SimDiffRec: Semantic Similarity-Guided Diffusion for Contrastive Sequential Recommendation

Donghyeon Park; Jinkyeong Choi; Yejin Noh

REVIEW 5 major objections 5 minor 51 references

SimDiffRec: Semantic Similarity-Guided Diffusion for Contrastive Sequential Recommendation

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

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

Pith's one-line read SimDiffRec claims that replacing random diffusion noise with semantically similar item-embedding averages, plus confidence-based position selection and hard negative sampling, improves contrastive sequential recommendation accuracy on…

desk verdict A plausible augmentation plugin with a potentially fatal no-op positive-sample problem; the empirical story is currently unverified. read the letter →

arxiv 2507.11866 v2 pith:LDU26XTZ submitted 2025-07-16 cs.IR

Jinkyeong Choi , Yejin Noh , Donghyeon Park This is my paper · ORCID

classification cs.IR

keywords sequentialrecommendationdiffusionmodeldataaugmentationcontrastivelearningsemanticsimilaritynoisehardnegativesamplingconfidence-basednext-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 data augmentation for contrastive sequential recommendation should be guided by semantic similarity rather than randomness. It proposes a diffusion-based augmentation plug-in that corrupts a user history with the averaged embeddings of the most similar items, augments only positions the diffusion model reconstructs with high confidence, and uses rank-selected hard negatives to sharpen contrastive learning. If the claim is right, a standard Transformer-based sequential recommender can be upgraded by swapping in this augmentation, and the accuracy gains reported on five benchmarks would follow without extra inference cost.

What carries the argument

The central object is a deterministic, semantically guided noise vector: for each item in a user sequence, the model takes the dot product of the item embedding with the full item embedding matrix, masks self-similarity, selects the top $k$ most similar embedding vectors, and uses their average as the noise injected during the diffusion forward process. Augmentation positions are selected by the diffusion model's reconstruction confidence, defined as the maximum softmax probability over item logits at each position after denoising. At a selected position, the highest-probability reconstructed item becomes the positive sample and the $k_{\text{sample}}$-th ranked item becomes the hard negative; the two views feed an InfoNCE-style contrastive loss that is added to the recommendation loss and the diffusion loss.

What would settle it

Run the same framework with the semantic noise replaced by same-norm Gaussian noise or by the average of k randomly chosen item embeddings, keeping all other components fixed; if HR@10 on Beauty stays near 0.0932 or the gap over the strongest baseline persists, the semantic-similarity claim is not the driver. Separately, inspect the augmented pairs: if a large share of top-1 positive replacements equal the original item, the positive views are degenerate and the contrastive gain must come from elsewhere.

Watch

Extended reading notes

Core claim

The paper reports that random noise and random augmentation positions are the reason diffusion-based augmentation damages sequential context, and that replacing both with similarity-guided choices improves contrastive sequential recommendation. Concretely, the framework paired with a Transformer-based recommendation backbone beats nine baselines on five benchmarks, with Beauty HR@10 reaching 0.0932 versus 0.0867 for the strongest baseline, and HR@5 and NDCG@10 improving on every dataset. Each removed component, similarity noise, confidence-based position selection, or hard negatives, degrades performance in the ablations, which the paper reads as evidence that all three modules contribute. The framework is not a standalone recommender but a training-time augmentation that can be attached to existing sequential recommendation models.

Load-bearing premise

The claim stands on the assumption that averaging the top-k most similar item embeddings yields noise that is semantically consistent with the original sequence and useful as a diffusion corruption signal at every stage of training; early in training these embeddings are uninformative, and later they may simply reinforce the model's own biases.

Editorial extensions

If this is right

Applied to a standard Transformer-based sequential recommender, the framework reports higher HR@5, HR@10, NDCG@5, and NDCG@10 than all nine baselines on Beauty, Toys, Sports, Yelp, and ML-1m.
Removing any one of the three components, similarity-based noise, confidence-based position selection, or hard negative sampling, lowers performance in the ablation study, so each component is load-bearing in the reported gain.
The diffusion module is not used at inference time, so the augmented model runs at roughly the same speed as the backbone recommender, making the approach a drop-in training-time upgrade.
The hard-negative rank study shows that smaller $k_{\text{sample}}$ values, which produce negatives most similar to the positive sample, give the strongest performance, indicating that negative difficulty drives part of the benefit.

Reading between the lines

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

A testable extension the paper does not run is to replace the semantic similarity noise with random Gaussian noise of matched magnitude or with the average of randomly chosen item embeddings; if the Beauty HR@10 gap narrows or disappears, the semantic mechanism, not the added training signal, is the active ingredient.
Because the noise is built from the co-trained item embedding matrix, the method implicitly assumes the embeddings are semantically meaningful throughout training; a warmup schedule or fixed pretrained embeddings could change whether early-training noise quality matters.
Confidence-based position selection relies on the softmax max-probability as a trust score; a calibrated uncertainty estimate or an ensemble variance could be tested as a drop-in alternative to see whether calibration, not peakiness, is what makes positions useful for augmentation.
The paper's framework suggests that other diffusion-based augmentation pipelines in sequence modeling could replace random noise with neighborhood averages in the same embedding space, but that extension would need validation outside the recommender setting.

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.

Desk Editor's Note

read the letter

Two things to know about this paper. First, the two new pieces are real: averaging the top-k similar item embeddings as diffusion noise, and selecting augmentation positions by reconstruction confidence. They are modest but legitimate extensions of CaDiRec and SeeDRec, and the paper is honest that the method is an augmentation plugin, not a new recommender. Second, though, the core empirical claim is not actually verified. The positive sample used in the contrastive loss is the top-1 reconstruction at the high-confidence positions. Because the diffusion model is trained to reconstruct the original sequence, high-confidence positions are exactly where that top-1 is likely the original item. The paper never reports the frequency with which the positive view is identical to the original sequence. If that happens often, the contrastive positive is a no-op and the loss only receives gradient from the hard negative and in-batch negatives. The cosine similarity term in Eq. 17 then does no work. That would mean the reported gains come from the auxiliary diffusion loss or the negative sampling, not from the semantic augmentation the paper claims. This is a load-bearing issue, not a side detail.

The experimental reporting has other holes: no error bars or significance tests anywhere, a variance-based stopping rule that is not a statistical protocol, and missing values for the number of augmented positions k and the temperature tau. The code link has no commit hash. Each of these is fixable, but together they make the 1-13% relative gains over CaDiRec hard to interpret.

What the paper does well: the ablation study is complete enough to show each component helps, the hyperparameter analysis covers alpha and beta, and the complexity table puts the method's cost in line with comparable augmentation approaches. The writing is clean and the positioning against CaDiRec and SeeDRec is accurate.

The circularity concern about the semantic noise being computed from the same embeddings that are being trained is real but minor; plenty of methods do this and it often works, but it deserves a comment on why the noise direction remains meaningful during training.

Who is this for? Researchers building diffusion-based augmentation for sequential recommendation. It's a useful data point, and the no-op issue is a genuinely instructive design trap.

Verdict: a serious referee should see this, but the paper needs a substantial revision before acceptance. The authors should report the no-op rate, add error bars, and run proper significance tests.

Referee Report

5 major / 5 minor

Summary. The paper proposes SimDiffRec, a data-augmentation framework for sequential recommendation. It replaces random Gaussian noise in a diffusion module with a deterministic "semantic similarity" noise defined as the average of the top-k most similar item embeddings, selects augmentation positions by the diffusion model's softmax confidence, and defines positive and hard-negative views from the top-1 and k_sample-th ranked reconstructed items. The total objective (Eq. 18) combines the sequential recommendation loss, an InfoNCE contrastive loss, and the diffusion loss. The paper reports experiments on Beauty, Toys, Sports, Yelp, and ML-1m, comparing against nine baselines and reporting consistent HR/NDCG improvements, together with ablations and hyperparameter studies.

Significance. If validated, SimDiffRec is an attractive contribution: it is a plug-in augmentation module for Transformer-based sequential recommendation backbones, it incurs almost no inference-time overhead, and the authors provide a public code link. The reported gains are consistent across five datasets, and the ablations in Table 3 give some support to each proposed component. However, the current evidence does not rule out the possibility that the gains come from trivial positives or from the auxiliary diffusion/contrastive losses rather than from the proposed semantic augmentation, and the lack of error bars, significance tests, and several key hyperparameter values prevents a reliable assessment of the central empirical claim.

major comments (5)

[§4.3–4.4, Eq. (17)] The positive sample is defined as the top-1 denoised item at a position selected for high confidence, and no check is made against the original item. When the top-1 reconstruction equals the original item, the "Augmentation View 1" is identical to the original sequence, so e_v+ = e_u and the InfoNCE numerator is a trivial self-similarity term. The reported gains then cannot be attributed to semantically consistent augmentation. Please report the frequency of such no-op replacements on each dataset, and rerun the key comparison and ablation with the original item excluded from the positive candidate set.
[Table 2 and §5.1.4] The caption of Table 2 states that bold entries indicate a "statistically significant improvement," but the paper reports no standard deviations, confidence intervals, or significance tests. The stated protocol of running each experiment until the variance drops below 0.01 is not a valid statistical basis for that claim, since early stopping on variance can bias results. Please report per-run standard deviations or variance for the ten runs and provide paired significance tests for the headline improvement over the best baseline.
[§5.1.4, §5.5, Eq. (17)] Several hyperparameters that directly control the method are not specified in the implementation details: the number of augmented positions k, the default value of k_sample used in the main results, the temperature τ in Eq. (17), and the noise schedule α_t, β_t used in Eqs. (8)–(9). Without these values the experiments cannot be reproduced, and the sensitivity of the main result to k and τ cannot be assessed. Please give the values used for Table 2 and Table 3, and include a sensitivity analysis for k and τ.
[§4.2.1 and Eq. (13)] The forward process in Eq. (8) injects the same deterministic semantic noise at every diffusion step, which is not the Gaussian corruption assumed by the standard DDPM objective, yet Eq. (13) uses the standard reconstruction loss without derivation or verification that the learned reverse process is well calibrated under this non-Gaussian corruption. Please justify this design choice theoretically, or provide empirical evidence of reconstruction quality (e.g., reconstruction accuracy broken down by training epoch) showing that the diffusion model actually recovers the original sequence before its outputs are used to select positive samples.
[§4.1 and Eq. (6)] The semantic similarity noise is computed from item embeddings that are the same embeddings co-trained by L_sr and L_cl, so the "semantic" neighborhoods are not stable and early in training are likely uninformative. The paper does not compare the proposed semantic noise against random noise at the same injected noise strength, nor against similarity noise computed from fixed/pretrained embeddings. Please add such a control to show that the semantic composition of the noise, rather than the extra regularization and auxiliary losses, is responsible for the observed gains.

minor comments (5)

[Eq. (3)] In Eq. (3), the summation index n is not defined in Section 3.2; please clarify that it is the length of the user sequence (or the padding length used in training).
[Table 2] The column labeled "improv." is not defined in the table or caption; please state explicitly that it is the relative improvement over the underlined second-best baseline.
[Figure 4] The axis labels in Figure 4 are crowded and run together (e.g., "HR@10 (Beauty & Toys)HR@10 (Sports)"); please reformat the panels with separate, readable labels and consistent subfigure captions.
[§5.7] The statement that "simple augmentation that only increases the amount of data" was excluded from the timing comparison is vague; please specify exactly which techniques were included in each model's measured training time.
[Eq. (13)] The final term -log p_θ(s | z_0) in Eq. (13) is a cross-entropy term with a different scale than the preceding L2 terms; please indicate its coefficient and whether it is averaged over sequence positions.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'positive' augmentation view is the diffusion model's own reconstruction of the original item, so at the selected high-confidence positions it can be identical to the input, making the claimed semantic contrastive augmentation a no-op.

fitted input called prediction [Section 4.3-4.4, Eqs. (13), (16), (17)]
"First, for a given augmentation position, we select the item reconstructed with the highest probability as the positive sample, as detailed in Section 4.3. ... A high confidence score c_i indicates that the context at that position is well captured and the model can reconstruct the correct item with high probability. ... Eq. (13): Ld = Σ ∥z0−fθ(zt,t)∥² + ∥e−fθ(z1,1)∥² − log pθ(s|z0)."

The diffusion loss (Eq. 13) trains fθ to map noised states back to the original embeddings z0 and to maximize log pθ(s|z0) for the original discrete sequence. The positive sample is then defined (Sec. 4.4) as the item with highest reconstructed probability, and the positions are chosen (Sec. 4.3) where this probability is highest, i.e., where the model 'can reconstruct the correct item with high probability.' Thus the 'augmented' item is the model's best reconstruction of the original item, not an independent semantic view. At those high-confidence positions the argmax is trained to be the original item, so replacing the original item with the positive item can leave the sequence unchanged and e_v+ = e_u in Eq. 17.

full rationale

Score 6: The framework has independent ingredients—hard-negative sampling, an auxiliary diffusion reconstruction loss, and a position-selection heuristic—and there are no load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling citations. The noise-generation step uses the co-trained item embedding space as a similarity oracle, which is a heuristic rather than a formal circular reduction. The circularity is concentrated in the construction of the positive view: the diffusion model is trained (Eq. 13) to reconstruct the original sequence, the positive item is the argmax of that reconstruction (Sec. 4.4), and the augmented positions are selected exactly where the model is most confident about that reconstruction (Sec. 4.3). Consequently the 'semantic augmentation' can reduce to the identity view e_v+ = e_u in Eq. 17, in which case the contrastive loss contains no non-trivial positive-pair information. The paper does not report the frequency of such degenerate replacements, and the evaluation is presented without paired significance tests; those are correctness concerns, but the identity-positive construction is the specific reduction that makes the central claim partially circular.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central claim depends on several custom, hand-tuned choices (k_noise, k_sample, alpha, beta) and on the untested assumption that deterministic similarity-based corruption can be reversed with a standard diffusion objective. The shared embedding matrix creates a feedback loop between the noise, the confidence, and the loss, but the held-out test evaluation provides independent grounding for the performance claim.

free parameters (7)

k_noise = tuned per dataset, best in 1 to 50
Number of top-similar item embeddings averaged to form the semantic noise (Eq. 7). Reported in Section 5.5, but exact per-dataset values are not given.
k_sample = not reported
Rank index used to pick the hard negative from the softmax distribution (Section 4.4). The paper says small values work best but does not give the values used in the main experiments.
alpha = 0.1 to 0.5
Weight of the contrastive loss in Eq. 18, tuned on the validation set over [0.1, 0.5].
beta = 0.1 to 0.5
Weight of the diffusion loss in Eq. 18, tuned on the validation set over [0.1, 0.5].
k (number of augmented positions) = not specified
The top k positions selected for augmentation in Section 4.3; no value is given in the implementation details.
temperature tau = not specified
Temperature in the InfoNCE contrastive loss (Eq. 17); not reported.
noise schedule beta_t = not specified
The variance schedule in Eq. 8 is not defined (e.g., linear or cosine), leaving the forward process underspecified.

assumptions (4)

ad hoc to paper Reverse diffusion with deterministic non-Gaussian noise can be learned with the standard DDPM-style loss.
Section 4.2 replaces Gaussian noise with an average of similar item embeddings but keeps the DDPM reverse equation (Eq. 10) and the DiffuSeq simplification (Eq. 13) without proving the posterior is tractable or that the deterministic corruption is reversible.
domain assumption Dot product between item embeddings reflects semantic similarity between items.
Section 4.1 uses the dot product e_u * W^T to rank similar items; this is assumed without evaluating the quality of the embedding space.
ad hoc to paper The softmax confidence of the diffusion reconstruction identifies positions safe for augmentation.
Section 4.3 selects positions with the highest max-softmax probability, assuming that high confidence implies the context is preserved and that replacing the item there yields a valid positive sample.
domain assumption The item embedding matrix W is shared and co-trained with the SR loss and diffusion loss.
Figure 2 and Section 5.1.4 show a shared embedding layer. The noise and confidence are therefore computed from the same representations being optimized, creating a feedback loop.

invented entities (1)

Semantic similarity-based noise
purpose: Deterministic corruption signal in the diffusion forward process, defined as the average of the top-k similar item embeddings (Eq. 7).
The construct is internal to the paper; its semantic consistency is only assessed through the final recommendation metrics, with no external falsifiable handle.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SimDiffRec: Semantic Similarity-Guided Diffusion for Contrastive Sequential Recommendation." pith.science (2026). https://pith.science/paper/LDU26XTZ

@misc{pith2026250711866,
  author       = {Pith},
  title        = {Pith review of: SimDiffRec: Semantic Similarity-Guided Diffusion for Contrastive Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LDU26XTZ}},
  note         = {Machine review of arXiv:2507.11866}
}

read the original abstract

In sequential recommendation systems, data augmentation and contrastive learning techniques have recently been introduced using diffusion models to achieve robust representation learning. However, most of the existing approaches use random augmentation, which risks damaging the contextual information of the original sequence. Accordingly, we propose SimDiffRec: a Semantic Similarity-Guided Diffusion for Contrastive Sequential Recommendation. Our framework leverages the similarity between item embedding vectors to generate semantically consistent noise. Moreover, we utilize high confidence scores in the denoising process to select our augmentation positions. This approach more effectively reflects contextual and structural information compared to augmentation at random positions. From a contrastive learning perspective, the proposed augmentation technique, combined with hard negative sampling, provides more discriminative positive and negative samples, simultaneously improving training efficiency and recommendation performance. Experimental results on five benchmark datasets show that SimDiffRec outperforms the existing baseline models. The code of our framework is available at https://github.com/zingyon/SimDiffRec.

Figures

Figures reproduced from arXiv: 2507.11866 by the authors.

**Figure 2.** Overview of our proposed SimDiffRec: (1) Embed user–item sequences with a Transformer-based SR model’s embedding layer and insert semantic similarity–based noise. (2) Preserve contextual information during the diffusion denoising process and select augmentation positions with high confidence. (3) Apply hard negative sampling for contrastive learning. included in noise generation, we apply self-masking to exclude an … view at source ↗

**Figure 3.** Hyperparameter study of 𝛼, 𝛽 on three datasets. across all datasets, indicating that the augmentation position selection itself is a critical factor governing performance. This module contributes to learning stability by guiding augmentation toward contexts the model already confidently understands, thereby inducing the generation of reliable positive samples. Finally, "w/o 𝑘𝑠𝑎𝑚𝑝𝑙𝑒 " is the setting where hard nega… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Performance changes based on hard negative sam [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: T-SNE visualization of sequence embeddings trained [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

[1]

Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley, and Caiming Xiong. 2022. Intent Contrastive Learning for Sequential Recommendation. InProceedings of WSDM ’26, February 22–26, 2026, Boise, ID, USA Jinkyeong Choi, Yejin Noh, and Donghyeon Park the ACM Web Conference 2022. 2172–2182

work page 2022
[2]

Zhuowei Chen, Lianxi Wang, Yuben Wu, Xinfeng Liao, Yujia Tian, and Junyang Zhong. 2024. An Effective Deployment of Diffusion LM for Data Augmentation in Low-Resource Sentiment Classification. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 1844–1856

work page 2024
[3]

Ziqiang Cui, Haolun Wu, Bowei He, Ji Cheng, and Chen Ma. 2024. Context Matters: Enhancing Sequential Recommendation with Context-aware Diffusion- based Contrastive Learning. InProceedings of the 33rd ACM International Confer- ence on Information and Knowledge Management. 404–414

2024
[4]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion Models Beat GANs on Image Synthesis. InAdvances in Neural Information Processing Systems, Vol. 34. 8780–8794

work page 2021
[5]

Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. InProceedings of the 2021 Conference on Em- pirical Methods in Natural Language Processing. 6894–6910

2021
[6]

Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. 2023. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. InThe Eleventh International Conference on Learning Representations

work page 2023
[7]

Zhengfu He, Tianxiang Sun, Qiong Tang, Kuanning Wang, Xuanjing Huang, and Xipeng Qiu. 2023. DiffusionBERT: Improving Generative Masked Language Models with Diffusion Models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 4521–4534

work page 2023
[8]

Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent Neural Networks with Top-k Gains for Session-based Recommendations. InProceedings of the 27th ACM international conference on information and knowledge management. 843–852

work page 2018

Show all 51 references

[9]

Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk
[10]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems, Vol. 33. 6840–6851

2020
[11]

Khawar Islam and Naveed Akhtar. 2025. Context-Guided Responsible Data Augmentation with Diffusion Models. arXiv:2503.10687

2025 arXiv
[12]

Yuxin Jiang, Linhan Zhang, and Wei Wang. 2022. Improved Universal Sentence Embeddings with Prompt-based Contrastive Learning and Energy-based Learn- ing. InFindings of the Association for Computational Linguistics: EMNLP 2022. 3021–3035

2022
[13]

Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Rec- ommendation. In2018 IEEE International Conference on Data Mining (ICDM). 197–206

2018
[14]

Wuchao Li, Rui Huang, Haijun Zhao, Chi Liu, Kai Zheng, Qi Liu, Na Mou, Guorui Zhou, Defu Lian, Yang Song, Wentian Bao, Enyun Yu, and Wenwu Ou. 2025. DimeRec: A Unified Framework for Enhanced Sequential Recommendation via Generative Diffusion Models. InProceedings of the Eighte...

2025
[15]

Hashimoto

Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B. Hashimoto. 2022. Diffusion-LM Improves Controllable Text Generation. InPro- ceedings of the 36th International Conference on Neural Information Processing Systems (NIPS ’22). 16 pages

2022
[16]

Zihao Li, Aixin Sun, and Chenliang Li. 2023. DiffuRec: A Diffusion Model for Sequential Recommendation.ACM Trans. Inf. Syst.42, 3 (2023)

2023
[17]

Qidong Liu, Fan Yan, Xiangyu Zhao, Zhaocheng Du, Huifeng Guo, Ruiming Tang, and Feng Tian. 2023. Diffusion Augmentation for Sequential Recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1576–1586

2023
[18]

Yu, Julian McAuley, and Caiming Xiong

Zhiwei Liu, Yongjun Chen, Jia Li, Philip S. Yu, Julian McAuley, and Caiming Xiong. 2021. Contrastive Self-Supervised Sequential Recommendation with Robust Augmentation. arXiv:2108.06479

2021 arXiv
[19]

Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Zhan- hui Kang. 2024. Plug-In Diffusion Model for Sequential Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 8886–8894

2024
[20]

Haokai Ma, Ruobing Xie, Lei Meng, Yimeng Yang, Xingwu Sun, and Zhanhui Kang. 2024. SeeDRec: Sememe-based Diffusion for Sequential Recommendation. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24. 2270–2278

2024
[21]

Wenyu Mao, Shuchang Liu, Haoyang Liu, Haozhe Liu, Xiang Li, and Lantao Hu. 2025. Distinguished Quantized Guidance for Diffusion-based Sequence Recommendation. InProceedings of the ACM on Web Conference 2025. 425–435

2025
[22]

Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved Denoising Diffusion Probabilistic Models. InProceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139). PMLR, 8162–8171

2021
[23]

Zipeng Qi, Lichen Bai, Haoyi Xiong, and Zeke Xie. 2024. Not All Noises Are Created Equally: Diffusion Noise Selection and Optimization. arXiv:2407.14041

2024 arXiv
[24]

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 ...

2023
[25]

Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. 2022. Contrastive Learn- ing for Representation Degeneration Problem in Sequential Recommendation. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. 813–823

2022
[26]

Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing Personalized Markov Chains for Next-Basket Recommendation. InProceed- ings of the 19th International Conference on World Wide Web. 811–820

2010
[27]

Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. 2021. Contrastive Learning with Hard Negative Samples. InInternational Conference on Learning Representations

2021
[28]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

2022
[29]

Sara Mah- davi, Raphael Gontijo-Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi

Chitwan Saharia, William Chan, Saurabh Saxena, Lala Lit, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mah- davi, Raphael Gontijo-Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. 2022. Photorealistic Text-to-Image Di...

2022
[30]

Joo-yeong Song and Bongwon Suh. 2022. Data Augmentation Strategies for Improving Sequential Recommender Systems. arXiv:2203.14037

2022 arXiv
[31]

Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang
[32]

Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining. 565–573

2018
[33]

Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748

2018 arXiv
[34]

Laurens Van Der Maaten. 2014. Accelerating t-SNE using Tree-Based algorithms. Journal of Machine Learning Research15, 1 (2014), 3221–3245

2014
[35]

Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua
[36]

Jason Wei and Kai Zou. 2019. EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Pr...

2019
[37]

Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive Learning for Sequential Recommendation. In2022 IEEE 38th International Conference on Data Engineering. IEEE, 1259–1273

2022
[38]

An Yan, Shuo Cheng, Wang-Cheng Kang, Mengting Wan, and Julian McAuley
[39]

Zhengyi Yang, Jiancan Wu, Zhicai Wang, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2023. Generate What You Prefer: Reshaping Sequential Recommen- dation via Guided Diffusion. InProceedings of the 37th International Conference on Neural Information Processing Systems. 15 pages

2023
[40]

Jose, and Xiangnan He

Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. 2019. A Simple Convolutional Generative Network for Next Item Recommendation. InProceedings of the Twelfth ACM International Conference on Web Search and Data Mining. 582–590

2019
[41]

Bin Zhao, Xuelong Li, and Xiaoqiang Lu. 2017. Hierarchical Recurrent Neural Network for Video Summarization. InProceedings of the 25th ACM International Conference on Multimedia. 863–871

2017
[42]

InProceedings of the 28th ACM International Conference on Information and Knowledge Management

CosRec: 2D Convolutional Neural Networks for Sequential Recommenda- tion. InProceedings of the 28th ACM International Conference on Information and Knowledge Management. 2173–2176
[43]

Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Com...

2021
[44]

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. InPro- ceedings of the 29th ACM International Conference on Inf...

2020
[45]

Peilin Zhou, Jingqi Gao, Yueqi Xie, Qichen Ye, Yining Hua, Jaeboum Kim, Shoujin Wang, and Sunghun Kim. 2023. Equivariant Contrastive Learning for Sequential SimDiffRec: Semantic Similarity-Guided Diffusion for Contrastive Sequential Recommendation WSDM ’26, February 22–26, 202...

2023
[46]

Jujia Zhao, Wang Wenjie, Yiyan Xu, Teng Sun, Fuli Feng, and Tat-Seng Chua. 2024. Denoising Diffusion Recommender Model. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1370–1379

2024
[47]

Sharare Zolghadr, Ole Winther, and Paul Jeha. 2024. Generative Diffusion Models for Sequential Recommendations. arXiv:2410.19429

2024 arXiv
[50]

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 Web Conference 2024. 3854–3863

2024
[2016]

In Proceedings of the International Conference on Learning Representations

Session-Based Recommendations with Recurrent Neural Networks. In Proceedings of the International Conference on Learning Representations
[2019]

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
[2023]

InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval

Diffusion Recommender Model. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 832–841

Pith tools

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

[1] [1]

Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley, and Caiming Xiong. 2022. Intent Contrastive Learning for Sequential Recommendation. InProceedings of WSDM ’26, February 22–26, 2026, Boise, ID, USA Jinkyeong Choi, Yejin Noh, and Donghyeon Park the ACM Web Conference 2022. 2172–2182

work page 2022

[2] [2]

Zhuowei Chen, Lianxi Wang, Yuben Wu, Xinfeng Liao, Yujia Tian, and Junyang Zhong. 2024. An Effective Deployment of Diffusion LM for Data Augmentation in Low-Resource Sentiment Classification. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 1844–1856

work page 2024

[3] [3]

Ziqiang Cui, Haolun Wu, Bowei He, Ji Cheng, and Chen Ma. 2024. Context Matters: Enhancing Sequential Recommendation with Context-aware Diffusion- based Contrastive Learning. InProceedings of the 33rd ACM International Confer- ence on Information and Knowledge Management. 404–414

2024

[4] [4]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion Models Beat GANs on Image Synthesis. InAdvances in Neural Information Processing Systems, Vol. 34. 8780–8794

work page 2021

[5] [5]

Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. InProceedings of the 2021 Conference on Em- pirical Methods in Natural Language Processing. 6894–6910

2021

[6] [6]

Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. 2023. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. InThe Eleventh International Conference on Learning Representations

work page 2023

[7] [7]

Zhengfu He, Tianxiang Sun, Qiong Tang, Kuanning Wang, Xuanjing Huang, and Xipeng Qiu. 2023. DiffusionBERT: Improving Generative Masked Language Models with Diffusion Models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 4521–4534

work page 2023

[8] [8]

Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent Neural Networks with Top-k Gains for Session-based Recommendations. InProceedings of the 27th ACM international conference on information and knowledge management. 843–852

work page 2018