configuration, with the largest gains on GCE-GNN: MRR@5 rises 42.56% and 65.33% respectively. The paper also reports that gains concentrate at small list sizes and that P@20 can decline, which it attributes to noise introduced by LLM inference on longer lists.

Load-bearing premise

The load-bearing premise is that the top-50 candidate list produced by the pre-trained GNN is a valid and sufficient knowledge base, because the prompt forces the LLM to choose every inferred intent from that list and the alignment loss then trains the GNN toward those choices.

Editorial extensions

If this is right

  • The same inference module can be attached to any existing GNN session recommender without altering the graph architecture, so the reported gains are not tied to one backbone.
  • Because the largest improvements appear at K=5 and shrink as K grows, the semantic channel chiefly sharpens the top of the ranking, which is where session recommenders are deployed.
  • Removing either the explicit or the latent intent stream degrades performance, so the distinction between the two intent classes carries predictive information, not just labeling overhead.
  • KL alignment with a moderate auxiliary weight (around $\sigma=0.2$) is the best fusion rule among the tested alternatives, with contrastive alignment competitive on the ML-1M MRR metric.
  • The occasional P@20 drop implies that LLM intents add signal for short lists but can also add noise, so list-length-aware gating of the semantic channel is a natural next step.

Reading between the lines

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

  • A testable implication not drawn in the paper: because the LLM's intents are constrained to the GNN's own top-50 list, part of the gain may be a self-distillation effect, where the semantic channel re-ranks candidates the structural channel already surfaced rather than finding genuinely new items; a control that replaces the LLM with a deterministic re-ranker over the same candidates would separat
  • The KL loss is one-directional, pulling semantic embeddings toward the graph embedding; a symmetric alignment that also lets semantic intents reshape the graph representation might recover some of the P@20 loss.
  • The larger gains on ML-1M than on Beauty suggest title quality and item heterogeneity modulate the effect, so experiments on datasets with sparse or generic item titles would bound where the method helps.
  • Production latency is not addressed in the paper; caching per-session LLM intents or distilling them into a small encoder would be needed before the approach is practical at scale.
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 / 5 minor

Summary. The paper proposes LLM-DMsRec, a three-stage framework for session-based recommendation. In the first stage, a pre-trained GNN selects a top-K candidate item set from the session ID sequence (Eq. 2). In the second stage, Qwen2.5-7B-Instruct is prompted to infer explicit and latent user intents from the session text and the candidate set (Prompt 1/2, Eqs. 3-7), and BERT encodes these intents. In the third stage, the encoded semantic intents are concatenated with the GNN's structural intent and a KL-divergence alignment loss is used to align the two modalities (Eqs. 8-15). Experiments on Beauty and ML-1M with five GNN backbones report consistent improvements, most notably in MRR@5 (e.g., 42.56% and 65.33% for GCE-GNN on the two datasets), with smaller or sometimes negative changes in P@20.

Significance. If the claimed attribution is valid, the proposed method would be a practical way to inject LLM-derived semantic knowledge into GNN-based session recommenders while controlling LLM hallucination via a candidate knowledge base. The paper reports experiments across five backbone models, includes an ablation study, hyperparameter sensitivity, and a code release, which are valuable. However, the central claim that the performance gains come from the LLM's independent semantic understanding is not established by the current experimental design, because the candidate set constrains the LLM to items already ranked highly by the same GNN that is later trained with the alignment loss. The method may still work as a re-ranking or self-reinforcement mechanism, but the paper does not separate this from genuine semantic grounding. The dimension inconsistency in Eq. (9) and the absence of significance testing further weaken the technical presentation.

major comments (3)
  1. [§4.1–§4.3, Prompt 1 (Requirement 2), Eqs. (2), (5), (10)–(12)] The experimental design is circular with respect to the claim that LLM-derived semantic intent drives the gains. The candidate set is generated by the pre-trained GNN (Eq. 2), and Prompt 1 explicitly requires that “the inferred intents must be selected from the Candidate item set.” The LLM therefore cannot propose any item outside the backbone's own top-50 predictions. The explicit/latent classification (Eq. 5) and the BERT embeddings (Eqs. 6–7) all derive from those same items, and the KL alignment loss (Eqs. 10–12) trains the GNN structural representation to match the distribution of those same candidate items. Consequently, the observed MRR improvements could reflect a re-ranking or self-reinforcement effect within the candidate list rather than independent semantic knowledge. The ablations in Table 3 all keep the same candidate set, so they cannot separate the semantic contribution from the candidate-selection contribution. I would need a control experiment in which the candidate set is generated by a source independent of the backbone (e.g., popularity, random retrieval, or a separate LLM-based retriever), or a comparison that allows the LLM to propose intents outside the top-K and measures whether that changes performance.
  2. [§4.3, Eq. (9)] Equation (9) has a dimension mismatch. It defines S = W [G, E_e, E_l] with W in R^{d×2d}, but the concatenation contains three objects. The text earlier says E_e, E_l ∈ R^{d×v} and G ∈ R^{d×d}, which is also inconsistent with the later use of softmax over S^T v_i in Eq. (13). If G, E_e, E_l are all d-dimensional vectors, W should be R^{d×3d}; if they are matrices, the concatenation and the subsequent dot product must be defined consistently. Please correct the notation and specify the actual tensor shapes used in the implementation.
  3. [§5.2, Table 2] The reported results have no error bars, no multiple runs, and no significance tests. The headline improvements are large (MRR@5 +42.56% and +65.33% for GCE-GNN), while P@20 declines by 2.91% and 4.41% on the two datasets. This pattern is consistent with a re-ranking artifact inside a fixed candidate list. Please report means and standard deviations over several random seeds and run a paired significance test (e.g., bootstrap or paired t-test over sessions). Additionally, report a metric that captures recall-oriented list quality, such as Recall@20 or nDCG@20, to clarify whether the P@20 declines are offset by ranking improvements.
minor comments (5)
  1. [§2.3] The citation for “Guo et al. proposed the LLMGR framework” is given as [25], but reference [25] is Qiao et al., LLM4SBR. Please correct the citation or the text.
  2. [Table 3] The row label “LLM-DMsRec(GEC-GNN)” appears to be a typo; it should be “GCE-GNN”.
  3. [Figure 3] The y-axis label contains a garbled glyph (“/glyph1197umber of Intents”); it should read “Number of Intents”.
  4. [§4.3] The notation for the linear transformations is inconsistent: \(\bar{E}_e = wE_e + b\) and \(\bar{E}_l = wE_l + b\) use the same symbol \(w,b\) but then state \(\bar{E}_e, \bar{E}_l \in \mathbb{R}^{d\times d}\). If the BERT embeddings are pooled to d-dimensional vectors, the output should be d-dimensional; please clarify.
  5. [§3.2, Prompt 1] The requirement “the number of recommended intents should be appropriate” is vague; please specify how the number of intents is determined or constrained in practice.

Circularity Check

3 steps flagged · score 6.0 of 10

The LLM's semantic intents are restricted to the pre-trained GNN's own top-50 candidate set, and the alignment loss trains that same GNN toward those intents, so the reported gains cannot be attributed to independent LLM semantics.

  1. fitted input called prediction [Section 4.1 Eq. (2); Section 3.2 Prompt 1; Section 4.2 Eq. (4)]
    "a pre-trained GNN model is employed to select the top-K highest-ranked items as the candidate item set. This process can be formalized as follows: {I1,I2,...,Ik} = Pre-trained GNN(st, W) ... Requirements: ... 2. The inferred intents must be selected from the Candidate item set."

    The candidate set that bounds the LLM's intents is exactly the pre-trained GNN's top-K output. Prompt 1, requirement 2 forces every inferred intent to be one of those items, so the semantic signal can never contain an item outside the backbone's own ranking. The later fine-tuning (Eqs. 10-15) trains the same GNN to align with BERT embeddings of those very items. The claimed source of improvement, LLM-derived semantic intent, is therefore a re-expression of the pre-trained GNN's candidate list; gains could come from the candidate-selection loop alone, and no control with a non-GNN candidate set is provided.

  2. self definitional [Section 4.2, Eq. (5)]
    "The categorization rule is as follows: C_k = ( explicit intent, if C_k ∈ s_t^T ; latent intent, otherwise )"

    The paper presents explicit versus latent intent as a semantically meaningful distinction and ablates each type, but Eq. (5) defines the split purely by set membership of the inferred item in the session text, not by meaning. Because the inferred intents are already restricted to the GNN's candidate set, explicit and latent are simply the GNN's top-K items that do or do not appear in the session. The Table 3 ablation therefore compares two arbitrary partitions of the same GNN-derived candidate list, making the conclusion that both types matter a property of the definition rather than independent evidence.

1 more flagged steps
  1. self definitional [Section 4.3 Eqs. (10)-(12); Eq. (15)]
    "we introduce the Kullback-Leibler (KL) divergence to measure the difference between the two distributions. ... L_e = KL(G || E_e), L_l = KL(G || E_l) ... L = L_r + σ L_info"

    The alignment loss in Eqs. (10)-(12) optimizes the GNN's structural intent G toward BERT embeddings of intents that, by Prompt 1, must lie in the Eq. (2) candidate set generated by the pre-trained GNN. Equation (15) then jointly trains the GNN with this loss, so the auxiliary target is a function of the pre-trained GNN's own top-50 output. The model is effectively trained to move its session representation toward BERT images of its own recommendations. The large MRR gains with P@20 declines (e.g., GCE-GNN on Beauty: MRR@5 +42.56%, P@20 -2.91%) are consistent with re-ranking inside a fixed candidate list, so the attribution to independent semantic understanding is not identifiable.

full rationale

This is not a case of self-citation: the paper does not rely on its authors' prior theorems, and the final evaluation uses held-out ground-truth clicks, which prevents a score of 8-10. However, the central attribution claim is structurally circular. Eq. (2) defines the candidate item set as the pre-trained GNN's top-K output; Prompt 1, requirement 2 forces all LLM-inferred intents to be selected from that set; Eq. (5) relabels them as explicit or latent by session membership; and Eqs. (10)-(15) align and train the GNN toward BERT embeddings of those same GNN-selected items. The 'semantic' signal is therefore not an independent source of knowledge about user intent; it is a transformation of the backbone's own ranking. Because the ablations in Table 3 all keep the same candidate set, they cannot separate the contribution of LLM semantics from the contribution of the candidate-selection mechanism itself. The reported MRR improvements alongside P@20 declines are the expected signature of re-ranking within a fixed candidate list. The paper may well contain a useful self-distillation technique, but the claim that LLM semantic understanding causes the gains is not established by the experimental design.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities beyond the 'multi-semantic intent' label. The main assumptions are domain-specific: that the GNN candidate set is a valid grounding for semantic inference, that LLM outputs are semantically faithful, and that KL alignment is an appropriate bridge. The free parameters are standard hyperparameters, but the candidate-set size K is particularly consequential because it defines the circularity boundary.

free parameters (5)
  • Candidate set size K = 50
    The number of items passed to the LLM is a hand-chosen hyperparameter that directly bounds the space of possible intents and strongly influences the feedback loop.
  • Explicit intent coefficient alpha = 0.1 (optimal for P@20 on ML-1M)
    Weight for the KL alignment loss on explicit intent, tuned empirically as shown in Figure 6.
  • Latent intent coefficient beta = varied, optimal near 0.7 for MRR@20 on ML-1M
    Weight for the KL alignment loss on latent intent, tuned empirically as shown in Figure 6.
  • Auxiliary loss weight sigma = 0.2 (optimal for MRR on Beauty)
    Scales the contribution of the alignment loss relative to the main recommendation loss, tuned in Figure 5.
  • Embedding dimension d = 100
    The dimension of all embeddings and the GNN representation is a fixed setting, not derived from theory.
assumptions (4)
  • domain assumption The GNN-selected candidate item set is a reliable knowledge base for LLM intent inference.
    Every intent the LLM can return is drawn from the GNN's top-50 predictions, so the entire semantic signal is constrained by the GNN's prior successes or failures. This is introduced in Section 4.1.
  • domain assumption KL divergence can align BERT-encoded semantic intents with GNN structural intents.
    The alignment mechanism in Eq. 10-12 assumes the two embedding spaces can be treated as probability distributions and that matching them helps recommendation. No evidence for this assumption is given beyond the empirical results.
  • domain assumption The LLM's inferred intents, when encoded with BERT, preserve the semantic meaning needed for recommendation.
    The model relies on Qwen2.5-7B to produce useful intents and on BERT to encode them, without validating whether the intermediate representations correspond to genuine user preferences. This is implicit in Sections 4.2 and 4.3.
  • domain assumption Fine-tuning the backbone GNN with the additional alignment loss does not degrade its structural modeling ability.
    The model jointly trains the GNN with the alignment loss and the recommendation loss, assuming the two objectives do not conflict destructively. This is stated in Algorithm 1 and Section 4.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating LLM-Derived Multi-Semantic Intent into Graph Model for Session-based Recommendation." pith.science (2026). https://pith.science/paper/3HLTO6OU

@misc{pith2026250720147,
  author       = {Pith},
  title        = {Pith review of: Integrating LLM-Derived Multi-Semantic Intent into Graph Model for Session-based Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3HLTO6OU}},
  note         = {Machine review of arXiv:2507.20147}
}
read the original abstract

Session-based recommendation (SBR) is mainly based on anonymous user interaction sequences to recommend the items that the next user is most likely to click. Currently, the most popular and high-performing SBR methods primarily leverage graph neural networks (GNNs), which model session sequences as graph-structured data to effectively capture user intent. However, most GNNs-based SBR methods primarily focus on modeling the ID sequence information of session sequences, while neglecting the rich semantic information embedded within them. This limitation significantly hampers model's ability to accurately infer users' true intention. To address above challenge, this paper proposes a novel SBR approach called Integrating LLM-Derived Multi-Semantic Intent into Graph Model for Session-based Recommendation (LLM-DMsRec). The method utilizes a pre-trained GNN model to select the top-k items as candidate item sets and designs prompts along with a large language model (LLM) to infer multi-semantic intents from these candidate items. Specifically, we propose an alignment mechanism that effectively integrates the semantic intent inferred by the LLM with the structural intent captured by GNNs. Extensive experiments conducted on the Beauty and ML-1M datasets demonstrate that the proposed method can be seamlessly integrated into GNNs framework, significantly enhancing its recommendation performance.

Figures

Figures reproduced from arXiv: 2507.20147 by the authors.

Figure 1
Figure 1. a. Traditional GNN-based SBR method; b. Our method; c. An example of a session sequence with [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the LLM-DMsRec model. Phase I employs a pre-trained GNN model to [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. The distribution of explicit and implicit intentions. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Impact of different alignment mechanisms on performance [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Impact of different 𝜎 on performance 0.3 0.4 0.5 0.6 0.7 Latent Coefficient 0.1 0.2 0.3 0.4 Explicit Coefficient 12.96 12.90 12.91 12.90 12.95 12.95 12.76 12.81 12.83 12.85 12.78 12.79 12.71 12.71 12.93 12.66 12.66 12.93 12.79 12.96 0.3 0.4 0.5 0.6 0.7 Latent Coefficie…
Figure 6
Figure 6. Figure 6: Impact of explicit intent coefficient 𝛼 and latent intent coefficient 𝛽 on performance. we designed three alignment mechanisms for the experiment: contrastive learning, InfoNCE[21], and DirectAU[34] [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 31 canonical work pages

  1. [25]

    Shutong Qiao, Chen Gao, Junhao Wen, Wei Zhou, Qun Luo, Peixuan Chen, and Yong Li. 2024. LLM4SBR: A Lightweight and Effective Framework for Integrating Large Language Models in Session-based Recommendation. arXiv preprint arXiv:2402.13840 (2024)

  2. [11]

    Naicheng Guo, Hongwei Cheng, Qianqiao Liang, Linxun Chen, and Bing Han. 2024. Integrating Large Language Models with Graphical Session-Based Recommendation. arXiv preprint arXiv:2402.16539 (2024)

  3. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

  4. [2]

    Susan C Anyosa, João Vinagre, and Alípio M Jorge. 2018. Incremental matrix co-factorization for recommender systems with implicit feedback. In Companion Proceedings of the The Web Conference 2018 . 1413–1418

  5. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al

  6. [4]

    Minjin Choi, Hye-young Kim, Hyunsouk Cho, and Jongwuk Lee. 2024. Multi-intent-aware Session-based Recommen- dation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2532–2536

  7. [5]

    KR1442 Chowdhary and KR Chowdhary. 2020. Natural language processing. Fundamentals of artificial intelligence (2020), 603–649

  8. [6]

    Amine Dadoun and Raphaël Troncy. 2020. Many-to-one recurrent neural network for session-based recommendation. arXiv preprint arXiv:2008.11136 (2020)

Show all 47 references
  1. [7]

    Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

  2. [8]

    Ricardo Dias and Manuel J Fonseca. 2013. Improving music recommendation in session-based collaborative filtering by using temporal context. In 2013 IEEE 25th international conference on tools with artificial intelligence . IEEE, 783–788

  3. [9]

    P Kingma Diederik. 2014. Adam: A method for stochastic optimization. (No Title) (2014)

  4. [10]

    Jiayan Guo, Yaming Yang, Xiangchen Song, Yuan Zhang, Yujing Wang, Jing Bai, and Yan Zhang. 2022. Learning multi-granularity consecutive user intent unit for session-based recommendation. In Proceedings of the fifteenth ACM International conference on web search and data mining...

  5. [12]

    Xiangnan He, Hanwang Zhang, Min-Yen Kan, and Tat-Seng Chua. 2016. Fast matrix factorization for online rec- ommendation with implicit feedback. In Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval . 549–558

  6. [13]

    B Hidasi. 2015. Session-based Recommendations with Recurrent Neural Networks. arXiv preprint arXiv:1511.06939 (2015)

  7. [14]

    Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent neural networks with top-k gains for session-based recommendations. In Proceedings of the 27th ACM international conference on information and knowledge management . 843–852. ACM Trans. Knowl. Discov. Data., Vol. 1, No....

  8. [15]

    Solomon Kullback and Richard A Leibler. 1951. On information and sufficiency. The annals of mathematical statistics 22, 1 (1951), 79–86

  9. [16]

    Jing Li, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Tao Lian, and Jun Ma. 2017. Neural attentive session-based recommendation. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management . 1419– 1428

  10. [17]

    Dawen Liang, Jaan Altosaar, Laurent Charlin, and David M Blei. 2016. Factorization meets the item embedding: Regularizing matrix factorization with item co-occurrence. In Proceedings of the 10th ACM conference on recommender systems. 59–66

  11. [18]

    Anjing Luo, Pengpeng Zhao, Yanchi Liu, Fuzhen Zhuang, Deqing Wang, Jiajie Xu, Junhua Fang, and Victor S Sheng

  12. [19]

    Sichun Luo, Yuxuan Yao, Bowei He, Yinya Huang, Aojun Zhou, Xinyi Zhang, Yuanzhang Xiao, Mingjie Zhan, and Linqi Song. 2024. Integrating large language models into recommendation via mutual augmentation and adaptive aggregation. arXiv preprint arXiv:2401.13870 (2024)

  13. [20]

    Gabriel de Souza P Moreira, Dietmar Jannach, and Adilson Marques da Cunha. 2020. Hybrid session-based news recommendation using recurrent neural networks. arXiv preprint arXiv:2006.13063 (2020)

  14. [21]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

  15. [22]

    Begum Ozbay, Resul Tugay, and Sule Gunduz Oguducu. 2024. A GNN Model with Adaptive Weights for Session-Based Recommendation Systems. In Proceedings of the 2024 9th International Conference on Machine Learning Technologies . 258–264

  16. [23]

    Sung Eun Park, Sangkeun Lee, and Sang-goo Lee. 2011. Session-based collaborative filtering for predicting the next song. In 2011 First ACIS/JNU International Conference on Computers, Networks, Systems and Industrial Engineering . IEEE, 353–358

  17. [24]

    Dunlu Peng and Shuo Zhang. 2022. GC–HGNN: A global-context supported hypergraph neural network for enhancing session-based recommendation. Electronic Commerce Research and Applications 52 (2022), 101129

  18. [26]

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Represen- tation learning with large language models for recommendation. In Proceedings of the ACM on Web Conference 2024 . 3464–3475

  19. [27]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factorizing personalized markov chains for next-basket recommendation. In Proceedings of the 19th international conference on World wide web . 811–820

  20. [28]

    J Ben Schafer, Dan Frankowski, Jon Herlocker, and Shilad Sen. 2007. Collaborative filtering recommender systems. In The adaptive web: methods and strategies of web personalization . Springer, 291–324

  21. [29]

    Guy Shani, David Heckerman, Ronen I Brafman, and Craig Boutilier. 2005. An MDP-based recommender system. Journal of machine Learning research 6, 9 (2005)

  22. [30]

    Shiming Sun, Yuanhe Tang, Zemei Dai, and Fu Zhou. 2019. Self-attention network for session-based recommendation with streaming data input. IEEE Access 7 (2019), 110499–110509

  23. [31]

    Zhu Sun, Hongyang Liu, Xinghua Qu, Kaidong Feng, Yan Wang, and Yew Soon Ong. 2024. Large language models for intent-driven session recommendations. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 324–334

  24. [32]

    Yong Kiam Tan, Xinxing Xu, and Yong Liu. 2016. Improved recurrent neural networks for session-based recommenda- tions. In Proceedings of the 1st workshop on deep learning for recommender systems . 17–22

  25. [33]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

  26. [34]

    Chenyang Wang, Yuanqing Yu, Weizhi Ma, Min Zhang, Chong Chen, Yiqun Liu, and Shaoping Ma. 2022. Towards representation alignment and uniformity in collaborative filtering. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining . 1816–1825

  27. [35]

    Jianling Wang, Kaize Ding, Ziwei Zhu, and James Caverlee. 2021. Session-based recommendation with hypergraph attention networks. In Proceedings of the 2021 SIAM international conference on data mining (SDM) . SIAM, 82–90

  28. [36]

    Meirui Wang, Pengjie Ren, Lei Mei, Zhumin Chen, Jun Ma, and Maarten De Rijke. 2019. A collaborative session- based recommendation approach with parallel memory modules. In Proceedings of the 42nd international ACM SIGIR conference on research and development in information ret...

  29. [37]

    Xinyuan Wang, Liang Wu, Liangjie Hong, Hao Liu, and Yanjie Fu. 2024. LLM-Enhanced User-Item Interactions: Leveraging Edge Information for Optimized Recommendations. arXiv preprint arXiv:2402.09617 (2024). ACM Trans. Knowl. Discov. Data., Vol. 1, No. 1, Article . Publication da...

  30. [38]

    Ziyan Wang, Yingpeng Du, Zhu Sun, Haoyan Chua, Kaidong Feng, Wenya Wang, and Jie Zhang. 2024. Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation. arXiv preprint arXiv:2403.16427 (2024)

  31. [39]

    Ziyang Wang, Wei Wei, Gao Cong, Xiao-Li Li, Xian-Ling Mao, and Minghui Qiu. 2020. Global context enhanced graph neural networks for session-based recommendation. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval ...

  32. [40]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. In Proceedings of the AAAI conference on artificial intelligence , Vol. 33. 346–353

  33. [41]

    Xin Xia, Hongzhi Yin, Junliang Yu, Qinyong Wang, Lizhen Cui, and Xiangliang Zhang. 2021. Self-supervised hyper- graph convolutional networks for session-based recommendation. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 4503–4511

  34. [42]

    Feng Yu, Yanqiao Zhu, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. 2020. TAGNN: Target attentive graph neural networks for session-based recommendation. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval . 1921–1924

  35. [43]

    Jiahao Yuan, Zihan Song, Mingyou Sun, Xiaoling Wang, and Wayne Xin Zhao. 2021. Dual sparse attention network for session-based recommendation. In Proceedings of the AAAI conference on artificial intelligence , Vol. 35. 4635–4643

  36. [44]

    Peiyan Zhang, Jiayan Guo, Chaozhuo Li, Yueqi Xie, Jae Boum Kim, Yan Zhang, Xing Xie, Haohan Wang, and Sunghun Kim. 2023. Efficiently leveraging multi-level user intent for session-based recommendation via atten-mixer network. In Proceedings of the sixteenth ACM international c...

  37. [45]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 1435–1448. ACM...

  38. [2020]

    In IJCAI

    Collaborative Self-Attention Network for Session-based Recommendation.. In IJCAI. 2591–2597

  39. [2023]

    arXiv preprint arXiv:2309.16609 (2023)

    Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)

Pith tools

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