Pith. sign in

REVIEW 4 major objections 4 minor 57 references

Intent-Interest Disentanglement and Item-Aware Intent Contrastive Learning for Sequential Recommendation

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read IDCLRec claims that separating stable user tastes from dynamic intents improves sequential recommendation, and that the residual after causal cross-attention is the right place to model intents.

desk verdict A well-built empirical model that likely beats SOTA on sequential recommendation, but the headline margin is not statistically pinned down and the disentanglement evidence is self-referential. read the letter →

arxiv 2501.07096 v1 pith:QB4YXMQH submitted 2025-01-13 cs.IR

Yijin Choi , Chiehyeon Lim This is my paper · ORCID
classification cs.IR
keywords sequentialrecommendationintentmodelinginterestdisentanglementcausalcross-attentionitem-awarecontrastivelearningimportance-weightedattentionnext-itemprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that sequential recommendation models should stop treating user interaction histories as a single behavioral signal. It argues that those histories mix stable personal tastes with evolving motivations, and that only by separating them can a model understand and predict user choices. The proposed model, IDCLRec, performs the separation by using causal cross-attention to extract interests and treating everything left over as intents, then improves intents with item-aware contrastive learning. The paper reports that this design outperforms all compared baselines on three real-world e-commerce datasets, with gains in hit rate and ranking quality that are largest on the sparsest dataset. If correct, the result would make two-factor behavior decomposition a standard component of intent-based sequential recommenders.

What carries the argument

The load-bearing object is the residual identity $I^u = H^u - R^u$, which defines intents as whatever remains after the causal cross-attention module extracts stable interests from the sequence representations. The causal cross-attention uses a causal mask so each current behavior attends only to previous behaviors, and layer normalization after the residual connection retains only features shared between current and previous steps, forcing $R^u$ to encode consistent patterns. The similarity-adjustment loss $\mathcal{L}_d = \frac{1}{N-1}\sum_{t=1}^{N-1}(1 - \mathrm{sigmoid}(\mathrm{sim}(i^u_t, i^u_{t+1})))$ enforces gradual intent drift, and the importance-weighted attention module selects intents whose similarity to the most recent intent exceeds a threshold $\delta$, computes sigmoid relevance scores between them, and softmax-normalizes summed relevance to weight each intent's contribution to the final categorical intent vector. Two InfoNCE contrastive objectives connect intents to items: $\mathcal{L}_{CL1}$ aligns intents of sequences sharing the same target item, and $\mathcal{L}_{CL2}$ aligns each intent with the averaged embedding of items interacted with under that intent. These mechanisms together carry the argument from raw sequences to the final user representation $h^u = r_N^u + i^u$.

What would settle it

Generate synthetic sequences from a known generative process in which each user has a fixed interest vector and a piecewise-constant intent vector that switches at known timestamps; train IDCLRec and measure whether the residual intent representations are more similar within an intent block than across blocks and whether the inferred switch points align with the injected ones. If the residual does not track the injected intents, the claim that $I^u = H^u - R^u$ encodes dynamic intent is refuted.

Watch

Extended reading notes

Core claim

IDCLRec's central claim is that user behavior representations are composed of two separable signals: stable interests and dynamic intents. The model extracts interests with a causal cross-attention mechanism that aggregates only previous behaviors relevant to the current one, then adds the current behavior back with layer normalization to form $R^u = \mathrm{LayerNorm}(\hat{R}^u + H^u)$. Intents are defined as the residual $I^u = H^u - R^u$. A similarity-adjustment loss makes adjacent intents evolve smoothly, and an importance-weighted attention module aggregates previous intents most similar to the most recent intent into a user-specific categorical intent without any preset number of categories. Two item-aware contrastive losses further sharpen this intent representation: one aligns intents from sequences whose next target item matches, and the other pulls an intent toward the average embedding of the items interacted with under that same intent. On the Sports, Beauty, and Toys datasets, the paper reports the highest hit rate and ranking quality across all baselines, with 5.47% to 14.62% improvements in HR and 4.75% to 12.70% improvements in NDCG over the best baseline.

Load-bearing premise

The load-bearing premise is that subtracting the extracted stable-interest component from behavior vectors leaves genuine dynamic intents; if that residual is mostly noise or still contains interest information, the performance gains could come from extra training regularization rather than from true intent disentanglement.

Editorial extensions

If this is right

  • If the paper is right, models that claim to capture user intents should be compared against a two-factor baseline that separates stable interests from dynamic intents, since the ablations attribute a large part of the gain to this separation.
  • Fixed global numbers of intent categories can be replaced by per-user similarity-based aggregation; the paper shows performance remains stable across thresholds, suggesting the threshold itself may be removable.
  • Explicitly tying intents to the items that express them through contrastive learning becomes a default ingredient: both same-target-item alignment and intent-to-item-combination alignment improve ranking in the reported experiments.
  • The larger relative gains on the sparser Sports dataset suggest that the disentanglement and item-aware contrastive objectives help most when user histories are short, a setting where sequential recommenders usually struggle.

Reading between the lines

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

  • Inference: the causal cross-attention mechanism can be viewed as a smoothing operation over behavior vectors, so part of the improvement attributed to disentanglement may be regularization that stabilizes representations rather than a faithful split into taste and intent.
  • Inference: a decisive external test of the split would use synthetic sequences generated from known piecewise-constant intents and fixed user interest vectors, checking whether the residual intent representations switch at the injected points; such validation is absent from the paper.
  • Inference: the same interest/intent decomposition could be transplanted to session-based recommendation, where there is no persistent user identity, by treating session-level recurring themes as interests and transition-level purposes as intents.
  • Inference: the threshold $\delta$ may be unnecessary; the paper's own $\delta=0$ test shows nearly identical performance, so dropping the threshold and keeping only importance-weighted attention would produce a simpler model with one fewer hyperparameter.
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

4 major / 4 minor

Summary. The paper proposes IDCLRec, a sequential recommendation model that disentangles user behavior representations into stable interests (extracted by a causal cross-attention mechanism) and dynamic intents (modeled as the residual), aggregates similar intents into a user-level categorical intent via importance-weighted attention, and applies two item-aware contrastive losses. The model is evaluated on three Amazon datasets (Sports, Beauty, Toys) against eleven baselines, reporting improvements in HR and NDCG across all settings, with ablations supporting each component.

Significance. If the performance results are reliable, IDCLRec would be a strong new state-of-the-art on these standard benchmarks, and the intent-interest disentanglement idea is a useful conceptual contribution. The paper provides a thorough experimental protocol: three datasets, eleven baselines, ablation studies, and hyperparameter sensitivity analyses, and the method is clearly described. However, the lack of statistical significance testing and the self-referential nature of the disentanglement analysis limit the strength of the central claims.

major comments (4)
  1. [Section 5.1.4 and Table 3] The claim that IDCLRec 'significantly outperforms' baselines is not statistically substantiated. The paper reports only three-seed averages without standard deviations, confidence intervals, or significance tests. The smallest margin—Beauty NDCG@5, 0.0507 vs. 0.0484 (4.75% relative)—is small enough that it could plausibly lie within run-to-run variance for transformer-based recommenders. Please report per-seed values, include error bars in Tables 3 and 4, and provide paired significance tests (e.g., bootstrap or t-test) across all datasets and metrics. Without this, the central SOTA claim is under-supported.
  2. [Section 5.5 and Eq. (7)] The t-SNE visualization does not independently validate the disentanglement of intents and interests. Because I^u is constructed as the residual H^u - R^u and R^u is produced by a different attention mechanism, the two representations are expected to occupy different regions of the latent space by construction. The figure therefore shows that the model separates these two vectors, but not that they correspond to semantically distinct user characteristics (dynamic motivations vs. stable tastes). Please provide external validation, such as evaluating the temporal stability of interest representations, probing intent representations with known item-category or time-window signals, or comparing against a control in which the residual is replaced by a random vector.
  3. [Eq. (20)] The training loss for next-item prediction is written as L_rec = -exp(y_hat_g) / sum_i exp(y_hat_i), which is the negative softmax probability, not the cross-entropy loss. The correct cross-entropy is -log(exp(y_hat_g) / sum_i exp(y_hat_i)). As written, the loss is not a proper classification loss and would not behave as standard cross-entropy during training. Please correct the equation (and verify that the implementation matches the corrected version), or clarify if a different loss is intended.
  4. [Tables 3 and 4] The ablation results are also reported as single numbers without variance. The conclusions in Section 5.3, such as the relative contributions of L_d, L_CL1, and L_CL2, would be more convincing with error bars or significance tests, given that the reported differences between some ablations (e.g., IDCLRec vs. A on Sports NDCG@20: 0.0426 vs. 0.0425) are very small. Please include per-seed standard deviations for the ablation experiments, or at least assess whether the observed differences are stable across seeds.
minor comments (4)
  1. [Section 5.4.1] The paper itself notes that the threshold delta may be unnecessary and suggests setting it to 0 to consider all intents. This observation undermines the design choice of introducing an extra hyperparameter; please either remove the threshold from the final model or provide a stronger justification for keeping it.
  2. [Section 4.6.1] The construction of the augmented sequence S_u' is unclear: the paper describes it as 'the interaction history of user u' with the target item as v_{N+1}', but does not specify how such a sequence is obtained from the sequence-splitting procedure. Please define the augmentation operation explicitly.
  3. [General] No code or per-seed training logs are provided, which prevents independent verification of the reported results and the claimed standard deviations (if any). Please release code and seeds to facilitate reproducibility.
  4. [Section 3, Figure 2] The text refers to Figure 2-(a), (b), (c) but the figure caption does not label these subpanels; please add subfigure labels for clarity.

Circularity Check

1 steps flagged · score 2.0 of 10

No core circularity: only a self-referential t-SNE validation of the definitional intent-interest split.

  1. self definitional [Section 4.5, Eq. (7); Section 5.5, Intent-Interest Analysis]
    "I^u = H^u - R^u (Eq. 7); Section 5.5: "the intent and interest representations of each user at their most recent time step are projected onto a 2D space. The resulting visualization reveals a strikingly clear separation between intent and interest representations with minimal overlap. This separation demonstrates that intents and interests are inherently different user characteristics.""

    Intent is defined as the residual after subtracting interest (I^u = H^u - R^u). The t-SNE plot in Section 5.5 projects exactly these two internally constructed vectors, so the 'clear separation' between I^u and R^u is a property of the model's own residualization, not independent evidence that intents and interests are inherently different user characteristics. Moreover, since I^u + R^u = H^u, the two are deterministically complementary, so the claimed 'distributional independence' is not demonstrated. This is a self-referential validation of the disentanglement claim; it is not load-bearing for the held-out performance comparison in Table 3.

full rationale

The main result of the paper is an empirical comparison on held-out test data (Table 3), not a derivation from first principles; therefore there is no equation-level circularity in the SOTA claim. The paper contains no load-bearing self-citations: the cited intent models (ICLRec, ICSRec) and the importance-extraction module [33] are external prior work with independent implementations, and the claimed 5.47-14.62% HR and 4.75-12.70% NDCG improvements are computed from the reported test metrics, not from fitted parameters renamed as predictions. The one circular element is the t-SNE 'Intent-Interest Analysis' (Section 5.5), which uses the model's own definitional residual I^u = H^u - R^u as evidence that intents and interests are inherently distinct; this is self-referential but does not affect the empirical benchmark result. The absence of error bars or significance tests is a statistical robustness concern, not circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The central claim is an empirical performance gain, so the main free parameters are the tuned hyperparameters that shape the intent representation. The key axioms are domain assumptions about how intents and interests relate to behavior and items, none of which are externally validated. The invented entities are the two latent factors, which lack independent evidence.

free parameters (5)
  • similarity threshold delta = 0.7 (Sports, Toys), 0.8 (Beauty)
    Controls which past intents are aggregated into the categorical intent; tuned by grid search on validation NDCG@20.
  • intent difference loss weight lambda_d = 0.3 (Sports, Toys), 0.4 (Beauty)
    Weight for L_d in total loss; tuned per dataset.
  • intent-intent contrastive weight lambda_CL1 = 0.5 (Sports, Beauty), 0.3 (Toys)
    Weight for L_CL1; tuned per dataset; higher values generally help.
  • intent-item contrastive weight lambda_CL2 = 0.1 (all datasets)
    Weight for L_CL2; optimal at the smallest tried value and highly sensitive.
  • temperature tau = 1
    Set by hand, not tuned; controls contrastive loss sharpness.
assumptions (6)
  • domain assumption User behaviors decompose into stable interests and dynamic intents
    Core premise of the method, introduced in Section 4.4; not derived from data or theory.
  • domain assumption The residual after subtracting interest representation from behavior representation equals intent
    Equation (7) defines intents as H^u - R^u; assumes all non-interest variation is intent.
  • domain assumption Intents similar to the most recent intent belong to the same intent category that guides the next interaction
    Section 4.5.2 uses this to aggregate past intents and form the categorical intent.
  • domain assumption Users interacting with the same target item share the same intent
    Basis for positive pairs in L_CL1 (Section 4.6.1); the paper itself notes some users may differ, so false negatives are removed only within the batch.
  • domain assumption Item combinations interacted under similar intents form a meaningful positive signal for intent representation
    Basis for L_CL2; item centroid is averaged from item embeddings at positions whose intents are selected.
  • standard math InfoNCE contrastive loss and transformer attention are valid building blocks
    Used as off-the-shelf tools; not proved in this paper.
invented entities (2)
  • Interest representation R
    purpose: Captures stable personal tastes across the interaction sequence.
    Defined only by the model's causal cross-attention; no external data or behavioral measure validates that it corresponds to real interests. The t-SNE plot in Section 5.5 is generated from the model itself.
  • Intent representation I
    purpose: Captures dynamic motivations by subtracting interest from behavior.
    A latent residual with no independent verification; its categorical nature is inferred from the model's own similarity computations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Intent-Interest Disentanglement and Item-Aware Intent Contrastive Learning for Sequential Recommendation." pith.science (2026). https://pith.science/paper/QB4YXMQH

@misc{pith2026250107096,
  author       = {Pith},
  title        = {Pith review of: Intent-Interest Disentanglement and Item-Aware Intent Contrastive Learning for Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QB4YXMQH}},
  note         = {Machine review of arXiv:2501.07096}
}
read the original abstract

Recommender systems aim to provide personalized item recommendations by capturing user behaviors derived from their interaction history. Considering that user interactions naturally occur sequentially based on users' intents in mind, user behaviors can be interpreted as user intents. Therefore, intent-based sequential recommendations are actively studied recently to model user intents from historical interactions for a more precise user understanding beyond traditional studies that often overlook the underlying semantics behind user interactions. However, existing studies face three challenges: 1) the limited understanding of user behaviors by focusing solely on intents, 2) the lack of robustness in categorizing intents due to arbitrary fixed numbers of intent categories, and 3) the neglect of interacted items in modeling of user intents. To address these challenges, we propose Intent-Interest Disentanglement and Item-Aware Intent Contrastive Learning for Sequential Recommendation (IDCLRec). IDCLRec disentangles user behaviors into intents which are dynamic motivations and interests which are stable tastes of users for a comprehensive understanding of user behaviors. A causal cross-attention mechanism is used to identify consistent interests across interactions, while residual behaviors are modeled as intents by modeling their temporal dynamics through a similarity adjustment loss. In addition, without predefining the number of intent categories, an importance-weighted attention mechanism captures user-specific categorical intent considering the importance of intent for each interaction. Furthermore, we introduce item-aware contrastive learning which aligns intents that occurred the same interaction and aligns intent with item combinations occurred by the corresponding intent. Extensive experiments conducted on real-world datasets demonstrate the effectiveness of IDCLRec.

Figures

Figures reproduced from arXiv: 2501.07096 by the authors.

Figure 1
Figure 1. An example of user interactions and corresponding [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Similarity distributions between intent representa [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Average similarity between item combinations in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed IDCLRec. (a) illustrates the architecture of IDCLRec including the modules for each process. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance comparisons of the proposed model [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visualization of intent and interest representations. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 22 canonical work pages

  1. [1]

    Jimmy Lei Ba. 2016. Layer normalization. arXiv preprint arXiv:1607.06450 (2016)

    arXiv 2016

  2. [2]

    Russell W Belk. 1975. Situational variables and consumer behavior. Journal of Consumer research 2, 3 (1975), 157–164

    work page 1975

  3. [3]

    Jesús Bobadilla, Fernando Ortega, Antonio Hernando, and Abraham Gutiérrez

  4. [4]

    Renqin Cai, Jibang Wu, Aidan San, Chong Wang, and Hongning Wang. 2021. Category-aware collaborative sequential recommendation. In Proceedings of the 44th international ACM SIGIR conference on research and development in informa- tion retrieval. 388–397

    work page 2021

  5. [5]

    Jianxin Chang, Chen Gao, Yu Zheng, Yiqun Hui, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. 2021. Sequential recommendation with graph neural networks. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 378–387

    work page 2021

  6. [6]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Interna- tional conference on machine learning . PMLR, 1597–1607. arXiv’25, , Choi et al

    work page 2020

  7. [7]

    Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley, and Caiming Xiong. 2022. Intent contrastive learning for sequential recommendation. In Proceedings of the ACM Web Conference 2022. 2172–2182

    work page 2022

  8. [8]

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

    arXiv 2018

Show all 57 references
  1. [9]

    James F Engel, David T Kollat, and Roger D Blackwell. 1969. Personality measures and market segmentation: Evidence favors interaction view. Business Horizons 12, 3 (1969), 61–70

    1969

  2. [10]

    Ziwei Fan, Zhiwei Liu, Yu Wang, Alice Wang, Zahra Nazari, Lei Zheng, Hao Peng, and Philip S Yu. 2022. Sequential recommendation via stochastic self-attention. In Proceedings of the ACM web conference 2022 . 2036–2047

    2022

  3. [11]

    Hongchao Fang, Sicheng Wang, Meng Zhou, Jiayuan Ding, and Pengtao Xie. 2020. Cert: Contrastive self-supervised learning for language understanding. arXiv preprint arXiv:2005.12766 (2020)

    2020 arXiv

  4. [12]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821 (2021)

    2021 arXiv

  5. [13]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 9729–9738

    2020

  6. [14]

    Ruining He and Julian McAuley. 2016. Fusing similarity models with markov chains for sparse sequential recommendation. In 2016 IEEE 16th international conference on data mining (ICDM) . IEEE, 191–200

    2016

  7. [15]

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

    2016

  8. [16]

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

    2015 arXiv

  9. [17]

    Dietmar Jannach and Markus Zanker. 2024. A Survey on Intent-aware Recom- mender Systems. ACM Transactions on Recommender Systems (2024)

    2024

  10. [18]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems (TOIS) 20, 4 (2002), 422–446

    2002

  11. [19]

    Di Jin, Luzhi Wang, Yizhen Zheng, Guojie Song, Fei Jiang, Xiang Li, Wei Lin, and Shirui Pan. 2023. Dual intent enhanced graph neural network for session- based new item recommendation. In Proceedings of the ACM Web Conference 2023. 684–693

    2023

  12. [20]

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

    2018

  13. [21]

    Yan Kang, Yancong Yuan, Bin Pu, Yun Yang, Lei Zhao, and Jing Guo. 2024. HICL: Hierarchical Intent Contrastive Learning for sequential recommendation. Expert Systems with Applications 251 (2024), 123886

    2024

  14. [22]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  15. [23]

    Walid Krichene and Steffen Rendle. 2020. On sampled metrics for item recom- mendation. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining . 1748–1757

    2020

  16. [24]

    Haoyang Li, Xin Wang, Ziwei Zhang, Jianxin Ma, Peng Cui, and Wenwu Zhu. 2021. Intention-aware sequential recommendation with structured intent transition. IEEE Transactions on Knowledge and Data Engineering 34, 11 (2021), 5403–5414

    2021

  17. [25]

    Jiacheng Li, Tong Zhao, Jin Li, Jim Chan, Christos Faloutsos, George Karypis, Soo-Min Pantel, and Julian McAuley. 2022. Coarse-to-fine sparse sequential recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and development in information ret...

    2022

  18. [26]

    Xuewei Li, Aitong Sun, Mankun Zhao, Jian Yu, Kun Zhu, Di Jin, Mei Yu, and Ruiguo Yu. 2023. Multi-intention oriented contrastive learning for sequential recommendation. In Proceedings of the sixteenth ACM international conference on web search and data mining . 411–419

    2023

  19. [27]

    Yue Liu, Shihao Zhu, Jun Xia, Yingwei Ma, Jian Ma, Wenliang Zhong, Xinwang Liu, Guannan Zhang, and Kejun Zhang. 2024. End-to-end learnable clustering for intent learning in recommendation. arXiv preprint arXiv:2401.05975 (2024)

    2024 arXiv

  20. [28]

    Zhiwei Liu, Yongjun Chen, Jia Li, Philip S Yu, Julian McAuley, and Caiming Xiong. 2021. Contrastive self-supervised sequential recommendation with robust augmentation. arXiv preprint arXiv:2108.06479 (2021)

    2021 arXiv

  21. [29]

    Jie Lu, Dianshuang Wu, Mingsong Mao, Wei Wang, and Guangquan Zhang

  22. [30]

    Jianxin Ma, Chang Zhou, Hongxia Yang, Peng Cui, Xin Wang, and Wenwu Zhu

  23. [31]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel

  24. [32]

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

    2018 arXiv

  25. [33]

    Zhiqiang Pan, Fei Cai, Yanxiang Ling, and Maarten de Rijke. 2020. Rethinking item importance in session-based recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 1837–1840

    2020

  26. [34]

    Xiuyuan Qin, Huanhuan Yuan, Pengpeng Zhao, Junhua Fang, Fuzhen Zhuang, Guanfeng Liu, Yanchi Liu, and Victor Sheng. 2023. Meta-optimized contrastive learning for sequential recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development...

    2023

  27. [35]

    In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52

  28. [36]

    Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. 2022. Contrastive learning for representation degeneration problem in sequential recommendation. In Proceedings of the fifteenth ACM international conference on web search and data mining. 813–823

    2022

  29. [37]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  30. [38]

    Qiaoyu Tan, Jianwei Zhang, Jiangchao Yao, Ninghao Liu, Jingren Zhou, Hongxia Yang, and Xia Hu. 2021. Sparse-interest network for sequential recommendation. In Proceedings of the 14th ACM international conference on web search and data mining. 598–606

    2021

  31. [39]

    Xiuyuan Qin, Huanhuan Yuan, Pengpeng Zhao, Guanfeng Liu, Fuzhen Zhuang, and Victor S Sheng. 2024. Intent Contrastive Learning with Cross Subsequences for Sequential Recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining . 548–556

    2024

  32. [40]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining . 565–573

    2018

  33. [41]

    Md Mehrab Tanjim, Congzhe Su, Ethan Benjamin, Diane Hu, Liangjie Hong, and Julian McAuley. 2020. Attentive sequential models of latent intent for next item recommendation. In Proceedings of The Web Conference 2020 . 2528–2534

    2020

  34. [42]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  35. [43]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  36. [44]

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

    2016

  37. [45]

    Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z Sheng, and Mehmet Orgun. 2019. Sequential recommender systems: challenges, progress and prospects. arXiv preprint arXiv:2001.04830 (2019)

    2019 arXiv

  38. [46]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval . 165–174

    2019

  39. [47]

    Yuling Wang, Xiao Wang, Xiangzhou Huang, Yanhua Yu, Haoyang Li, Mengdi Zhang, Zirui Guo, and Wei Wu. 2024. Intent-aware recommendation via disen- tangled graph contrastive learning. arXiv preprint arXiv:2403.03714 (2024)

    2024 arXiv

  40. [48]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive learning for sequential recommendation. In 2022 IEEE 38th international conference on data engineering (ICDE) . IEEE, 1259– 1273

    2022

  41. [49]

    Lei Wang, Ee-Peng Lim, Zhiwei Liu, and Tianxiang Zhao. 2022. Explanation guided contrastive learning for sequential recommendation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management . 2017–2027

    2022

  42. [50]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-rec: Self-supervised learning for se- quential recommendation with mutual information maximization. In Proceedings of the 29th ACM international conference on in...

    2020

  43. [51]

    Peilin Zhou, Jingqi Gao, Yueqi Xie, Qichen Ye, Yining Hua, Jaeboum Kim, Shoujin Wang, and Sunghun Kim. 2023. Equivariant contrastive learning for sequential recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems. 129–140

    2023

  44. [52]

    Peilin Zhou, You-Liang Huang, Yueqi Xie, Jingqi Gao, Shoujin Wang, Jae Boum Kim, and Sunghun Kim. 2024. Is Contrastive Learning Necessary? A Study of Data Augmentation vs Contrastive Learning in Sequential Recommendation. In Proceedings of the ACM on Web Conference 2024 . 3854–3863

    2024

  45. [54]

    Yipeng Zhang, Xin Wang, Hong Chen, and Wenwu Zhu. 2023. Adaptive disen- tangled transformer for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 3434–3445

    2023

  46. [2013]

    Knowledge-based systems 46 (2013), 109– 132

    Recommender systems survey. Knowledge-based systems 46 (2013), 109– 132

    2013

  47. [2015]

    Decision support systems 74 (2015), 12–32

    Recommender system application developments: a survey. Decision support systems 74 (2015), 12–32

    2015

  48. [2019]

    In Proceedings of the 28th ACM international conference on information and knowledge management

    BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management . 1441–1450

  49. [2020]

    In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

    Disentangled self-supervision in sequential recommenders. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 483–491

Pith tools

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