Flow Matching for Collaborative Filtering

Chengkai Liu; James Caverlee; Jianling Wang; Rex Ying; Yangtian Zhang

REVIEW 3 major objections 4 minor 63 references

Flow Matching for Collaborative Filtering

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read FlowCF claims to beat VAE and diffusion recommenders with two inference steps by using a behavior-guided prior and a discrete flow.

desk verdict Solid application of discrete flow matching to CF with a genuinely useful behavior-guided prior; the accuracy gains are real but the two-step inference protocol is under-ablated. read the letter →

arxiv 2502.07303 v2 pith:VCC5TNXC submitted 2025-02-11 cs.IR

Chengkai Liu , Yangtian Zhang , Jianling Wang , Rex Ying , James Caverlee This is my paper · ORCID

classification cs.IR

keywords flowmatchingcollaborativefilteringimplicitfeedbackgenerativerecommendersystemsbehavior-guidedpriordiscretebinaryinteractiondataefficientinference

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 flow matching, which learns a vector field that pushes a source distribution into a target distribution along straight trajectories, can be adapted to collaborative filtering to beat both variational-autoencoder and diffusion recommenders. The adaptation has two parts: a behavior-guided prior that starts the flow from Bernoulli samples matched to global item frequencies, and a discrete flow framework that keeps every intermediate state binary while its expectation evolves linearly. On MovieLens-1M, MovieLens-20M, and Amazon-Beauty, the resulting model FlowCF improves Recall@10 over the best baseline by 4.96%, 1.98%, and 4.62%, respectively, using only two sampling steps at inference. The practical stake is that generative recommender accuracy need not come at the cost of slow inference.

What carries the argument

The load-bearing object is the discretized linear interpolation with its expectation-level vector field. The equation $X_t = M_t \odot X_1 + (1 - M_t) \odot X_0$ keeps every intermediate state binary, while $\mathbb{E}[X_t] = t X_1 + (1 - t) X_0$ makes the flow linear in expectation; differentiating gives $u_t = X_1 - X_0$, so the model can be trained to predict the clean interaction vector $X_1$ directly rather than estimating a vector field. The behavior-guided prior, $\text{Bernoulli}(1_{|U|} \otimes f)$ with $f$ the global item frequency vector, is what makes the discrete flow stable, since ablations show random binary, uniform, or Gaussian priors degrade or collapse, especially on sparse data.

What would settle it

Initialize inference instead from an honest masked sample at the same time step (draw $M_t \sim \text{Bernoulli}(t)$ and form $X_t$ from the observed vector and a prior sample), then compare Recall@10 with the paper's two-step numbers; a large drop would show the reported accuracy relies on seeing the unmasked target at the starting step rather than on the learned flow.

Watch

Extended reading notes

Core claim

The central discovery claimed is that the discrete, binary structure of implicit feedback can be preserved inside a flow-matching generative model without giving up the straight-path efficiency of continuous flows. FlowCF samples $X_0$ from a Bernoulli prior built on global item frequencies, forms training interpolations $X_t = M_t \odot X_1 + (1 - M_t) \odot X_0$ with $M_t \sim \text{Bernoulli}(t)$, and learns an MLP $f_\theta(X_t, t)$ to predict $X_1$ directly by mean squared error. Because $\mathbb{E}[X_t] = t X_1 + (1 - t) X_0$, the vector field $d\mathbb{E}[X_t]/dt = X_1 - X_0$ is linear, and the model can be evaluated by stepping from the observed interaction vector for two Euler steps with binary thresholding. The paper claims this yields state-of-the-art Recall and NDCG on three benchmark datasets and the fastest inference among generative baselines.

Load-bearing premise

The inference procedure starts from the exact observed interaction vector as though it were a random masked mixture at a late time step, but training only ever presents such mixtures to the model; the two-step accuracy result assumes this transfer is benign.

Editorial extensions

If this is right

If FlowCF's two-step inference holds across datasets, generative recommenders can serve live traffic with inference cost near that of a single forward pass, unlike diffusion recommenders that need many denoising steps.
The simplified loss reduces flow training to predicting the clean interaction vector from a masked mixture, so the same infrastructure as denoising autoencoders can be reused.
The behavior-guided prior's ablation results imply that prior choice, not just the flow update, determines whether a discrete generative CF model trains at all on sparse data.
The expectation-level vector field gives a theoretical bridge between binary masks and continuous probability paths, so future discrete generative CF models can use ODE-style stepping while keeping binary states.

Reading between the lines

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

Going beyond the paper, the same two-part recipe of a frequency-calibrated prior plus masked interpolation should transfer to related sparse binary prediction tasks such as top-k subset selection, bundle recommendation, or multi-label classification, where the same training/inference masking mismatch would need scrutiny.
The large gap between behavior-guided and random Bernoulli priors suggests that the prior supplies the correct marginal density of positives; an ablation varying the prior's sparsity while holding the flow model fixed could isolate how much of the gain is prior calibration versus path geometry.
A natural stress test is sequential or cold-start settings, where the observed interaction vector is shorter or noisier; the paper does not report these, and two-step performance could degrade when the starting state is far from the late-training-time mixtures.

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 genuinely new piece is the behavior-guided prior: sampling X0 from a Bernoulli with item frequencies and interpolating via a mask preserves the binary structure and avoids the Gaussian-prior mismatch. The derivation from conditional flow matching to the simplified MSE loss (Eq 20) is correct; dropping the (1-t)^-2 weighting is a standard training convenience and doesn't undermine the method. The paper ships code, reports significance stars, and ablates the prior choice (Table 3), which shows the prior is load-bearing—behavior-guided prior beats uniform/Gaussian/random binary, and the discrete flow collapses on Amazon-Beauty without it. That is real evidence, not noise.

The stress-test concern is fair: Algorithm 2 starts from the exact observed X at t=(N-2)/N while training only ever sees a masked mixture at that t (with ~1/5 of entries replaced by prior samples for N=9). The paper offers no ablation with s=0, no corrupted-start experiment, and no analysis of how the distribution shift interacts with the 'preserve observed' OR-mask. That means the two-step inference is closer to a one-shot denoiser with a prescribed input than to a generative flow that truly integrates from the prior. For typical users this may be benign—the model has seen heavily overlapped inputs at t close to 1—but for long-tail users the masked training inputs at that t look different from the clean observed vector, so the global Recall gains might hide per-user degradation. This is a soft spot, not a fatal flaw; the method still learns something useful, but the state-of-the-art claim over diffusion baselines should come with an ablation showing where the performance actually comes from. Also missing: any discrete-diffusion baseline (CF-Diff, GiffCF are cited but not compared) and error bars in Table 1.

The paper deserves peer review. The math is coherent, the prior idea is transferable, and the empirical claims are concrete and verifiable. For a reader working on generative CF, it's worth reading and citing; for a referee, I'd ask for the s=0 and corrupted-start ablation and a discrete diffusion baseline before relying on the SOTA claim.

Referee Report

3 major / 4 minor

Summary. The paper proposes FlowCF, a flow-matching model for collaborative filtering. It introduces two components: a behavior-guided prior, defined as a Bernoulli distribution over the global item-frequency vector, and a discrete flow framework that interpolates binary interaction vectors via masked mixtures. The training objective is simplified to a mean squared error between the predicted and true interaction vectors, and inference is performed with only two Euler steps starting from the observed user vector. Experiments on MovieLens-1M, MovieLens-20M, and Amazon-Beauty report state-of-the-art Recall@K and NDCG@K against nine baselines, with claimed improvements of roughly 1–5% over the best baselines and faster inference than DiffRec.

Significance. If the reported empirical gains hold under closer scrutiny, FlowCF is a meaningful contribution: it is, to my knowledge, the first flow-matching recommender system, it explicitly addresses the discreteness of implicit feedback, and the two-step inference is a practical efficiency win over diffusion-based alternatives. The authors release code, which aids reproducibility. The theoretical derivation from conditional flow matching to the simplified training loss is internally consistent, and the discrete interpolation with expectation-preserving dynamics is a clean adaptation. However, the central claim of state-of-the-art accuracy depends on an empirical evaluation that currently lacks error bars and contains a training–inference distribution shift that is not ablated.

major comments (3)

[Algorithm 2 / §3.5] The inference procedure starts from the clean observed interaction vector X at t=(N-2)/N and applies two update steps, but the model is trained only on corrupted mixtures X_t = M_t ⊙ X1 + (1-M_t) ⊙ X0 (Eq. 13). For N=9, the training distribution at t=7/9 has roughly 22% of entries replaced by prior samples, whereas the inference input has 0% corruption at the same nominal t. The paper provides no ablation with a starting step s=0 or with a corrupted starting state that matches the training distribution, so the transfer of the learned denoiser to clean inputs is unverified. This is load-bearing because the 'two sampling steps' efficiency claim and the reported accuracy both rely on this off-distribution starting point.
[Table 1 / §4.2] The central empirical claim—state-of-the-art accuracy with improvements of 4.96%, 1.98%, and 4.62% over best baselines—is reported without error bars or variance across runs, even though all FlowCF results are marked with a statistically significant p<0.05. The text does not state the number of seeds, the standard deviation, or the test procedure used for significance testing. For the ML-20M dataset, the R@10 improvement over DiffRec is only 1.98%, and without variance information it is impossible to judge whether this gain is meaningful. This is a fundamental reproducibility issue for the headline result.
[§3.4] The simplification from Eq. (19) to Eq. (20) drops the (1-t)^{-2} weighting that arises from the flow-matching loss. The paper states that Eq. (20) is the 'simplified loss' but does not analyze how this change affects training: the unweighted objective upweights small-t regions and downweights the near-1 steps that are exclusively used at inference. If the goal is to claim that FlowCF trains with the flow-matching objective, the dropped weighting is a meaningful deviation; if it is a deliberate design choice, it should be justified or ablated. As written, the theoretical link between the learned predictor and the flow vector field is weaker than claimed.

minor comments (4)

[§4.1 / Algorithm 2] The starting step s in Algorithm 2 is left as a free parameter, but the text only says 'two sampling steps'. For reproducibility, the paper should state explicitly that s = N-2 for the reported experiments, and ideally show the sensitivity of the results to s.
[§3.4 / Eq. (20)] After the loss simplification, the model is effectively trained as a multi-label classifier directly on X1; the paper should discuss this connection to denoising autoencoders and explain why the flow-matching framing adds value beyond the simplified objective.
[§4.2 / Figures 3 and 4] The natural noise and random noise experiments lack a precise description of how noise is injected into the validation set and how the noise-corrupted positive/negative labels are used in evaluation; this should be clarified.
[§4.3 / Table 3] The ablation shows that FlowCF with non-behavior-guided priors can fail on Amazon-Beauty (R@10 ≈ 0.002), but the paper does not report the number of runs or discuss whether these failures are due to optimization instability or to the prior choice; a loss curve or training diagnostic would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the flow-matching loss reduces to a direct denoising objective, but that reduction is not circular and the paper's central performance claims are empirical comparisons.

full rationale

The paper's derivation chain is self-contained. The discretized interpolation (Eq. 13) defines X_t as a Bernoulli mixture of the target X_1 and prior sample X_0; the expectation is linear (Eq. 15), giving the conditional vector field u_t = (X_1 - E[X_t])/(1-t) (Eq. 17). Substituting the predicted field v_t with hat X_1 yields Eq. 19, which simplifies to Eq. 20, E||hat X_1 - X_1||^2. This is a standard denoising / masked-reconstruction objective: the model is trained to recover the conditioning target X_1 from a corrupted version X_t. Nothing in this chain defines X_1 in terms of the model's own output; the target is the observed interaction matrix, and the prediction is the model's function of X_t. The behavior-guided prior is estimated from global training-item frequencies and used as a source distribution; it is an input to training, not a renamed prediction of held-out interactions. The ablation in Table 3 shows the prior matters empirically, but that is a sensitivity result, not circular reasoning. At inference, Algorithm 2 initializes X_t to the clean observed vector X, whereas training corrupts X_t by mixing in prior samples; this train/inference distribution mismatch is a legitimate correctness concern (the model may not generalize from masked training inputs to the clean starting state), but it is not a logical circularity: no fitted parameter is being reported as a prediction, and no load-bearing claim is justified by definition. The paper's self-citations (e.g., [20], [23]-[25]) appear only in related-work context or as prior methodology and are not used to justify the central result. The reported Recall/NDCG gains are empirical comparisons against external baselines on held-out test splits, so the central claim is not forced by the paper's own construction.

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

The central learning objective depends on one tuned discretization parameter N, a chosen starting step for inference, a data-derived popularity prior, and several modeling assumptions inherited from flow matching and discrete step interpolations. These are listed above.

free parameters (5)

Number of discretization steps N = 9 for ML-1M and Amazon-Beauty; 50 for ML-20M
Selected by validation NDCG@10; controls the granularity of the interpolation set (Figure 5).
Inference starting step s = N-2 (second-to-last step)
Chosen to minimize sampling steps; the two-step inference claim depends on this choice, and no sensitivity analysis is provided.
Number of sampling steps at inference = 2
Fixed for all datasets; a key factor in the reported inference speed.
MLP hidden sizes = [300, 300] for ML datasets; [600] for Amazon-Beauty
Hyperparameters set by hand; no ablation is reported.
Dropout rate = 0 for FlowCF
Fixed at 0 while baseline models are tuned within [0, 0.5], a potential fairness concern in comparison.

assumptions (5)

standard math Flow matching objective equivalence: L_CFM has identical gradients to L_FM (Lipman et al., 2022)
Invoked in Section 2.2, Equations 6 and 7, as the basis for the training loss.
standard math Discrete flow matching with masking defines a valid probability path on binary states
Basis for Equations 13 to 15; cited from Gat et al. 2024 and Campbell et al. 2024.
domain assumption Global item frequencies f_i are a sufficient statistic for a behavior-guided prior
Section 3.2 hypothesizes that a prior close to the true user behavior distribution improves training; the paper tests this only indirectly through ablations.
domain assumption Optimizing the vector field of expectation E[X_t] is sufficient to learn a useful discrete flow
Section 3.3 derives u_t from E[X_t] and trains with the simplified loss Equation 20; no proof is given that this yields a correct probability path for the discrete samples.
ad hoc to paper The inference procedure in Algorithm 2 transfers from masked training inputs to the observed interaction vector
Section 3.5 starts from the observed X at t=(N-2)/N and applies the argmax update; this distribution shift is untested in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Flow Matching for Collaborative Filtering." pith.science (2026). https://pith.science/paper/VCC5TNXC

@misc{pith2026250207303,
  author       = {Pith},
  title        = {Pith review of: Flow Matching for Collaborative Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VCC5TNXC}},
  note         = {Machine review of arXiv:2502.07303}
}

read the original abstract

Generative models have shown great promise in collaborative filtering by capturing the underlying distribution of user interests and preferences. However, existing approaches struggle with inaccurate posterior approximations and misalignment with the discrete nature of recommendation data, limiting their expressiveness and real-world performance. To address these limitations, we propose FlowCF, a novel flow-based recommendation system leveraging flow matching for collaborative filtering. We tailor flow matching to the unique challenges in recommendation through two key innovations: (1) a behavior-guided prior that aligns with user behavior patterns to handle the sparse and heterogeneous user-item interactions, and (2) a discrete flow framework to preserve the binary nature of implicit feedback while maintaining the benefits of flow matching, such as stable training and efficient inference. Extensive experiments demonstrate that FlowCF achieves state-of-the-art recommendation accuracy across various datasets with the fastest inference speed, making it a compelling approach for real-world recommender systems. The code is available at https://github.com/chengkai-liu/FlowCF.

Figures

Figures reproduced from arXiv: 2502.07303 by the authors.

**Figure 2.** Overview of FlowCF. (a) A set of interpolations [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Performance of noisy training with different pro [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Performance comparison under natural noise set [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]

**Figure 5.** Figure 5: Effect of the number of discretization steps [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Learning curves and running time per epoch com [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 24 canonical work pages

[1]

Michael S Albergo and Eric Vanden-Eijnden. 2022. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571 (2022)

arXiv 2022
[2]

Yoshua Bengio, Li Yao, Guillaume Alain, and Pascal Vincent. 2013. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems 26 (2013)

work page 2013
[3]

Avishek Joey Bose, Tara Akhound-Sadegh, Guillaume Huguet, Kilian Fatras, Jarrid Rector-Brooks, Cheng-Hao Liu, Andrei Cristian Nica, Maksym Korablyov, Michael Bronstein, and Alexander Tong. 2023. Se (3)-stochastic flow matching for protein backbone generation. arXiv preprint arXiv:2310.02391 (2023)

arXiv 2023
[4]

Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. 2024. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. arXiv preprint arXiv:2402.04997 (2024)

arXiv 2024
[5]

Hao Chen, Zefan Wang, Feiran Huang, Xiao Huang, Yue Xu, Yishi Lin, Peng He, and Zhoujun Li. 2022. Generative adversarial framework for cold-start item recommendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 2565–2571

work page 2022
[6]

Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. 2018. Neural ordinary differential equations. Advances in neural information processing systems 31 (2018)

work page 2018
[7]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. [n. d.]. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URL https://arxiv. org/abs/2403.03206 2 ([n. d.])

arXiv 2024
[8]

Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. 2024. Discrete flow matching. arXiv preprint arXiv:2407.15595 (2024)

arXiv 2024

Show all 63 references

[9]

F Maxwell Harper and Joseph A Konstan. 2015. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis) (2015)

2015
[10]

Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval ...

2020
[11]

Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web . 173–182

2017
[12]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

2020
[13]

Yu Hou, Jin-Duk Park, and Won-Yong Shin. 2024. Collaborative filtering based on diffusion models: Unveiling the potential of high-order connectivity. In Proceed- ings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1360–1369

2024
[14]

Bowen Jing, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi Jaakkola
[15]

Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

2013 arXiv
[16]

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

2014 arXiv
[17]

Yehuda Koren, Steffen Rendle, and Robert Bell. 2021. Advances in collaborative filtering. Recommender systems handbook (2021), 91–142

2021
[18]

Zihao Li, Aixin Sun, and Chenliang Li. 2023. Diffurec: A diffusion model for sequential recommendation. ACM Transactions on Information Systems 42, 3 (2023), 1–28

2023
[19]

Dawen Liang, Rahul G Krishnan, Matthew D Hoffman, and Tony Jebara. 2018. Variational autoencoders for collaborative filtering. In Proceedings of the 2018 world wide web conference . 689–698

2018
[20]

Jianghao Lin, Jiaqi Liu, Jiachen Zhu, Yunjia Xi, Chengkai Liu, Yangtian Zhang, Yong Yu, and Weinan Zhang. 2024. A Survey on Diffusion Models for Recom- mender Systems. arXiv preprint arXiv:2409.05033 (2024)

2024 arXiv
[21]

Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving graph collaborative filtering with neighborhood-enriched contrastive learning. In Proceedings of the ACM web conference 2022 . 2320–2329

2022
[22]

Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le
[23]

Chengkai Liu, Jianghao Lin, Hanzhou Liu, Jianling Wang, and James Caver- lee. 2024. Behavior-Dependent Linear Recurrent Units for Efficient Sequential Recommendation. arXiv preprint arXiv:2406.12580 (2024)

2024 arXiv
[24]

arXiv preprint arXiv:2210.02747 (2022)

Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

2022 arXiv
[25]

Chengkai Liu, Jianling Wang, and James Caverlee. 2024. TwinCL: A Twin Graph Contrastive Learning Model for Collaborative Filtering. arXiv preprint arXiv:2409.19169 (2024). KDD ’25, August 3–7, 2025, Toronto, ON, Canada Chengkai Liu, Yangtian Zhang, Jianling Wang, Rex Ying, and...

2024 arXiv
[26]

Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee
[27]

Youzhi Luo, Chengkai Liu, and Shuiwang Ji. 2023. Towards symmetry-aware generation of periodic materials. Advances in Neural Information Processing Systems 36 (2023), 53308–53329

2023
[28]

Jianxin Ma, Chang Zhou, Peng Cui, Hongxia Yang, and Wenwu Zhu. 2019. Learn- ing disentangled representations for recommendation. Advances in neural infor- mation processing systems 32 (2019)

2019
[29]

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

2022 arXiv
[30]

George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. 2021. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research 22, 57 (2021), 1–64

2021
[31]

Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. 2024. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720

2024 arXiv
[32]

Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel
[33]

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 . 10684–10695

2022
[34]

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

2007
[35]

Ilya Shenbin, Anton Alekseev, Elena Tutubalina, Valentin Malykh, and Sergey I Nikolenko. 2020. Recvae: A new variational autoencoder for top-n recommenda- tions with implicit feedback. In Proceedings of the 13th international conference on web search and data mining . 528–536

2020
[36]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme
[37]

Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems

2019
[38]

Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

2020 arXiv
[39]

Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. 2023. Improving and gener- alizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482 (2023)

2023 arXiv
[40]

Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol
[41]

Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli
[42]

In International conference on machine learning

Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning . PMLR, 2256–2265
[43]

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

2019
[44]

Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised graph learning for recommendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval . 726–735

2021
[45]

Yao Wu, Christopher DuBois, Alice X Zheng, and Martin Ester. 2016. Collabora- tive denoising auto-encoders for top-n recommender systems. In Proceedings of the ninth ACM international conference on web search and data mining . 153–162

2016
[46]

Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola
[47]

Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. 2022. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923 (2022)

2022 arXiv
[48]

Wenjie Wang, Fuli Feng, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2021. Denoising implicit feedback for recommendation. In Proceedings of the 14th ACM international conference on web search and data mining . 373–381

2021
[49]

Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua
[50]

Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. 2022. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and development in info...

2022
[51]

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

2024
[52]

Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, et al. 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. In proceedings of the 30th acm international...

2021
[53]

Yunqin Zhu, Chao Wang, Qi Zhang, and Hui Xiong. 2024. Graph signal diffusion model for collaborative filtering. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . A Appendix Algorithm 3 Inference with CFlowCF In...

2024
[57]

Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiang- nan He. 2024. Generate what you prefer: Reshaping sequential recommendation via guided diffusion. Advances in Neural Information Processing Systems

2024
[58]

Zixuan Yi, Xi Wang, and Iadh Ounis. 2024. A Directional Diffusion Graph Transformer for Recommendation. arXiv preprint arXiv:2404.03326 (2024)

2024 arXiv
[600]

The initial dropout rate of the MLP in FlowCF is fixed at 0, while for other generative CF baseline models, it is tuned within [0, 0.5]

for Amazon-Beauty. The initial dropout rate of the MLP in FlowCF is fixed at 0, while for other generative CF baseline models, it is tuned within [0, 0.5]. We run and tune our baselines using the widely used recommendation library RecBole [52] and report the best results to en...
[2008]

In Proceedings of the 25th international conference on Machine learning

Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning . 1096–1103
[2012]

arXiv preprint arXiv:1205.2618 (2012)

BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618 (2012)

2012 arXiv
[2015]

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

Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52
[2021]

arXiv preprint arXiv:2110.06197 (2021)

Crystal diffusion variational autoencoder for periodic material generation. arXiv preprint arXiv:2110.06197 (2021)

2021 arXiv
[2022]

Torsional diffusion for molecular conformer generation.Advances in Neural Information Processing Systems 35 (2022), 24240–24253

2022
[2023]

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

Diffusion recommender model. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval . 832–841
[2024]

arXiv preprint arXiv:2403.03900 (2024)

Mamba4rec: Towards efficient sequential recommendation with selective state space models. arXiv preprint arXiv:2403.03900 (2024)

2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Michael S Albergo and Eric Vanden-Eijnden. 2022. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571 (2022)

arXiv 2022

[2] [2]

Yoshua Bengio, Li Yao, Guillaume Alain, and Pascal Vincent. 2013. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems 26 (2013)

work page 2013

[3] [3]

Avishek Joey Bose, Tara Akhound-Sadegh, Guillaume Huguet, Kilian Fatras, Jarrid Rector-Brooks, Cheng-Hao Liu, Andrei Cristian Nica, Maksym Korablyov, Michael Bronstein, and Alexander Tong. 2023. Se (3)-stochastic flow matching for protein backbone generation. arXiv preprint arXiv:2310.02391 (2023)

arXiv 2023

[4] [4]

Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. 2024. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. arXiv preprint arXiv:2402.04997 (2024)

arXiv 2024

[5] [5]

Hao Chen, Zefan Wang, Feiran Huang, Xiao Huang, Yue Xu, Yishi Lin, Peng He, and Zhoujun Li. 2022. Generative adversarial framework for cold-start item recommendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 2565–2571

work page 2022

[6] [6]

Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. 2018. Neural ordinary differential equations. Advances in neural information processing systems 31 (2018)

work page 2018

[7] [7]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. [n. d.]. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URL https://arxiv. org/abs/2403.03206 2 ([n. d.])

arXiv 2024

[8] [8]

Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. 2024. Discrete flow matching. arXiv preprint arXiv:2407.15595 (2024)

arXiv 2024