acceptodds
Under review as a conference paper at ICLR 2027

OT-FlowRec: Optimal Transport-Guided Multi-Bridge Flow for Multimodal Sequential Recommendation

Abstract

E-commerce user preferences evolve dynamically within sessions, exhibiting concurrent and multi-centric interest distributions. While existing multimodal sequential recommendation methods leverage global attention to compress historical interactions, they inadequately capture the concurrent nature of multi-interest behaviors and fail to model the underlying multi-centric distribution of user intent. To tackle this issue, we propose OT-FlowRec, an optimal transport-guided multi-bridge flow matching framework that explicitly models multi-interest distributions to facilitate multimodal sequential recommendation. Unlike single-path generative models, OT-FlowRec constructs a mixture of parallel, interest-specific transport flows to model the multi-centric distribution characteristics, where each flow preserves a distinct user interest cluster via an optimal transport plan. Then, a geometric transport mechanism is designed to establish minimum-cost paths from the latent multi-interest distribution to target preference prototypes to provide multi-interest awareness. Moreover, a progressive time-sampling warmup strategy dynamically biases optimization weights toward the final states of the generation process, bridging the gap between continuous global distribution fitting and practical top- ranking metrics. This approach achieves explicit interest disentanglement, ensuring the preservation of a multi-peaked preference representation instead of collapsing it into a unimodal average. Extensive experiments on three public datasets demonstrate that OT-FlowRec outperforms existing state-of-the-art baselines.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multimodal Sequential Recommendation, Flow Matching, Optimal Transport

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_R89zOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Existing sequential recommendation methods typically compress user history into a single representation, failing to account for the multi-interest nature and multi-centric distribution of user preferences. The paper proposes OT-FlowRec, which (i) constructs a mixture of interest-specific transport flows, each representing a distinct user interest cluster; (ii) designs a geometric transport mechanism to align latent interest distributions with target preference prototypes; and (iii) integrates a progressive time-sampling warmup strategy to bridge the gap between continuous global distributions and ranking metrics. Experimental results show that the proposed OT-FlowRec achieves clearly performance gains over baselines on benchmark datasets.

Critical strengths
  1. The paper addresses an important and practical problem in multimodal sequential recommendation. Since user interests are inherently multi-centric, modeling this characteristic can better capture user preferences and improve recommendation performance.

  2. The proposed OT-FlowRec model has a sound architecture and reasonable component design. The paper is also well-written and easy to understand.

  3. OT-FlowRec consistently outperforms the baselines across the selected datasets, demonstrating its effectiveness.

Critical weaknesses
  1. Prototype-based representation learning for disentangling multiple user interests for recommendation has been extensively studied in prior work [a–f]. However, the paper provides insufficient evidence to distinguish its approach from existing methods, limiting its novelty.

  2. Even if the novelty is considered from the optimal transport perspective, related work [g] has already explored this approach. The paper should discuss the differences and better highlight its key contributions.

  3. The comparisons primarily involve single-interest modeling baselines, which are insufficient to convincingly demonstrate the proposed method's advantages. The improvements may largely stem from multi-interest modeling, whose effectiveness has already been established in prior work. Furthermore, despite leveraging multimodal content for interest modeling, the paper lacks comparisons with stronger multimodal sequential recommendation baselines, such as MISSRec [h]. Table 3 shows that each proposed component contributes only marginally to overall performance, suggesting that the main gains may come from multi-interest modeling.

  4. The paper lacks an evaluation of interest disentanglement, leaving it unclear whether multiple user interests are effectively disentangled.

  5. The paper does not discuss the computational costs of optimal transport and multi-interest modeling, making it difficult to assess the trade-off between effectiveness and efficiency.

  6. The paper lacks discussion and comparisons with closely related VAE-based approaches [d, g, i, j].

[a] Controllable Multiinterest Framework for Recommendation. KDD 2020.

[b] Re4: Learning to Re-Contrast, Re-Attend, Re-Construct for Multi-Interest Recommendation. TheWebConf (WWW) 2022.

[c] User-Aware Multi-Interest Learning for Candidate Matching in Recommenders. SIGIR 2022.

[d] Aligning Dual Disentangled User Representations from Ratings and Textual Content. KDD 2022.

[e] Disentangled Multi-Interest Representation Learning for Sequential Recommendation. KDD 2024.

[f] Parameter-Efficient Variational AutoEncoder for Multimodal Multi-Interest Recommendation. ACM MM 2025.

[g] Optimal Transport Alignment of User Preferences from Ratings and Texts. UAI 2025.

[h] MISSRec: Pre-training and Transferring Multi-modal Interest-aware Sequence Representation for Recommendation. ACM MM 2022.

[i] TopicVAE: Topic-Aware Disentanglement Representation Learning for Enhanced Recommendation. ACM MM 2022.

[j] Align MacridVAE: Multimodal Alignment for Disentangled Recommendations. ECIR 2024.

Questions for authors

Please see the review.

4 comments

As the concerns raised in the review have not been addressed in the rebuttal phase, I would like to maintain my original score, as it continues to reflect my assessment of the paper.

Thanks for your insightful and constructive comments. We responded them point-by-point and updated the relevant content in the manuscript.

1. Distinction from prototype-based and multi-interest methods

Thank you for the constructive suggestion. We distinguished between OT-FlowRec and seven prototype-based methods you mentioned in the table below. In our model, OT-FlowRec used parallel interest transport flows to prevent interest distribution collapse. The OT-MBFM module constructs minimum-cost transport paths toward preference prototypes, while PTW emphasizes terminal generative states to align flow learning with top-kk ranking.

Following your suggestion, we have added a comparison result with prototype-based methods you mentioned in following table(full results on page 8 in table 2 and page 17 in table 6). Our method achieved the best performance on all four metrics across three datasets.

MethodHow the baseline worksOT-FlowRec (ours)Main difference
ComiRecUsed dynamic routing to extract multiple interest vectors from past interactions. It retrieved items for each vector and merged the candidates.Updated the current sequence state separately for each selected prototype and scored items with the resulting vectors.ComiRec used extracted interest vectors directly for retrieval. OT-FlowRec used prototypes to guide how the user state is updated before scoring.
Re4Used re-contrast to distinguish interests, re-attention to align attention weights with matching scores, and reconstruction to recover representative historical items for each interest.Used OT to define targets for state updates and trained the prediction network with the ranking loss for the observed next item.Re4’s auxiliary tasks refined the interest vectors extracted from the history. OT-FlowRec’s flow-matching loss supervised the generation of prediction vectors.
ADDVAELearned interest factors separately from user interactions and related text, then aligned the two sets through attention so that text supplements the interaction signal.Encoded interactions in chronological order and used the current user state to select the prototypes that guide next-item prediction.ADDVAE learned user preferences from a set of interactions and related text. OT-FlowRec modeled next-item preferences based on the ordered interaction history..
TopicVAELearned topics and their associations with items from item content, decomposed user interactions by topic, and combined the predictions across topics.Selected relevant prototypes based on the current sequence and generated one prediction vector conditioned on each selected prototype.TopicVAE used topics to decompose observed interactions. OT-FlowRec used prototypes to guide predictions along distinct interest directions.
Align MacridVAEAligned image and text representations of the same item, groups items into categories in a shared space, and used a VAE to learn user preferences within each category.Fused item ID, image, and text features before sequence encoding. The current user state determined which prototypes guide prediction.The prototypes defined semantic item categories in Align MacridVAE and distinct prediction directions for a user’s current sequence in OT-FlowRec.
BANDVAEUsed OT to align interest factors from ratings and text, then combined the preferences from both sources through bidirectional barycentric mapping.Used OT to compute soft matches between recent interaction representations and interest prototypes. These matched define the flow-matching targets.BANDVAE’s transport plan combined preferences from two sources. OT-FlowRec’s transport plan provided supervision for the state updated used in prediction.
MISSRecClusters image and text features to obtain interest prototypes. It learned transferable sequence representations through interest-aware encoding and decoding and contrastive pretraining. At recommendation time, it scored candidates with the vector from the final decoder position.Retained the prediction vectors from the selected interest branches and assigned each candidate its highest matching score across branches.MISSRec scored all candidates with one sequence vector. In OT-FlowRec, each candidate received its score from the branch that best matched it.

2. Distinction from BANDVAE’s use of OT

Thank you for the constructive suggestion. OT-FlowRec and BANDVAE (related work g) differed in two points, (i) the learning mechanism, i.e. how they used OT during training; (ii) Input modality differences.

(i) Learning mechanism. In OT-FlowRec, we used flow-matching targets from the OT coupling between the source and target distributions. These targets directly supervised the direction of the state updates that generated next-item prediction vectors. In contrast,BANDVAE used OT to align interest factors learned from ratings and text, then combined preference information from the two sources through bidirectional barycentric mapping.

(ii) Input modality differences. OT-FlowRec input item IDs, images, and text. Item ID embeddings captured collaborative information. Images provided visual cues such as color and shape, and text describes product attributes such as specifications. The sequence encoder incorporated these complementary modality for each item. In contrast,BANDVAE only used ratings and text as the same modality, failing to fully leverage multimodal information.

3. Stronger baselines and the sources of improvement

Thank you for your constructive suggestion. Following your suggestion, we added ComiRec(KDD 2020) and Re4(WWW 2022) as multi-interest baselines and MISSRec(MM 2023) as a stronger multimodal sequential recommendation baseline. OT-FlowRec outperformed all three baselines in the table below (full results on page 8 in table 8). Its relatived NDCG@20 gains over MISSRec were 15.98%, 7.44%, and 2.05% across the three datasets.

We also evaluated contributions of the individual components. We removed OT-MBFM and PTW one at a time from the single-interest variant of OT-FlowRec. According to result,removing either component lowered every metric on all datasets (page19 table 9). These results indicate that each component also provided a certain performance gain in single-interest modeling. Hence, all components contribute to the performance gains, rather than the improvements arising from multi-interest modeling.

BaselineBeauty
H@10H@20N@10N@20
ComiRec0.07600.10570.04280.0503
Re40.05570.07920.03200.0379
MISSRec0.07580.11090.03760.0463
ours0.08910.12670.04420.0537
VariantBeauty
H@10H@20N@10N@20
Single-interest0.08800.12620.04370.0532
Single-interest & w/o OT-MBFM0.08330.12160.04140.0510
Single-interest & w/o PTW0.08650.12430.04280.0523
OT-FlowRec0.08910.12670.04420.0537

4. Interest Disentanglement Evaluation

Thank you for raising this point. To evaluate the interest disentanglement, we used Jensen–Shannon (JS) divergence[1] to evaluate the level of interest disentanglement. Higher JS divergence indicates greater differences between the interest distributions. We computed the average pairwise JS divergence between different interest prototypes and showed results in below table. The experimental results indicate that our method got a well separation among the interest representations.

[1] Si Z, Sun Z, Zhang X, et al. When search meets recommendation: Learning disentangled search representation for recommendation[C]//Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 2023: 1313-1323.

DatasetJS Divergence ↑
Beauty0.5877
Toys0.6423

5. Efficiency Analysis

Thank you for the suggestion. We added a complexity analysis of OT and multi-interest modeling and measurements of inference time. Based on our analysis, the time complexity are O(LM(d+J))\mathcal{O}(LM(d+J)) for OT and O(Md+k(Sd2+∣I∣d))\mathcal{O}\left(Md+k(Sd^2+|\mathcal{I}|d)\right) for multi-interest modeling. Here, LL is the maximum history length, MM is the number of interest prototypes, dd is the embedding dimension, and JJ is the number of Sinkhorn iterations. kk is the number of bridges, SS is the number of Euler steps per bridge, and I\mathcal{I} is the item set.

We ran inference on an RTX 3090 and the inference time over five runs. The table below shows sizmilar inference times across the component ablations. Together with the recommendation results, these measurements showed that OT-FlowRec balances ranking performance and inference cost.

DatasetMethodFull-test time (ms)
ToysFull385.6
w/o MPSD371.7
w/o OT-MBFM375.4
w/o PTW372.7

6. Comparison with related VAE methods

Thank you for your constructive suggestion. Following your suggestion, we added a discussion of VAE baselines and compared them with OT-FlowRec.

ADDVAE, TopicVAE, and Align MacridVAE learned user preferences from sets of interactions and used content in different ways. ADDVAE used attention to align interest factors learned from interactions with those learned from text, so text could complement interaction signals. TopicVAE learned topics from item content, groups user interactions by topic, and aggregated the topic-level predictions. Align MacridVAE first aligned the image and text representations of each item, then groups items into semantic categories and learned user preferences within each category. OT-FlowRec preserved interaction order and selects interest prototypes based on the current sequence state. Different orderings of the same items can therefore lead to different prototype selections and next-item predictions. OT-FlowRec used OT coupling to construct flow-matching targets that supervise state updates for next-item prediction, whereas BANDVAE used OT to align rating- and text-based interests through bidirectional barycentric mapping. For input modalities, OT-FlowRec integrated item IDs, images, and text, while BANDVAE only used ratings and text, without fully leveraging multimodal information.

As shown in the table below, OT-FlowRec outperformed all four VAE baselines on all metrics across the three datasets(page 17 table 6). Relative to BANDVAE, the strongest VAE baseline on NDCG@20, OT-FlowRec improved that metric by 14.74% on Beauty, 21.31% on Video Games, and 10.48% on Toys.

BaselineBeauty
H@10H@20N@10N@20
ADDVAE0.06310.09330.03300.0407
TopicVAE0.06340.10030.03280.0420
Align MacridVAE0.05570.08550.02870.0362
BANDVAE0.07090.10200.03900.0468
ours0.08910.12670.04420.0537
Reviewer_gSeFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes OT-FlowRec, an optimal transport-guided multi-bridge flow matching framework for multimodal sequential recommendation. It targets two limitations of prior work: the collapse of multi-centric interest distributions into single deterministic embeddings, and the lack of directional constraint in transport trajectories due to independently sampled priors. MPSD selects top-k interest prototypes to preserve multi-peaked user preferences and initialize parallel interest-specific bridges; OT-MBFM computes Sinkhorn optimal couplings to establish minimum-cost paths from the latent multi-interest distribution to target preference prototypes; and PTW shifts Beta-distributed time-step sampling toward terminal states to align continuous distribution fitting with discrete top-k ranking.

Critical strengths
  1. The paper is well-motivated, and the motivation is critical for MMSR. It highlights the limitations of existing unimodal modeling methods and designs OT-FlowRec to address them.
  2. Each module clearly corresponds to the problem, and the overall framework is internally consistent.
  3. The experiments are relatively comprehensive. The main results and the ablation study support the core idea, while the t-SNE visualization further provides evidence that MPSD preserves multi-peaked interest structure.
Critical weaknesses
  1. Some SR baselines with intent prototypes are missing, such as ICLRec, MISSRec, and ELCRec. If the authors could add such baseline comparisons, the main experiment will be more convincing.
  2. The datasets for evaluation are all from Amazon-Review. Therefore, the generality of the model effect cannot be fully proved. Adding a dataset from other sources is encouraged.
  3. The ablation study only discusses the significance of components. The effects of the loss functions in Eq.8 are not fully analyzed, especially for LprotoL_{proto} and LOTL_{OT}.
Questions for authors

Please see the weaknesses above.

1 comment

Thanks for your insightful and constructive comments. We responded them point-by-point and updated the relevant content in the manuscript.

1. Additional intent prototype baselines

Thank you for your constructive suggestion. Following your suggestion, we compared three sequential recommendation baselines that use intent prototypes ICLRec(WWW 2022), MISSRec(MM 2023), and ELCRec(NeurIPS 2024). OT-FlowRec achieved the best performance among these methods on three datasets (complete result on page 8 in table 2). These results demonstrated its performance advantage over the prototype-based baselines on the evaluated datasets.

BaselineBeautyVideo Games
H@10H@20N@10N@20H@10H@20N@10N@20
ICLRec0.07300.10490.04080.04880.11170.16820.05930.0735
MISSRec0.07580.11090.03760.04630.11940.18200.05720.0726
ELCRec0.07320.10240.04200.04930.11490.17250.06050.0750
ours0.08910.12670.04420.05370.12760.19640.06070.0780

2. Additional experiments on ML-1M and Yelp

Thank you for your constructive suggestion. Following your suggestion, we evaluated our method on two additional datasets from different domains (ML-1M for movie recommendation and Yelp for local business recommendation). We reported the dataset statistics and experimental results below (complete result on page 18 in table 8). Our method performed best on both datasets, providing further evidence of its applicability beyond Amazon.

DatasetUsersItemsInter.SparsityAvg. Len.
ML-1M6,0403,417999,61195.16%165.5
Yelp277,629112,3934,250,45799.99%15.3
BaselineML-1MYelp
H@10H@20N@10N@20H@10H@20N@10N@20
ADRec0.20780.29370.11750.13910.06150.09850.03120.0405
FlowCF0.10500.16850.05080.06670.04600.07270.02390.0306
FindRec0.25580.36290.14020.16720.05690.09090.02920.0378
MISSRec0.24230.34680.12090.14650.06370.09870.03140.0400
ours0.27430.37810.15050.17660.07440.11250.04150.0510

3. Additional loss ablations

Thank you for your constructive suggestion. Following your suggestion, we added an ablation study of the three loss terms in Eq. (8). The results below showed that all these three terms contribute to the recommendation task. Specifically, LProtoL_{\mathrm{Proto}} helped the model distinguish among user interests, and removing it reduced Hit@10 on Beauty by 3.48%. LOTL_{\mathrm{OT}} encouraged alignment between historical preferences and interest prototypes. Without this term, NDCG@10 on Video Games dropped by 10.21%. Removing LFML_{\mathrm{FM}}, which helped the model learn how preferences evolve, reduced NDCG@10 on Toys by 5.71%.

VariantBeautyVideo Games
H@10H@20N@10N@20H@10H@20N@10N@20
w/o LProtoL_{\mathrm{Proto}}0.08600.12420.04300.05260.12580.19480.05960.0770
w/o LOTL_{\mathrm{OT}}0.08690.12450.04410.05360.12200.18820.05450.0711
w/o LFML_{\mathrm{FM}}0.08800.12630.04340.05330.12600.19200.06060.0772
ours0.08910.12670.04420.05370.12760.19640.06070.0780
Reviewer_fHicOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

OT-FlowRec combines multimodal feature fusion with a bank of learned interest prototypes and prototype-conditioned flow-based representations. An entropic optimal-transport plan connects history-derived source representations with target prototypes. Training combines flow matching, next-item cross-entropy, transport regularization, and prototype orthogonality, while a progressive Beta time-sampling schedule emphasizes later trajectory states. At inference, several prototype-conditioned ODE trajectories start from the user state and their item scores are max-pooled. Experiments on three Amazon datasets report gains over sequential, diffusion, and flow-based baselines and ablate the three main modules.

Critical strengths
  1. Explicitly representing several interests is a relevant modeling direction. Using multiple prototype-conditioned terminal states rather than a single averaged representation provides an interpretable inductive bias for concurrent preferences.
  2. The paper attempts to separate the roles of its major components. The comparisons span several model families, and the ablations separately remove multiple prototypes, OT guidance, and time-sampling warmup. These are useful ingredients for an empirical evaluation, provided that the metric and implementation issues below are resolved.
Critical weaknesses
  1. Several main-table metrics are inconsistent with the stated single-next-item ranking task. For a common set of evaluation events with one relevant next item, let
\Delta N=N@20-N@10.$$ Only hits at ranks 11–20 contribute to these differences, so necessarily $$\frac{\Delta H}{\log_2 21} \leq \Delta N \leq \frac{\Delta H}{\log_2 12}.$$ In Table 2, PreferDiff on Beauty has $$\Delta H=0.0675-0.0594=0.0081,$$ requiring $\Delta N$ to lie approximately in [0.00184, 0.00226]. The reported difference is instead $$0.0471-0.0350=\mathbf{0.0121}.$$ ADRec on Toys similarly has $\Delta H=0.0205$ but $\Delta N=0.0091$, above the allowable upper bound of about 0.00572. These discrepancies are much larger than rounding error, and averaging over seeds preserves the inequality. Unless the metrics use different, currently undocumented protocols, the entries cannot simultaneously be correct. This prevents relying on the main empirical comparison; it is not an allegation about the cause of the discrepancy. 2. The training objective, transport marginals, and inference procedure do not yet specify a consistent conditional-flow model. Equation (9) trains on source–prototype pairs drawn from the history-based OT plan, while Eq. (13) initializes every inference branch at the summarized state $h_u$ and explicitly conditions the field on $p_k$; the corresponding training conditioning and source-state correspondence are not specified. Section 2.4 defines transport to the selected top-k prototypes, but Appendix A defines the target distribution over all $M$ prototypes and mixes an $\ell\times M$ coupling with a length-$|K_p|$ marginal constraint. If only top-k masses are retained, they must be renormalized to form a feasible balanced coupling with a unit-mass source. These are central algorithmic choices, and the available ablations do not resolve which model was actually optimized.
Questions for authors
  1. Evaluation clarification: Can you reconcile the Hit/NDCG contradictions in Table 2 and provide the exact evaluation protocol, metric implementation, and per-seed results computed from the same ranked outputs? Do the conclusions survive after correcting the affected entries?

  2. Method clarification: Please provide one consistent training/inference pseudocode specifying the source states, prototype conditioning, top-k target-mass normalization, and the definition of the terminal states used by cross-entropy. How is the training flow-matching distribution connected to the huh_u initialization used at inference?

1 comment

Thanks for your insightful and constructive comments. We responded them point-by-point and updated the relevant content in the manuscript.

1. Evaluation clarification

Thank you for valuable suggestion. The issue comes from differences in baseline evaluation protocols. We reviewed the evaluation protocols for all baselines and found that some differed from ours. In particular, the PreferDiff scores in the original Table 2 came from a per-user 8:1:1 split rather than leave-one-out evaluation, hence fell outside the bounds relating Hit and NDCG.

Following your suggestion, we conducted additional experiments and re-evaluated partial baseline methods using leave-one-out evaluation. For each user, we held out the last interaction as the test target and the second-to-last interaction as the validation target. We used the earlier interactions for training and selected model checkpoints based solely on validation performance.

The updated results have reported on page 8 in table 2. According to updated results, our method still achieved best performance on all four metrics across three datasets, which supporting the paper’s main conclusions.

Following to your suggested calculation method, we calculated ΔH and ΔN ranges, and all metrics are satisfied. The table below reports the metric consistency check for Video Games under this protocol. The equations that follow show how we computed Hit and NDCG.

Let rur_u denote the held-out item's rank for test user uu. All four metrics are computed from the same ranking outputs:

\mathrm{NDCG@}K=\frac{1}{|\mathcal U|}\sum_{u\in\mathcal U} \frac{\mathbf 1[r_u\le K]}{\log_2(r_u+1)},\quad K\in\{10,20\}.$$ | Baseline | ΔH | Reported ΔN | Allowed ΔN Range | Consistent | | ---------- | ------ | ----------- | ---------------- | ---------- | | ADRec | 0.0483 | 0.0121 | [0.0110, 0.0135] | ✓ | | PreferDiff | 0.0126 | 0.0033 | [0.0029, 0.0035] | ✓ | | FlowCF | 0.0388 | 0.0098 | [0.0088, 0.0108] | ✓ | **2. Method clarification** Thank you for valuable suggestion. We checked the manuscript against the implementation and confirmed that OT actually couples the $L$ historical states to all $M$ prototypes during training. Specifically, Top-$K$ selection is used only to choose the recommendation branches. To make the algorithm clearer, we revised Section 2.4, Eq. (9), and Appendix A. The changes are marked in blue in the manuscript. The explanation and pseudocode for training and inference follow. **About the corresponding training conditioning and source-state correspondence .** The source states for flow matching are the last $\ell$ encoder states, $x_{n-\ell+1},\ldots,x_n$, with normalized recency masses $\mathbf a$. Each recommendation branch starts at $x_0^{(k)}=h_u$. Prototype conditioning uses $q_u=\sum_{m=1}^{M}s_mp_m$ for flow matching and the selected $p_k$ for each recommendation branch. For top-k target-mass normalization, Sinkhorn uses the full softmax $\mathbf s$ over all $M$ prototypes, with total mass one. The terminal states used by cross-entropy are $x_1^{(k)}$, the $t=1$ outputs of the Euler integration procedure used in both training and inference (Eq. 13). Cross-entropy uses the max-pooled item logits computed from these states. **The relation between training flow-matching distribution and inference $h_u$ initialization**. Flow matching samples a source state $x_{n-\ell+j}$ according to $\mathbf a$ and learns the velocity $T_j-x_{n-\ell+j}$ along the interpolation to its OT barycenter $T_j$. The same encoder states yield $h_u=\tfrac12 x_n+\tfrac1{2n}\sum_{i=1}^{n}x_i$, which determines the prototype responses $\mathbf s$ and initializes the recommendation ODE. Flow matching provides velocity supervision to the shared vector field, and cross-entropy trains it directly through the $h_u$-initialized ODE used at inference. **Algorithm 1: Training and inference of OT-FlowRec** 1. Input: History $S_u$, mode, prototype bank $P$, and next item $i_{n+1}$. 2. Encode $S_u$ to obtain $x_1,\ldots,x_n$ and $h_u$. 3. Compute prototype responses $\mathbf{s}$ and select $K_p$ using Eqs. (3) and (4). 4. if mode is training then 5. Build $C\in\mathbb{R}^{\ell\times M}$ and normalized recency weights $\mathbf{a}$, with $\ell=\min(n,L)$. 6. Solve $\pi^*=\mathrm{Sinkhorn}(C,\mathbf{a},\mathbf{s})$. 7. Sample $j$ using $\mathbf{a}$ and $t\sim\mathrm{Beta}(\alpha,1)$; compute $\mathcal{L}_{\mathrm{FM}}$ by Eq. (9). 8. end if 9. Obtain $x_1^{(k)}$ by integrating Eq. (13) from $h_u$ with condition $p_k$, $k\in K_p$. 10. Compute $s_j=\max_{k\in K_p}\langle x_1^{(k)},e_j\rangle$. 11. if mode is inference then return $\{s_j:j\in\mathcal{I}\}$. 12. Compute $\mathcal{L}_{\mathrm{CE}}$, $\mathcal{L}_{\mathrm{OT}}$, $\mathcal{L}_{\mathrm{Proto}}$ using Eqs. (10)–(12). 13. Update model parameters by backpropagation to minimize loss.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.