pith. sign in

arxiv: 2504.19178 · v3 · submitted 2025-04-27 · 💻 cs.IR

Relative Contrastive Learning for Sequential Recommendation with Similarity-based Positive Pair Selection

Zhikai Wang , Yanyan Shen , Zexi Zhang , Li He , Yichun Li , Hao Gu , Yinghua Zhang This is my paper

Pith reviewed 2026-05-22 18:59 UTC · model grok-4.3

classification 💻 cs.IR
keywords sequential recommendationcontrastive learningpositive pair selectionsimilarity-based samplingself-supervised learningrepresentation learninguser intent preservation
0
0 comments X

The pith

Relative Contrastive Learning treats similar sequences with different targets as weak positives to strengthen sequential recommendation training.

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

The paper proposes Relative Contrastive Learning to overcome the scarcity of positive samples in contrastive training for sequential recommendation. Standard augmentation methods can distort user intent, while supervised contrastive approaches lack enough same-target sequences. RCL adds similar sequences as weak positives alongside strong same-target positives and uses a weighted loss to keep representations closer to the strong ones. This supplies richer self-supervision signals during training. The method is shown to improve performance when applied to existing deep sequential models.

Core claim

The authors claim that selecting same-target sequences as strong positives and similar sequences with different target items as weak positives, then applying a weighted relative contrastive loss that enforces greater similarity to strong positives than to weak ones, produces better user representations and yields an average 4.88 percent improvement over state-of-the-art sequential recommendation methods across five public datasets and one private dataset.

What carries the argument

The Relative Contrastive Learning framework, built around a dual-tiered positive sample selection module that identifies strong and weak positives and a relative contrastive learning module that applies a weighted loss to enforce ordering between them.

If this is right

  • Mainstream deep sequential recommendation models receive additional positive samples beyond the limited set of same-target sequences.
  • The weighted relative loss produces representations that respect a clear ordering between strong and weak positives.
  • Performance gains appear consistently across multiple public and private datasets.
  • The approach avoids the risk of altering user intent that data-augmentation strategies can introduce.

Where Pith is reading between the lines

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

  • Sequence similarity could serve as a general signal for related but non-identical preferences in other sequential modeling tasks.
  • Adaptive weighting that scales with measured sequence similarity might further refine the distinction between strong and weak positives.
  • The method may prove especially helpful in cold-start settings where same-target sequences are even rarer.
  • Effective similarity metrics for sequences remain an open design choice that could affect how reliably weak positives contribute.

Load-bearing premise

Similar sequences with different target items supply useful weak positive signals that improve representation learning without adding noise or conflicting information about user preferences.

What would settle it

An experiment in which sequences chosen as weak positives are shown to reflect conflicting user preferences with the anchor sequence and their inclusion measurably lowers recommendation accuracy.

Figures

Figures reproduced from arXiv: 2504.19178 by Hao Gu, Li He, Yanyan Shen, Yichun Li, Yinghua Zhang, Zexi Zhang, Zhikai Wang.

Figure 1
Figure 1. Figure 1: Similarity frequency histograms of sequence pairs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: In the General Contrastive Learning Framework (a), the typical components include a data or model based augmentation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance with different top similar sequence [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study of different similarity metrics and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE results of two sequences from Beauty and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The average dot products of the center sequence. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Contrastive Learning (CL) enhances the training of sequential recommendation (SR) models through informative self-supervision signals. Existing methods often rely on data augmentation strategies to create positive samples and promote representation invariance. Some strategies such as item reordering and item substitution may inadvertently alter user intent. Supervised Contrastive Learning (SCL) based methods find an alternative to augmentation-based CL methods by selecting same-target sequences (interaction sequences with the same target item) to form positive samples. However, SCL-based methods suffer from the scarcity of same-target sequences and consequently lack enough signals for contrastive learning. In this work, we propose to use similar sequences (with different target items) as additional positive samples and introduce a Relative Contrastive Learning (RCL) framework for sequential recommendation. RCL comprises a dual-tiered positive sample selection module and a relative contrastive learning module. The former module selects same-target sequences as strong positive samples and selects similar sequences as weak positive samples. The latter module employs a weighted relative contrastive loss, ensuring that each sequence is represented closer to its strong positive samples than its weak positive samples. We apply RCL on two mainstream deep learning-based SR models, and our empirical results reveal that RCL can achieve 4.88% improvement averagely than the state-of-the-art SR methods on five public datasets and one private dataset.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Relative Contrastive Learning (RCL) for sequential recommendation. It augments supervised contrastive learning via a dual-tiered positive sample selection module that designates same-target sequences as strong positives and similarity-selected sequences (different targets) as weak positives, followed by a weighted relative contrastive loss that enforces closer embeddings to strong than to weak positives. The framework is applied to two mainstream deep SR models and evaluated on five public datasets plus one private dataset, claiming an average 4.88% improvement over state-of-the-art SR methods.

Significance. If the central empirical claim holds after addressing validation gaps, the work offers a practical way to increase positive sample volume in contrastive SR without intent-altering augmentations such as reordering or substitution. The relative loss hierarchy and multi-dataset evaluation (including a private dataset) are strengths that could influence follow-up work on weak supervision in recommendation. The approach builds directly on existing SCL ideas while adding a controllable strength ordering.

major comments (2)
  1. [§3.2] §3.2 (Dual-tiered positive sample selection): The central claim that similarity-based sequences with mismatched targets serve as reliable weak positives rests on the untested assumption that sequence similarity proxies shared user intent. No target-item overlap statistics, preference-alignment scores, or conflict analysis between anchors and weak positives are reported; without such evidence the relative loss may optimize conflicting gradients rather than useful regularization.
  2. [§5] §5 (Experiments): The reported 4.88% average improvement is load-bearing for the contribution, yet the section provides insufficient detail on baseline re-implementations, hyperparameter search protocols, statistical significance testing (e.g., paired t-tests or Wilcoxon), and exact dataset split ratios. These omissions prevent independent verification that the gains are attributable to the RCL components rather than tuning differences.
minor comments (2)
  1. [Abstract] Abstract: The phrase '4.88% improvement averagely' should specify the primary metric (HR@K or NDCG@K) and the exact set of compared methods for immediate clarity.
  2. [§3.3] Notation: The weighting parameters in the relative contrastive loss (Eq. likely in §3.3) are introduced without an explicit statement of whether they are fixed, learned, or tuned per dataset; a short clarification would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will incorporate revisions to improve clarity and reproducibility.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Dual-tiered positive sample selection): The central claim that similarity-based sequences with mismatched targets serve as reliable weak positives rests on the untested assumption that sequence similarity proxies shared user intent. No target-item overlap statistics, preference-alignment scores, or conflict analysis between anchors and weak positives are reported; without such evidence the relative loss may optimize conflicting gradients rather than useful regularization.

    Authors: We acknowledge that the original manuscript does not report quantitative analyses such as target-item overlap rates or preference-alignment metrics between anchors and weak positives. The design rationale is that sequence similarity (computed via embedding or interaction overlap) serves as a proxy for latent user intent, and the weighted relative contrastive loss explicitly enforces a strict hierarchy (strong positives closer than weak positives) to limit gradient conflicts. Nevertheless, to address the concern directly, we will add an analysis subsection with overlap statistics, conflict examples, and ablation on weak-positive quality in the revised version. revision: yes

  2. Referee: [§5] §5 (Experiments): The reported 4.88% average improvement is load-bearing for the contribution, yet the section provides insufficient detail on baseline re-implementations, hyperparameter search protocols, statistical significance testing (e.g., paired t-tests or Wilcoxon), and exact dataset split ratios. These omissions prevent independent verification that the gains are attributable to the RCL components rather than tuning differences.

    Authors: We agree that additional experimental details are required for independent verification. The revised manuscript will expand Section 5 with: explicit descriptions of baseline re-implementations (including code-level adaptations), the hyperparameter search protocol and ranges, results of statistical significance tests (paired t-tests and Wilcoxon signed-rank across 5 runs), and the precise train/validation/test split ratios used for each of the six datasets. These additions will confirm that reported gains are attributable to the RCL components. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes an explicit new framework (RCL) consisting of a dual-tiered positive sample selection module (same-target sequences as strong positives, similarity-selected sequences as weak positives) and a weighted relative contrastive loss that enforces embedding proximity hierarchy. These components are defined directly as modeling choices extending standard contrastive and supervised contrastive learning, without any equations or steps that reduce predictions or results to fitted inputs by construction. No load-bearing self-citations, uniqueness theorems imported from prior author work, or ansatzes smuggled via citation are present in the provided text. Empirical improvements are reported on external datasets and models rather than derived tautologically. The derivation remains self-contained with independent content.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review is limited to the abstract, so the ledger captures only the high-level assumptions visible there; no free parameters or invented entities are explicitly described.

axioms (1)
  • domain assumption Similar sequences (different target items) can serve as valid weak positive samples for contrastive learning without distorting user intent.
    This premise underpins the dual-tiered positive sample selection module described in the abstract.

pith-pipeline@v0.9.0 · 5790 in / 1386 out tokens · 53750 ms · 2026-05-22T18:59:00.686464+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

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

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

40 extracted references · 40 canonical work pages · 3 internal anchors

  1. [1]

    Abdulaziz AlQatan, Leif Azzopardi, and Yashar Moshfeghi. 2020. Analyzing the Influence of Bigrams on Retrieval Bias and Effectiveness. InProceedings of the 2020 ACM SIGIR on International Conference on Theory of Information Retrieval (Virtual Event, Norway). Association for Computing Machinery, New York, NY, USA, 157–160. doi:10.1145/3409256.3409831

    work page doi:10.1145/3409256.3409831 2020

  2. [2]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PMLR, 1597–1607

    work page 2020

  3. [3]

    Xinlei Chen and Kaiming He. 2021. Exploring simple siamese representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 15750–15758

    work page 2021

  4. [4]

    Liu Chong, Xiaoyang Liu, Rongqin Zheng, Lixin Zhang, Xiaobo Liang, Juntao Li, Lijun Wu, Min Zhang, and Leyu Lin. 2023. CT4Rec: Simple yet Effective Consistency Training for Sequential Recommendation. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3901–3913

    work page 2023

  5. [5]

    Yizhou Dang, Enneng Yang, Guibing Guo, Linying Jiang, Xingwei Wang, Xiaoxiao Xu, Qinghui Sun, and Hong Liu. 2022. Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation.arXiv preprint arXiv:2212.08262(2022)

    work page arXiv 2022

  6. [6]

    Mathieu d’Aquin, Stefan Dietze, Claudia Hauff, Edward Curry, Philippe Cu- dre Mauroux, 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 Sequential Recommendation with Mutual Information Maximization.Proceedings of the 29th ACM International Conference on ...

    work page doi:10.1145/3340531.3411954 2020

  7. [7]

    Chengxin Ding, Jianhui Li, Tianhang Liu, and Zhongying Zhao. 2022. Graph- Augmented Multi-Level Representation Learning for Session-based Recommen- dation.2022 IEEE 8th International Conference on Cloud Computing and Intelligent Systems (CCIS)00 (2022), 576–580. doi:10.1109/ccis57298.2022.10016436

    work page doi:10.1109/ccis57298.2022.10016436 2022

  8. [8]

    Yike Guo, Faisal Farooq, Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction.AAAI(2018), 1059–1068

    work page 2018

  9. [9]

    Mohammad Al Hasan, Li Xiong, Shuqing Bian, Wayne Xin Zhao, Jinpeng Wang, and Ji-Rong Wen. 2022. A Relevant and Diverse Retrieval-enhanced Data Aug- mentation Framework for Sequential Recommendation.Proceedings of the 31st ACM International Conference on Information & Knowledge Management(2022), 2923–2932. doi:10.1145/3511808.3557071

    work page doi:10.1145/3511808.3557071 2022

  10. [10]

    Mohammad Al Hasan, Li Xiong, Jiangxia Cao, Xin Cong, Jiawei Sheng, Tingwen Liu, and Bin Wang. 2022. Contrastive Cross-Domain Sequential Recommendation. Proceedings of the 31st ACM International Conference on Information & Knowledge Management(2022), 138–147. doi:10.1145/3511808.3557262

    work page doi:10.1145/3511808.3557262 2022

  11. [11]

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

    work page 2020

  12. [12]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

    work page

  13. [13]

    Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939(2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  14. [14]

    R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. 2018. Learning deep represen- tations by mutual information estimation and maximization.arXiv preprint arXiv:1808.06670(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    Chengkai Huang, Shoujin Wang, Xianzhi Wang, and Lina Yao. 2023. Dual Con- trastive Transformer for Hierarchical Preference Modeling in Sequential Rec- ommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 99–109

    work page 2023

  16. [16]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

    work page 2018

  17. [17]

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-Interest Network with Dynamic Routing for Recommendation at Tmall. InProceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China)(CIKM ’19). Association for Computing Machiner...

    work page doi:10.1145/3357384.3357814 2019

  18. [18]

    Chong Liu, Xiaoyang Liu, Rongqin Zheng, Lixin Zhang, Xiaobo Liang, Juntao Li, Lijun Wu, Min Zhang, and Leyu Lin. 2021. C$^2$-Rec: An Effective Consistency Constraint for Sequential Recommendation.arXiv(2021). arXiv:2112.06668 doi:10.48550/arxiv.2112.06668 C2 Rec

    work page doi:10.48550/arxiv.2112.06668 2021

  19. [19]

    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)

    work page arXiv 2021

  20. [20]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  21. [21]

    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. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’23). Association for Computing Machinery, New Y...

    work page doi:10.1145/3539618.3591727 2023

  22. [22]

    Ruihong Qiu, Zi Huang, Tong Chen, and Hongzhi Yin. 2021. Exploiting positional information for session-based recommendation.ACM Transactions on Information Systems (TOIS)40, 2 (2021), 1–24

    work page 2021

  23. [23]

    Ruihong Qiu, Zi Huang, and Hongzhi Yin. 2021. Memory Augmented Multi- Instance Contrastive Predictive Coding for Sequential Recommendation.2021 IEEE International Conference on Data Mining (ICDM)00 (2021), 519–528. doi:10. 1109/icdm51629.2021.00063

    work page arXiv 2021

  24. [24]

    Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. 2021. Contrastive Learn- ing for Representation Degeneration Problem in Sequential Recommendation. arXiv(2021). arXiv:2110.05730 doi:10.48550/arxiv.2110.05730

    work page doi:10.48550/arxiv.2110.05730 2021

  25. [25]

    Ruihong Qiu, Jingjing Li, Zi Huang, and Hongzhi Yin. 2019. Rethinking the item order in session-based recommendation with graph neural networks. In Proceedings of the 28th ACM international conference on information and knowledge management. 579–588

    work page 2019

  26. [26]

    Ruihong Qiu, Hongzhi Yin, Zi Huang, and Tong Chen. 2020. Gag: Global at- tributed graph neural network for streaming session-based recommendation. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 669–678

    work page 2020

  27. [27]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Fac- torizing personalized Markov chains for next-basket recommendation.WWW (2010), 811–820

    work page 2010

  28. [28]

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

    work page

  29. [29]

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

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

    work page

  30. [30]

    Chenyang Wang, Weizhi Ma, Chong Chen, Min Zhang, Yiqun Liu, and Shaoping Ma. 2023. Sequential Recommendation with Multiple Contrast Signals.ACM Transactions on Information Systems41, 1 (2023), 1–27. doi:10.1145/3522673

    work page doi:10.1145/3522673 2023

  31. [31]

    Pengfei Wang, Jiafeng Guo, Yanyan Lan, Jun Xu, Shengxian Wan, and Xueqi Cheng. 2015. Learning Hierarchical Representation Model for NextBasket Rec- ommendation.SIGIR(2015), 403–412

    work page 2015

  32. [32]

    Zhikai Wang and Yanyan Shen. 2022. Time-aware Multi-interest Capsule Network for Sequential Recommendation. InProceedings of the 2022 SIAM International Conference on Data Mining (SDM). SIAM, 558–566

    work page 2022

  33. [33]

    Zhikai Wang and Yanyan Shen. 2023. Incremental Learning for Multi-Interest Sequential Recommendation. InICDE. IEEE, 1071–1083

    work page 2023

  34. [34]

    Zhikai Wang and Yanyan Shen. 2024. A Framework for Elastic Adaptation of User Multiple Intents in Sequential Recommendation.IEEE Transactions on Knowledge and Data Engineering(2024), 1–13. doi:10.1109/TKDE.2024.3354796

    work page doi:10.1109/tkde.2024.3354796 2024

  35. [35]

    Zhikai Wang, Yanyan Shen, Zibin Zhang, and Kangyi Lin. 2023. Feature Staleness Aware Incremental Learning for CTR Prediction. InIJCAI

    work page 2023

  36. [36]

    Xin Xia, Hongzhi Yin, Junliang Yu, Qinyong Wang, Lizhen Cui, and Xiangliang Zhang. 2021. Self-Supervised Hypergraph Convolutional Networks for Session- based Recommendation(Proceedings of the AAAI Conference on Artificial Intelli- gence, Vol. 35). 4503–4511. doi:10.1609/aaai.v35i5.16578

    work page doi:10.1609/aaai.v35i5.16578 2021

  37. [37]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive Learning for Sequential Recommendation. 2022 IEEE 38th International Conference on Data Engineering (ICDE)00 (2022), 1259–1273. doi:10.1109/icde53745.2022.00099

    work page doi:10.1109/icde53745.2022.00099 2022

  38. [38]

    Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. 2022. Use all the labels: A hierarchical multi-label contrastive learning framework. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16660–16669

    work page 2022

  39. [39]

    Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. 2022. Filter-enhanced MLP is All You Need for Sequential Recommendation. InProceedings of the ACM Web Conference 2022(Virtual Event, Lyon, France)(WWW ’22). Association for Computing Machinery, New York, NY, USA, 2388–2399. doi:10.1145/3485447. 3512111

    work page doi:10.1145/3485447 2022

  40. [40]

    Yu Zhu, Hao Li, Yikang Liao, Beidou Wang, Ziyu Guan, Haifeng Liu, and Deng Cai. 2017. What to Do Next: Modeling User Behaviors by Time-LSTM.IJCAI (2017), 3602–3608

    work page 2017