Flow Matching based Sequential Recommender Model

Chenliang Li; Dan Luo; Feng Liu; Liming Dong; Lixin Zou; Min Tang; Xiangyang Luo; Xiangyu Zhao

REVIEW 2 major objections 7 minor 43 references

Flow Matching based Sequential Recommender Model

T0 review · 2 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Flow matching, a straight-path generative model with a deterministic Euler reverse sampler, claims a 6.53% average improvement over state-of-the-art sequential recommenders on four benchmark datasets.

desk verdict Interesting flow-matching application with a load-bearing time-direction error in the reverse sampler; the paper is not acceptable as written, but the idea and code deserve referee scrutiny. read the letter →

arxiv 2505.16298 v1 pith:D4LA62JE submitted 2025-05-22 cs.IR

Feng Liu , Lixin Zou , Xiangyu Zhao , Min Tang , Liming Dong , Dan Luo , Xiangyang Luo , Chenliang Li This is my paper · ORCID

classification cs.IR

keywords flowmatchingsequentialrecommendationdiffusionmodelsstraighttrajectorydeterministicreversesamplertarget-predictionlossreconstructionnext-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 argues that the dominant way diffusion models are applied to sequential recommendation builds noise into exactly the wrong places. In the forward pass, a curved noise schedule lengthens the path from target item to noise, accumulating errors that the reverse pass must undo; in the reverse pass, a stochastic sampler injects randomness that can turn a 'watch' recommendation into 'milk.' FMRec replaces both. It uses a straight flow trajectory $z_t = (1-t)x_c + t x_n$, trains the network to predict the clean target embedding $x_c$ rather than the vector field, adds a cross-entropy regularizer and an interaction-reconstruction loss, and decodes at inference with a deterministic Euler solver. On Beauty, Steam, MovieLens-100k, and Yelp the model reports an average 6.53% improvement over state-of-the-art baselines, with HR and NDCG gains of roughly 3.3–9.9%.

What carries the argument

The load-bearing object is the straight-line interpolation of Equation (8), $z_t = (1-t)x_c + t x_n$, together with its reparameterized learning target. Rather than regressing the velocity field $u_t = x_n - x_c$, the paper's decoder predicts the clean target $x_c$, which is trained with $L_{FM} = \mathbb{E}\|f_{\Theta}(z_t, t) - x_c\|^2$, and the reverse vector field is reconstructed as $f_{\Theta}(z_t, t) - x_n$ for the Euler update $z_{t+\Delta t} = z_t + \Delta t(f_{\Theta}(z_t,t) - x_n)$. The straightness of the path makes the higher-order terms in the Taylor expansion vanish, so the Euler method becomes exact; the deterministic sampler fixes the output given the same historical sequence and initial noise. Two regularizers carry secondary weight: a cross-entropy term that keeps item embeddings distinguishable and an MSE term that reconstructs the user's interaction vector from the decoder's final hidden state.

What would settle it

A decisive check would be to invert the integration order: set $z_1 = x_n$ with t = 1, integrate $d z_t = -(f_{\Theta}(z_t,t) - x_n) dt$ from t=1 down to t=0 with step $-\Delta t$, and output $z_0$. If the reported HR@10/NDCG@10 gains largely vanish under this convention-correct reverse process, then the success of Algorithm 1 depends on its reversed time indexing rather than on the straight trajectory and deterministic sampler; if they persist, the mismatch is harmless.

Watch

Extended reading notes

Core claim

The central discovery, stated by the authors, is that a flow-matching formulation fits sequential recommendation better than a conventional diffusion formulation because it lets the model treat the next item's embedding as the clean endpoint of a straight, easily reversible corruption process. FMRec defines the forward process as $z_t = (1-t)x_c + t x_n$, a straight line from the target embedding $x_c$ to standard normal noise $x_n$, and the decoder learns $f_{\Theta}(z_t, t)$ that predicts $x_c$ directly. Three additional components are claimed to make this effective for recommendation: a cross-entropy loss that prevents the trainable item embeddings from collapsing onto one another; a reconstruction loss that decodes the user's historical interaction vector from the last hidden state, keeping preference information alive under noise; and a deterministic Euler-integration reverse process that removes the stochastic perturbations of SDE-based samplers. Because the trajectory is straight, the Euler update has zero truncation error, so the reverse process needs only a finite number of steps to match the continuous ODE. The paper's claim is that this combination produces the reported 6.53% average improvement over state-of-the-art methods, including clear wins over the diffusion-based DreamRec and DiffuRec.

Load-bearing premise

The method assumes that a model trained on a schedule in which t=0 means 'clean item' and t=1 means 'pure noise' still guides the reverse loop correctly when Algorithm 1 starts from noise at t=0 and advances to t=1, which is the opposite time direction from the one seen in training.

Editorial extensions

If this is right

The paper's reported results would make FMRec the strongest generative next-item predictor among the compared methods on all four datasets, with an average 6.53% advantage.
Because straight trajectories make Euler integration exact, the number of reverse steps could be reduced below the reported 30 without the discretization error that curved schedules incur.
With a deterministic reverse sampler, recommendation generation becomes reproducible: identical historical sequences and identical initial noise produce identical ranked lists, which is a directly testable property.
The ablation against v-prediction shows that the modified target-prediction loss, not just the flow-matching backbone, is claimed to be responsible for most of the accuracy gain.
The reconstruction loss and cross-entropy regularizer are argued to be essential, since removing or weakening either degrades performance on the Beauty dataset.

Reading between the lines

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

A natural next test is to run the same straight-trajectory flow model with a stochastic sampler (or the same deterministic sampler on the cosine trajectory) to separate the contribution of the deterministic reverse process from the contribution of the straight path; the paper does not run this cross-ablation.
Because the model predicts a clean item embedding from noise, the machinery could transfer to other dense-vector retrieval tasks that need a single deterministic decode, such as session-based search or candidate generation in advertising, provided a suitable history encoder exists.
The interaction-reconstruction loss currently uses a binary interaction vector over all items; a more scalable variant would be to replace the full-item MSE with a sampled contrastive loss, which would matter for catalogs much larger than those tested.
The deterministic Euler sampler makes the model a deterministic function of (history, initial noise); this means the method could be used as a fixed 'generator' inside a larger pipeline, such as reranking or reinforcement-learning loops, where reproducibility of the candidate set is valuable.

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

The new thing here is real: applying flow matching to sequential recommendation with a straight trajectory, a target-prediction loss (predict x_c instead of the vector field), a reconstruction loss, and a deterministic Euler sampler. That combination is new relative to the cited diffusion recommenders, and the empirical tables look internally consistent, with gains around 3–10% on four benchmarks. The L_FM derivation is standard flow-matching algebra, and the paper is generally clear about the forward process.

The problem is the reverse process. Under Eq. (8), z_t = (1-t)x_c + t x_n, so t=0 is the data endpoint and t=1 is the noise endpoint. Training supervises f_theta(z_t, t) exactly on those forward-time pairs. Algorithm 1, however, initializes z_0 = x_n and loops t from 0 to 1, calling f_theta(z_t, t) with increasing t. The first call is f_theta(noise, 0), a pair never seen in training; the last call is f_theta(data-ish, 1), also unseen. The straight-trajectory argument in Appendix C does not fix this, because the network is evaluated out-of-distribution. The same issue appears in Eq. (22)–(23), where the ODE is integrated forward in t. This is load-bearing: every reported improvement depends on that sampler. A correct deterministic reverse should start from noise at forward time t=1 and integrate to t=0, or equivalently call the network with 1-tau. This is not a minor typo; it is an internal contradiction with the paper’s own equations.

There are additional soft spots. Section 5.1 says s=1.0, while Appendix B.2 says s=0.4 is optimal. DimeRec is cited in related work but missing from the baseline table. And Eq. (23) subtracts the fixed initial noise x_n rather than the current state, which is not obviously the Euler discretization of the learned vector field even with correct time indexing.

The paper has enough novelty and enough reproducibility (code is released) to warrant a serious referee. But as written, the central inference procedure is incoherent. If the released code actually implements a time-reversed sampler, that would resolve the main concern; otherwise the empirical gains cannot be attributed to the described method. This deserves peer review, not a desk reject, but the authors need to fix the algorithm description and address the s-contradiction before the numbers can be trusted. Reader's take is about right; I would not cite it until the sampler is corrected.

Referee Report

2 major / 7 minor

Summary. The paper proposes FMRec, a sequential recommender model that replaces the diffusion forward/reverse processes with flow matching. The forward process uses a straight trajectory z_t=(1-t)x_c+t x_n, a modified loss that trains a network f_theta to directly predict the target item embedding x_c rather than the velocity, a cross-entropy regularizer, and a reconstruction loss over user history. The reverse process uses a deterministic Euler sampler. Experiments on Beauty, Steam, MovieLens-100k, and Yelp report an average 6.53% improvement over the best baseline in HR/NDCG, and ablations support the modified loss, straight trajectory, and regularized losses.

Significance. If the reported results are reproducible, the paper makes a useful contribution by showing that a simple straight-line flow matching formulation with deterministic sampling improves over diffusion-based sequential recommenders while simplifying inference. The paper releases code, includes ablations for the loss and trajectory choices, and the derivation of the modified flow-matching loss is standard up to a typo. The main obstacle is the reverse-sampler time mismatch, which currently makes the described inference procedure inconsistent with the training objective and must be resolved before the experimental claims can be credited.

major comments (2)

[Section 4.2 (Reverse Process), Eqs. (22)-(23), and Algorithm 1] The reverse sampler is executed from t=0 to t=1 starting at z_0=x_n. Under Eq. (8), t=0 is the data endpoint (z_0=x_c) and t=1 is the noise endpoint (z_1=x_n); the training loss in Eq. (13) therefore supervises f_theta only on states z_t=(1-t)x_c+t x_n. Algorithm 1 queries f_theta on exactly the reversed pairing: the first call is f_theta(x_n,0), the last is f_theta(x_c,1), and every intermediate call uses (z_t,t) with z_t=(1-t)x_n+t x_c, a state that training associates with time 1-t, not t. Because t also enters the model input through Eq. (14) via the term (z_t+t), this is a genuine distribution shift rather than a harmless reparametrization. As written, the paper does not describe the flow-matching inference procedure that produced Table 2; the straight-trajectory argument in Appendix C applies only when the learned vector field is evaluated at its training-time pairs. Please correct the algorithm (e.g., integrate t from 1 down to 0 starting at z_1=x_n, or introduce s=1-t and query f_theta(z_s,1-s)) and confirm that the released code implements the corrected procedure.
[Section 5.2 and Table 2] The paper's central quantitative claim is an average 6.53% improvement over the best baselines, computed by averaging the 24 Delta values in Table 2. However, no standard deviations, number of random seeds, or significance tests are reported. Several improvements are small (e.g., ML-100k NDCG@5, 3.34%; Yelp NDCG@20, 5.19%), so without variance information it is not possible to determine whether the gains are reliable. Please report the experimental variance and, if feasible, paired significance tests, or explicitly state the number of runs used for Table 2.

minor comments (7)

[Section 4.2, after Eq. (12)] The parenthetical 'The first derivation relies on the fact x_n = phi^{-1}_t(z_t|x_n)' should read x_c = phi^{-1}_t(z_t|x_n); the inverse of the flow returns the data point, not the noise.
[Throughout] The model name is inconsistent: FMREC in the abstract and text, FMRec in Figure 3 and the repository URL, and FM4Rec in Table 2. Please unify the notation.
[Algorithm 1] Algorithm 1 lists 'Scaling factor for Timesteps Schedules' as an input but never uses it; either remove it or explain its role in inference.
[Section 5.1 vs. Appendix B.2] Section 5.1 sets the mode-sampling scaling s=1.0, while Appendix B.2 reports that s=0.4 yields the optimal performance on the same Beauty and MovieLens-100k datasets; please clarify which configuration was used for Table 2 and whether s was tuned per dataset.
[Section 5.2] The sentence 'FMREC outperforms both DreamRec and DiffuRec, achieving superior results with an average improvement of 12.44% on HR@5' is ambiguous: 12.44% is the average improvement over DiffuRec alone, not over both baselines. Please state the comparison explicitly.
[Figures 4 and 7] Figures 4 and 7 contain undisplayed Unicode artifacts such as '/uni00000013/...'; these should be removed or properly rendered.
[Eq. (9) and Section 1] The mode-sampling formula in Eq. (9) is hard to read because the superscript is missing; please typeset t = 1 - k^{-s}(cos^2(pi*k/2) - 1 + k). Also, 'Flow Fatching' in the Introduction should be 'Flow Matching'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the flow-matching loss is a standard reparameterization, the trainable target is regularized by L_CE and L_MSE, and the central comparison is against external benchmarks; the reverse-time inconsistency is a correctness risk, not a circular step.

full rationale

FMRec's derivation chain is self-contained. The modified flow-matching loss in Eq. (13) is obtained from the standard conditional flow-matching objective (Eq. (4)) by substituting the straight path z_t=(1-t)x_c+t x_n and reparameterizing the learned vector field as -f_theta(z_t,t)+x_n; this is algebraic parameter substitution, not a fit of the evaluation metric. The trainable target embedding e_{m+1} is a deliberate modeling choice, and the paper explicitly introduces L_CE to prevent the collapse that would make prediction trivial ("This might lead to the problem that different x_c converge to a single embedding... To mitigate this issue, we introduce a cross-entropy loss L_CE"). The final HR/NDCG numbers are computed on held-out next items against eight external baselines, so the reported 6.53% improvement is not forced by construction. The straight-trajectory advantage in Appendix C is a mathematical property of linear interpolation (the second derivative vanishes), independent of any fitted parameter or self-citation. Self-citations to DiffuRec and other group work appear in related work and in borrowing the input fusion rule of Eq. (14), but they are not used as a uniqueness theorem or as the proof of the central claim; the central claim is tested against those very methods on public benchmarks. The reverse-sampler time direction in Algorithm 1 and Eq. (23) is inconsistent with the forward convention of Eq. (8): the loop starts from z_0=x_n at t=0 and queries f_theta at (noise,0), a pair never seen in training. This is a correctness/reproducibility risk that would invalidate the described procedure if implemented literally, but it does not make the reported gains equivalent to an input by construction, so it is not a circularity step under the quoted-evidence standard.

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

The central claim rests on tuned hyperparameters (delta, alpha, beta, s, q), a standard flow matching path, and several domain assumptions about how noise can be injected into item embeddings without destroying the signal needed for next-item ranking. The most fragile premise is the reverse-time integration in Algorithm 1, which assumes the learned field can be evaluated at time/state pairs not present in training.

free parameters (5)

delta (noise fusion variance) = 0.001
Controls lambda_i ~ N(delta,delta) in Eq (14); tuned on Beauty and MovieLens-100k per Figure 5.
alpha (L_CE weight) = 0.2
Weights the cross-entropy loss in Eq (21); tuned via grid search in Figure 4.
beta (L_MSE weight) = 0.4
Weights the reconstruction loss in Eq (21); tuned via grid search in Figure 4.
s (mode sampling scaling) = 1.0 (Section 5.1) vs 0.4 (Appendix B.2)
Shapes the timestep sampling density in Eq (9); the paper reports both values, creating ambiguity about which was used for Table 2.
q (Euler steps) = 30
Number of reverse integration steps in Algorithm 1; hand-set, not shown to be optimal.

assumptions (5)

domain assumption Flow matching probability path with a_t=1-t and b_t=t defines a valid generative process between item embeddings and Gaussian noise.
Used in Eq (8); assumes linear interpolation is a meaningful preference path for next-item recommendation.
standard math The inverse map phi_t^{-1}(z_t|x_n) recovers x_c for straight paths.
Used in Eq (10)-(13); standard algebra, though the text has a typo writing x_n instead of x_c.
domain assumption The last hidden state h_m of the unidirectional decoder contains enough information to reconstruct the user's full interaction vector r.
L_MSE in Eq (19)-(20) assumes a 128-d vector can be decoded to an |I|-dimensional interaction history.
domain assumption Adding noise to historical embeddings via E_i=e_i+lambda_i*(z_t+t) preserves user preferences.
Eq (14) is the input fusion rule; no proof that this mixture is information-preserving for ranking.
ad hoc to paper The reverse integration can start at pure noise and end at the target embedding with a deterministic Euler sampler using the same t index as forward training.
Algorithm 1 uses z_0=x_n with forward time t; this is inconsistent with Eq (8) unless time is reversed, so the paper implicitly assumes the learned field generalizes to unseen (state, t) pairs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Flow Matching based Sequential Recommender Model." pith.science (2026). https://pith.science/paper/D4LA62JE

@misc{pith2026250516298,
  author       = {Pith},
  title        = {Pith review of: Flow Matching based Sequential Recommender Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4LA62JE}},
  note         = {Machine review of arXiv:2505.16298}
}

read the original abstract

Generative models, particularly diffusion model, have emerged as powerful tools for sequential recommendation. However, accurately modeling user preferences remains challenging due to the noise perturbations inherent in the forward and reverse processes of diffusion-based methods. Towards this end, this study introduces FMRec, a Flow Matching based model that employs a straight flow trajectory and a modified loss tailored for the recommendation task. Additionally, from the diffusion-model perspective, we integrate a reconstruction loss to improve robustness against noise perturbations, thereby retaining user preferences during the forward process. In the reverse process, we employ a deterministic reverse sampler, specifically an ODE-based updating function, to eliminate unnecessary randomness, thereby ensuring that the generated recommendations closely align with user needs. Extensive evaluations on four benchmark datasets reveal that FMRec achieves an average improvement of 6.53% over state-of-the-art methods. The replication code is available at https://github.com/FengLiu-1/FMRec.

Figures

Figures reproduced from arXiv: 2505.16298 by the authors.

**Figure 2.** The framework of the FMREC. In the training process, our design incorporates the development of straight flow trajectories, modifications to the learning target LFM, the design of a decoder-based model, and the implementation of regularized loss functions LCE and LMSE. In the inference process, we present a deterministic reverse sampler that generates recommendations. define the time-dependent flow of the next item … view at source ↗

**Figure 3.** Performance comparison based on different flow trajec [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Comparison of model performance across various [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 6.** Figure 6: Performance comparison based on different timesteps [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: Comparison of model performance across various [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 28 canonical work pages

[1]

News session-based recommendations using deep neural networks

[de Souza Pereira Moreiraet al., 2018 ] Gabriel de Souza Pereira Moreira, Felipe Ferreira, and Adilson Marques Da Cunha. News session-based recommendations using deep neural networks. InProceedings of the 3rd workshop on deep learning for recommender systems, pages 15–23,

work page 2018
[5]

Scaling rectified flow transformers for high- resolution image synthesis

[Esseret al., 2024 ] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boe- sel, et al. Scaling rectified flow transformers for high- resolution image synthesis. InForty-first International Conference on Machine Learning,

work page 2024
[7]

Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933,

[Gonget al., 2022 ] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933,

arXiv 2022
[8]

Dualvae: Dual disentangled variational autoencoder for recommendation

[Guoet al., 2024 ] Zhiqiang Guo, Guohui Li, Jianjun Li, Chaoyang Wang, and Si Shi. Dualvae: Dual disentangled variational autoencoder for recommendation. InProceed- ings of the 2024 SIAM International Conference on Data Mining (SDM), pages 571–579. SIAM,

work page 2024
[9]

The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19,

[Harper and Konstan, 2015] F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19,

work page 2015
[11]

Session- based recommendations with recurrent neural networks

[Hidasiet al., 2015 ] Bal´azs Hidasi, Alexandros Karat- zoglou, Linas Baltrunas, and Domonkos Tikk. Session- based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939,

arXiv 2015
[12]

Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851,

[Hoet al., 2020 ] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851,

2020
[14]

Dual conditional diffusion models for sequential recommendation.arXiv preprint arXiv:2410.21967,

[Huanget al., 2024 ] Hongtao Huang, Chengkai Huang, Xi- aojun Chang, Wen Hu, and Lina Yao. Dual conditional diffusion models for sequential recommendation.arXiv preprint arXiv:2410.21967,

arXiv 2024

Show all 43 references

[15]

Dgrm: Diffusion-gan recommendation model to alleviate the mode collapse problem in sparse environments.Pattern Recognition, page 110692,

[Jiangzhouet al., 2024 ] Deng Jiangzhou, Wang Songli, Ye Jianmei, Ji Lianghao, and Wang Yong. Dgrm: Diffusion-gan recommendation model to alleviate the mode collapse problem in sparse environments.Pattern Recognition, page 110692,

2024
[16]

Self-attentive sequential recommendation

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

2018
[17]

Dif- furec: A diffusion model for sequential recommendation

[Liet al., 2023 ] Zihao Li, Aixin Sun, and Chenliang Li. Dif- furec: A diffusion model for sequential recommendation. ACM Transactions on Information Systems, 42(3):1–28,

2023
[18]

Dimerec: A unified framework for en- hanced sequential recommendation via generative diffu- sion models.arXiv preprint arXiv:2408.12153,

[Liet al., 2024b ] Wuchao Li, Rui Huang, Haijun Zhao, Chi Liu, Kai Zheng, Qi Liu, Na Mou, Guorui Zhou, Defu Lian, Yang Song, et al. Dimerec: A unified framework for en- hanced sequential recommendation via generative diffu- sion models.arXiv preprint arXiv:2408.12153,

arXiv
[19]

Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

[Lipmanet al., 2022 ] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

2022 arXiv
[20]

Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

[Liuet al., 2022 ] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

2022 arXiv
[21]

Autoseqrec: Autoencoder for efficient sequential recommendation

[Liuet al., 2023 ] Sijia Liu, Jiahao Liu, Hansu Gu, Dong- sheng Li, Tun Lu, Peng Zhang, and Ning Gu. Autoseqrec: Autoencoder for efficient sequential recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 1493– 1502,

2023
[23]

Justifying recommendations using distantly- labeled reviews and fine-grained aspects

[Niet al., 2019 ] Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantly- labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conferen...

2019
[26]

Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

[Songet al., 2020 ] Jiaming Song, Chenlin Meng, and Ste- fano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

2020 arXiv
[27]

Bert4rec: Se- quential recommendation with bidirectional encoder rep- resentations from transformer

[Sunet al., 2019 ] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Se- quential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowl-...

2019
[28]

Person- alized top-n sequential recommendation via convolutional sequence embedding

[Tang and Wang, 2018] Jiaxi Tang and Ke Wang. Person- alized top-n sequential recommendation via convolutional sequence embedding. InProceedings of the eleventh ACM international conference on web search and data mining, pages 565–573,

2018
[29]

Unbiased recommendation through invariant representation learning

[Tanget al., 2024 ] Min Tang, Lixin Zou, Shujie Cui, Shiuan- ni Liang, and Zhe Jin. Unbiased recommendation through invariant representation learning. InJoint European Con- ference on Machine Learning and Knowledge Discovery in Databases, pages 280–296. Springer,

2024
[30]

Sequential rec- ommendation by reprogramming pretrained transformer

[Tanget al., 2025 ] Min Tang, Shujie Cui, Zhe Jin, Shiuan- ni Liang, Chenliang Li, and Lixin Zou. Sequential rec- ommendation by reprogramming pretrained transformer. Information Processing & Management, 62(1):103938,

2025
[32]

Diffusion recommender model

[Wanget al., 2023 ] Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua. Diffusion recommender model. InProceedings of the 46th Interna- tional ACM SIGIR Conference on Research and Develop- ment in Information Retrieval, pages 832–841,

2023
[33]

Conditional denoising diffusion for se- quential recommendation

[Wanget al., 2024 ] Yu Wang, Zhiwei Liu, Liangwei Yang, and Philip S Yu. Conditional denoising diffusion for se- quential recommendation. InPacific-Asia Conference on Knowledge Discovery and Data Mining, pages 156–169. Springer,

2024
[34]

Recurrent recom- mender networks

[Wuet al., 2017 ] Chao-Yuan Wu, Amr Ahmed, Alex Beu- tel, Alexander J Smola, and How Jing. Recurrent recom- mender networks. InProceedings of the tenth ACM inter- national conference on web search and data mining, pages 495–503,

2017
[35]

Pd-gan: Adversarial learning for personalized diversity-promoting recommen- dation

[Wuet al., 2019 ] Qiong Wu, Yong Liu, Chunyan Miao, Bin- qiang Zhao, Yin Zhao, and Lu Guan. Pd-gan: Adversarial learning for personalized diversity-promoting recommen- dation. InIJCAI, volume 19, pages 3870–3876,

2019
[36]

Ar-diffusion: Auto-regressive diffusion model for text generation.Advances in Neural Information Processing Systems, 36:39957–39974,

[Wuet al., 2023 ] Tong Wu, Zhihao Fan, Xiao Liu, Hai-Tao Zheng, Yeyun Gong, Jian Jiao, Juntao Li, Jian Guo, Nan Duan, Weizhu Chen, et al. Ar-diffusion: Auto-regressive diffusion model for text generation.Advances in Neural Information Processing Systems, 36:39957–39974,

2023
[37]

Sheng, Zhiming Cui, Xiaofang Zhou, and Hui Xiong

[Xuet al., 2019 ] Chengfeng Xu, Pengpeng Zhao, Yanchi Liu, Jiajie Xu, Victor S Sheng S. Sheng, Zhiming Cui, Xiaofang Zhou, and Hui Xiong. Recurrent convolutional neural network for sequential recommendation. InThe world wide web conference, pages 3398–3404,

2019
[38]

Gener- ate what you prefer: Reshaping sequential recommenda- tion via guided diffusion.Advances in Neural Information Processing Systems, 36,

[Yanget al., 2024 ] Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. Gener- ate what you prefer: Reshaping sequential recommenda- tion via guided diffusion.Advances in Neural Information Processing Systems, 36,

2024
[39]

Graph masked autoencoder for sequential recom- mendation

[Yeet al., 2023 ] Yaowen Ye, Lianghao Xia, and Chao Huang. Graph masked autoencoder for sequential recom- mendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, pages 321–330,

2023
[40]

Fastervd: on accel- eration of video diffusion models

[Yuet al., 2024 ] Pinrui Yu, Dan Luo, Timothy Rupprecht, Lei Lu, Zhenglun Kong, Pu Zhao, Yanyu Li, Octavia Camps, Xue Lin, and Yanzhi Wang. Fastervd: on accel- eration of video diffusion models. InProceedings of the Thirty-Third International Joint Conference on Artificial Int...

2024
[41]

Uniform graph pre- training and prompting for transferable recommendation

[Yuet al., 2025 ] Qing Yu, Lixin Zou, Xiangyang Luo, Xi- angyu Zhao, and Chenliang Li. Uniform graph pre- training and prompting for transferable recommendation. ACM Transactions on Information Systems,

2025
[42]

Reinforce- ment learning to optimize long-term user engagement in recommender systems

[Zouet al., 2019 ] Lixin Zou, Long Xia, Zhuoye Ding, Ji- axing Song, Weidong Liu, and Dawei Yin. Reinforce- ment learning to optimize long-term user engagement in recommender systems. InProceedings of the 25th ACM SIGKDD international conference on knowledge discov- ery & data...

2019
[43]

Neural interactive collaborative filtering

[Zouet al., 2020 ] Lixin Zou, Long Xia, Yulong Gu, Xiangyu Zhao, Weidong Liu, Jimmy Xiangji Huang, and Dawei Yin. Neural interactive collaborative filtering. InProceed- ings of the 43rd international ACM SIGIR conference on research and development in information retrieval, pa...

2020
[2015]

Flexible diffusion modeling of long videos.Advances in Neural Information Processing Systems, 35:27953–27965,

[Harveyet al., 2022 ] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos.Advances in Neural Information Processing Systems, 35:27953–27965,

2022
[2017]

Sequential user-based recurrent neural net- work recommendations

[Donkerset al., 2017 ] Tim Donkers, Benedikt Loepp, and J¨urgen Ziegler. Sequential user-based recurrent neural net- work recommendations. InProceedings of the eleventh ACM conference on recommender systems, pages 152– 160,

2017
[2018]

Gan-based matrix factorization for recom- mender systems

[Dervishaj and Cremonesi, 2022] Ervin Dervishaj and Paolo Cremonesi. Gan-based matrix factorization for recom- mender systems. InProceedings of the 37th ACM/SIGAPP Symposium on Applied Computing, pages 1373–1381,

2022
[2019]

Improved denoising diffusion proba- bilistic models

[Nichol and Dhariwal, 2021] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion proba- bilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR,

2021
[2020]

Video diffusion models.Advances in Neural Infor- mation Processing Systems, 35:8633–8646,

[Hoet al., 2022 ] Jonathan Ho, Tim Salimans, Alexey Grit- senko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advances in Neural Infor- mation Processing Systems, 35:8633–8646,

2022
[2021]

Autorec: Autoen- coders meet collaborative filtering

[Sedhainet al., 2015 ] Suvash Sedhain, Aditya Krishna Menon, Scott Sanner, and Lexing Xie. Autorec: Autoen- coders meet collaborative filtering. InProceedings of the 24th international conference on World Wide Web, pages 111–112,

2015
[2022]

Long and short-term recommendations with re- current neural networks

[Devooght and Bersini, 2017] Robin Devooght and Hugues Bersini. Long and short-term recommendations with re- current neural networks. InProceedings of the 25th con- ference on user modeling, adaptation and personalization, pages 13–21,

2017
[2023]

Step-by-step diffusion: An elementary tutorial.arXiv preprint arXiv:2406.08929,

[Nakkiranet al., 2024 ] Preetum Nakkiran, Arwen Bradley, Hattie Zhou, and Madhu Advani. Step-by-step diffusion: An elementary tutorial.arXiv preprint arXiv:2406.08929,

2024 arXiv
[2024]

Sequential recommendation via stochastic self-attention

[Fanet al., 2022 ] Ziwei Fan, Zhiwei Liu, Yu Wang, Al- ice Wang, Zahra Nazari, Lei Zheng, Hao Peng, and Philip S Yu. Sequential recommendation via stochastic self-attention. InProceedings of the ACM web conference 2022, pages 2036–2047,

2022
[2025]

Attention is all you need.Ad- vances in Neural Information Processing Systems,

[Vaswani, 2017] A Vaswani. Attention is all you need.Ad- vances in Neural Information Processing Systems,

2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

News session-based recommendations using deep neural networks

[de Souza Pereira Moreiraet al., 2018 ] Gabriel de Souza Pereira Moreira, Felipe Ferreira, and Adilson Marques Da Cunha. News session-based recommendations using deep neural networks. InProceedings of the 3rd workshop on deep learning for recommender systems, pages 15–23,

work page 2018

[2] [5]

Scaling rectified flow transformers for high- resolution image synthesis

[Esseret al., 2024 ] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boe- sel, et al. Scaling rectified flow transformers for high- resolution image synthesis. InForty-first International Conference on Machine Learning,

work page 2024

[3] [7]

Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933,

[Gonget al., 2022 ] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933,

arXiv 2022

[4] [8]

Dualvae: Dual disentangled variational autoencoder for recommendation

[Guoet al., 2024 ] Zhiqiang Guo, Guohui Li, Jianjun Li, Chaoyang Wang, and Si Shi. Dualvae: Dual disentangled variational autoencoder for recommendation. InProceed- ings of the 2024 SIAM International Conference on Data Mining (SDM), pages 571–579. SIAM,

work page 2024

[5] [9]

The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19,

[Harper and Konstan, 2015] F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19,

work page 2015

[6] [11]

Session- based recommendations with recurrent neural networks

[Hidasiet al., 2015 ] Bal´azs Hidasi, Alexandros Karat- zoglou, Linas Baltrunas, and Domonkos Tikk. Session- based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939,

arXiv 2015

[7] [12]

Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851,

[Hoet al., 2020 ] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851,

2020

[8] [14]

Dual conditional diffusion models for sequential recommendation.arXiv preprint arXiv:2410.21967,

[Huanget al., 2024 ] Hongtao Huang, Chengkai Huang, Xi- aojun Chang, Wen Hu, and Lina Yao. Dual conditional diffusion models for sequential recommendation.arXiv preprint arXiv:2410.21967,

arXiv 2024