acceptodds
Under review as a conference paper at ICLR 2027

Pathwise Alignment for Collaborative Filtering: Bilateral Conditional Flow Regularization

Abstract

Collaborative filtering predicts user-item interactions from implicit feedback. We focus on matrix factorization (MF) trained with the Bayesian Personalized Ranking loss. This loss updates each user embedding only through that user's observed items and sampled negatives, so sparse feedback gives MF few constraints. To constrain each user through other users, we propose Pathwise Alignment for Collaborative Filtering (PACF), which trains MF with one network shared by all training pairs. At points between user and item embeddings, a flow matching term trains the network to predict the user-to-item step, reversed and scaled for sampled negatives. A trajectory consistency term pulls each user's predicted path to end at the observed item. Both terms update the embeddings, so the network is discarded after training, thereby maintaining the efficiency of MF at inference. We prove that a network with bounded sensitivity to its input point accurately fits users who share an item only if their embeddings are close, so these users constrain one another. Against eleven discriminative and generative baselines, PACF achieves the highest performance on all seven datasets. Its relative gains over MF are largest where feedback is sparsest. The shared network thus supplies the missing constraints through other users.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Recommender Systems, Collaborative Filtering, Flow Matching, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZXkiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PACF, a training-time regularizer for matrix factorization under sparse implicit feedback. It combines BPR with a shared network that predicts signed user–item displacements and a trajectory-consistency loss that encourages positive trajectories to reach their items. Both auxiliary losses update the embeddings; the network is discarded at inference, preserving static inner-product scoring.

Critical strengths

S1. Important problem and practical design. PACF addresses limited supervision in sparse recommendation without complicating the deployed scorer. Its auxiliary network operates only during training, preserving MF’s inference complexity at the same embedding dimension.

S2. Interpretable theoretical contribution. Proposition 2 links flow-regression error to the dispersion of users sharing an item and label under bounded network sensitivity. This provides a concrete explanation of how shared regression can couple users, beyond simply reporting that an auxiliary loss improves accuracy.

Critical weaknesses

W1. The theoretical mechanism is not sufficiently validated in practice. The central bound becomes weaker as network sensitivity increases, but the paper does not establish a quantitatively useful sensitivity bound for its trained networks. Moreover, Proposition 2 provides a necessary condition for low regression error—not a guarantee of better ranking. The connection between the theory and observed improvements therefore remains incomplete.

W2. Novelty and necessity relative to simpler regularizers remain unclear. Closely related approaches—including DirectAU, UltraGCN, and graph-regularized MF—are discussed but not evaluated. Without these comparisons, the experiments do not establish whether pathwise regression offers a distinctive advantage over simpler embedding constraints.

W3. Experimental controls weaken attribution and confidence in some gains. Some endpoint and sampling controls receive only 150 tuning trials and report single-seed results, whereas PACF uses its 500-trial, five-seed main result. These differences confound component-level conclusions. Additionally, several improvements over the strongest baseline are small relative to seed variability, so paired statistical analysis is needed to establish reliable superiority. 40710_Pathwise_Alignment_for_C(…

Questions for authors

Please refer to the weaknesses.

4 comments

We are grateful for your insightful comments.

The manuscript has been revised accordingly. Each heading below notes the updated sections and tables.

W1. Quantitative probes of sensitivity, dispersion, and updates (Sec. 5.1, App. C.3, Figs. 3, 9–10, Tables 8, 9, 24)

Proposition 2 is a residual lower bound under a shared predictor and a sensitivity condition. On trained models we therefore measure the local signatures of that condition directly.

Eq. (37) turns flow residuals into user updates:

∇pu∥eu(t)∥22=2[(1−t)Jh+μ(y)I]⊤eu(t).\nabla _ {p _ u}\lVert e _ u(t) \rVert _ 2^2 = 2[(1-t)J _ h + \mu(y)I]^\top e _ u(t).

At t=0.5t=0.5 on sampled states we measure the state-Jacobian norm ∥Jh∥2\lVert J _ h \rVert _ 2, terminal co-user dispersion VV, the Adam drift of LfL _ f relative to BPR, the cosine CfC _ f, and the dispersion derivatives DkD _ k (units 10−610^{-6}):

Dataset∥Jh∥2\lVert J _ h \rVert _ 2BPR VVPACF VVAdam DriftCfC _ fDrD _ rDfD _ fDtD _ t
ML29.102.071.170.99+0.014+10.73-9.392-0.029
Food5.870.580.581.49+0.062+0.371-0.011-0.023
Baby1.780.900.891.02-0.080+0.717-0.019-0.016

These numbers make the residual mechanism concrete. Finite ∥Jh∥2\lVert J _ h \rVert _ 2, positive target–prediction gaps in Fig. 10, and E±>0E^\pm>0 in Tables 22–24 show shared-fitting conflict at probed states. Near-zero CfC _ f and Df,Dt<0D _ f,D _ t<0 show that the auxiliary terms supply a geometric direction distinct from BPR and oriented toward lower mean relative dispersion. Fig. 3 then records lower terminal VV than BPR. Together with the matched ablations in W3, the evidence chain is therefore: conflict is observed, auxiliary gradients have a different geometry, and deleting those training paths hurts ranking.

W2. Comparisons with DirectAU, UltraGCN, SCCF, & CoUser (Sec. 6.5.1, App. C.4, Tables 2, 3, 14, 15)

The full seven-dataset HPO comparison is reported in #14vB-Q1. Here we separate the mechanisms that produce that ranking lead.

CoUser adds κRc=κ∣E∣∑i∑u∈Si∥pu−pˉi∥22\kappa R _ c=\frac{\kappa}{|E|}\sum _ i\sum _ {u\in S _ i}\lVert p _ u-\bar{p} _ i \rVert _ 2^2 to BPR-MF. By the identity ∑u∈Si∥pu−qi∥22=∑u∈Si∥pu−pˉi∥22+∣Si∣∥pˉi−qi∥22\sum _ {u\in S _ i}\lVert p _ u-q _ i \rVert _ 2^2=\sum _ {u\in S _ i}\lVert p _ u-\bar{p} _ i \rVert _ 2^2+|S _ i|\lVert \bar{p} _ i-q _ i \rVert _ 2^2 CoUser contracts users toward pˉi\bar{p} _ i and leaves pˉi\bar{p} _ i free of qiq _ i. EP$^+$ instead restores an item anchor by a fixed distance pull. PACF fits common displacements into vθv _ \theta and updates embeddings through leftover residuals. Fig. 3 and Table 3 support that geometric contrast under matched hyperparameters, while Tables 2, 6, and 7 place PACF ahead of published graph and contrastive methods under their own HPO.

W3. Ablations & reliability (Sec. 6.5.1–6.5.6, App. C.1, C.4, Tables 3–5, 14–21)

Matched ablations under PACF's selected hyperparameters identify which residual-coupling modules move ranking. #14vB-Q4 reports the main-table Yelp paired gaps versus LightCCF under the same five-seed protocol.

With the same hyperparameters, splits, initializations, and 500-epoch budgets, the matched substitutes below change the named factor. The joint control β=0\beta=0 with t∼Beta(1,5)t\sim\mathrm{Beta}(1,5) from Table 17 is included as a simultaneous path/loss stress test. The table reports five-seed NDCG@10 in percent as mean ± STD:

SettingMLFoodBaby
BPR-only2.46±0.274.10±0.071.36±0.01
CoUser2.25±0.414.12±0.051.36±0.01
EP+\text{EP}^+3.00±0.244.31±0.071.37±0.01
EP±\text{EP}^{\pm}3.01±0.244.32±0.081.37±0.02
t=0t=02.23±0.274.09±0.051.36±0.02
t=0.5t=0.52.28±0.364.11±0.051.36±0.01
Grid2.05±0.214.09±0.041.35±0.01
β=0\beta=0, t∼Beta(1,5)t\sim\mathrm{Beta}(1,5)2.14±0.274.09±0.071.36±0.01
w/o LtL _ t2.41±0.374.09±0.071.36±0.01
w/o LfL _ f2.46±0.154.14±0.071.36±0.01
TargetDetach2.68±0.294.09±0.031.36±0.01
μ−=0\mu _ {-} = 01.98±0.304.09±0.081.39±0.01
Unsigned negative target2.23±0.304.09±0.051.35±0.01
w/o negative regression2.06±0.454.09±0.051.35±0.01
PACF3.11±0.154.46±0.041.44±0.01

These matched gaps show that shared field, path support, signed targets, and both pathwise losses each move ranking under one fixed protocol. Sec. 6.5.6 further shows that the static readout aa leads aΦa _ \Phi and adista _ {\mathrm{dist}}, so the retained gain is amortized into the scorer. On the main comparison, #14vB-Q4 reports Yelp mean leadership with positive paired gaps and CIs that cover both signs under the five-seed protocol.

AuthorsOpenReview ↗

Follow-up: manuscript revisions and updated references

Thank you again for your review.

In response to Reviewer #ATCM's follow-up, we have further revised the manuscript to clarify the link between signed regression, shared residual coupling and positive endpoint consistency, as detailed in #ATCM-F1.

Some references in our earlier rebuttal have shifted. The following guide points to the latest revision:

  • W1: the flow-gradient formula previously cited as Eq. (37) is now Eq. (42). The velocity residual norms E±E^\pm, previously referenced as Tables 22–24, are specifically in Table 24. Sec. 5.1 now derives the local target-gradient correction in Eq. (16) for users sharing an item and label under the stated sensitivity condition. Figs. 9 and 10 show local sensitivity and co-user fitting gaps, Table 9 measures mean embedding-gradient directions, and Fig. 3 shows terminal normalized co-user dispersion.
  • W2: the geometric controls are in Tables 3 and 14, and the independently tuned CoUser comparison is in Table 15. The revised mechanism distinguishes fixed geometric contraction from embedding corrections induced by a shared field's fitting residuals.
  • W3: Eq. (9) expresses endpoint error through residuals at generated integration stages. This connects reference-state regression to the generated-state constraint in Eq. (10). The loss-removal, negative-target and target-gradient controls are in Tables 17, 18 and 20. Sec. 6.5.6 and Table 5 compare the static scorer with candidate-conditioned endpoint readouts, and Eq. (12) gives the gradient route into the retained static embeddings.

PACF couples co-users through the residuals of a shared field. The target-gradient component contracts embeddings of users sharing an item and label under the local conditions in Eq. (16). Trajectory consistency penalizes accumulated residuals at generated states. Both auxiliary losses update the static embeddings through Eq. (12). The endpoint readouts instead integrate a separate trajectory for each candidate. As the endpoint approaches its conditioning item, its inner-product score approaches the item norm squared and its distance score approaches zero. These limits explain the distinction between endpoint agreement and the static ranking score directly supervised by BPR.


With the discussion period nearing its end, we would appreciate your taking a look at our rebuttal and these revisions. If they resolve your concerns, please consider revising your score upward. We would be happy to clarify any remaining questions.

AuthorsOpenReview ↗

Follow-up: brief summary before the discussion closes

Thank you again for your review.

Our rebuttal and follow-up answer all of your weaknesses.

App. D.7 of the revised manuscript shows that one shared velocity field aligns all user–item paths and unifies PACF's components, and Lemma 7 shows that PACF contains the EP+ and CoUser penalties of Table 3 as special cases. Tables 4 and 16 to 20 show that each pathwise component lowers performance when removed under PACF's selected hyperparameters.

With two days of discussion left, we would be grateful if you could consider raising your score should these responses resolve your concerns.

AuthorsOpenReview ↗

Summary of our responses during discussion phase

Thank you again for recognizing PACF's practical design and its interpretable theory. Your questions led us to add trained-model diagnostics, direct comparisons and matched controls to the revised manuscript.

As the discussion period closes today, we summarize all answers in the table below, together with their locations in our responses and in the revised manuscript, to help you check each point directly.

QuestionAnswerResponseRevised paper
W1. Practical validation of the theoryEq. (16) proves that the shared field contracts co-user embeddings near the item. Trained models confirm this prediction, since both auxiliary descent directions reduce co-user dispersion, whereas the BPR direction increases it. Eq. (E4) then links this contraction to ranking, and detaching the target gradient accordingly lowers performance.W1, #ATCM-F2.8Sec. 5.1, Figs. 9, 10, Tables 9, 20
W2. Simpler regularizersPACF leads DirectAU, UltraGCN, SCCF, LightCCF and CoUser on all seven datasets. Beyond this empirical lead, Lemma 7 shows that PACF contains EP+ and CoUser as two special cases on positive pairs, a zero field and a field that ignores the path state. PACF thus extends both with the path state and the negative regression, and accordingly exceeds both under the same hyperparameters.W2, #14vB-Q1Sec. 6.5.1, Lemma 7, Tables 2, 3, 15
W3. Experimental controlsEach control keeps PACF's selected hyperparameters, splits and seeds, so each comparison isolates the changed component. Every control lowers performance over five seeds, so each component contributes to PACF's performance. PACF also ranks first by five-seed mean on every cell.W3, #14vB-Q4Tables 4, 6, 7, 16 to 20

Other reviewers:

  • Reviewer #ATCM asked for a unified motivation of PACF's components, answered in F2.1 and F2.9 of Follow-up 2 to #ATCM. One shared velocity field aligns all user–item paths, and all six components of PACF serve this one mechanism.
  • Reviewer #14vB asked about explicit collaborative constraints, state sensitivity, the sign reversal, paired tests, an Amazon ablation, the solver and item popularity, answered in #14vB-Q1 to Q7.

We hope these answers address all of your questions. If they fully resolve your concerns, we would sincerely appreciate your considering an updated assessment.

Reviewer_ATCMOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Summary

PACF augments BPR-MF with a shared, label-conditioned velocity network used only during training. For each user–item pair, it constructs a straight interpolation path

ht=(1−t)pu+tqi,t∈[0,1],h_t=(1-t)p_u+tq_i,\qquad t\in[0,1],

and trains the network to regress a signed user-to-item displacement. Positive pairs use the target qi−puq_i-p_u, while sampled negatives use the reversed and scaled target −μ−(qi−pu)-\mu_-(q_i-p_u). Positive pairs additionally use RK4 integration and a trajectory-consistency loss that encourages the integrated endpoint to reach the observed item.

The full objective combines BPR ranking, flow matching, and trajectory consistency. Importantly, all auxiliary gradients update the static user/item embeddings. At inference, the velocity network and ODE solver are completely discarded, and recommendation remains standard MF scoring,

s(u,i)=pu⊤qi.s(u,i)=p_u^\top q_i.

The paper motivates the shared velocity network as a mechanism for introducing cross-user constraints under sparse feedback. Its main theoretical explanation is that, near a shared item endpoint, users associated with the same item present increasingly similar states while retaining different displacement targets. A shared network with bounded state sensitivity cannot fit all such targets accurately unless the corresponding user embeddings are sufficiently close, thereby inducing a form of co-user regularization.

Critical strengths

Strengths

  1. The core auxiliary-training idea has some novelty.
    Using a shared vector-valued regression task to generate structured residuals that reshape otherwise static MF embeddings is an interesting direction. The method is not simply another scalar ranking loss: the shared regressor creates interactions across training pairs, while its gradients are amortized into the user/item embeddings retained at inference.

  2. The paper attempts to analyze the training mechanism mathematically.
    The authors derive how the BPR, flow-matching, and trajectory losses contribute to the embedding gradients and provide a conditional analysis relating shared-item user dispersion to unavoidable regression residuals under bounded field sensitivity. Although I do not think this analysis fully justifies the complete design, it is useful that the paper attempts to explain the auxiliary optimization mechanism rather than presenting it as a purely empirical heuristic.

Critical weaknesses

Weaknesses

  1. There is a substantial mismatch between the flow-based narrative and the mechanism that appears to drive the method.
    PACF is not a flow-based recommender at inference: the velocity network and solver are discarded, and the final recommender remains static dot-product MF. The mechanism actually supported by the analysis is that a capacity-limited shared regressor produces residuals that reshape the embeddings, approximately inducing co-user smoothing near a shared item. This is a plausible regularization mechanism, but the paper does not establish that continuous-time transport or flow matching is the essential reason it works.

  2. The negative branch is highly heuristic, yet it is important to the reported gains.
    A sampled unobserved item is assigned the target

    −μ−(qi−pu),-\mu_-(q_i-p_u),

    despite having no meaningful transport endpoint or clear preference semantics. Moreover, users that happen to sample the same item as a negative are also coupled by the shared-field argument, although such users do not form an obvious collaborative group. The scale μ−\mu_- is tuned per dataset and varies widely across the selected configurations.

    This is not a minor component. In Table 8, setting μ−=0\mu_-=0 reduces ML Recall/NDCG from 7.48/3.737.48/3.73 to 7.08/3.337.08/3.33, and Food from 9.75/4.799.75/4.79 to 9.17/4.509.17/4.50. These values also fall below the strongest reported baseline on the corresponding metrics. Therefore, an important part of the final performance depends on a branch whose modeling semantics are not adequately justified.

  3. Conditioning the velocity network on the ground-truth interaction label is conceptually problematic.
    The network receives y∈{+1,−1}y\in\{+1,-1\} as an explicit input. Therefore, it does not learn whether the user prefers the candidate or even infer the sign of the preference direction; this information is supplied directly.

    This problem becomes particularly clear at inference. Since the true label of a candidate is unknown, the endpoint-readout experiment evaluates every candidate with y=+1y=+1. The model is therefore explicitly asked to transport the user toward every candidate. The resulting trajectory readouts perform extremely poorly: in Table 4, ML NDCG drops from 3.733.73 with the static scorer to 0.230.23 and 0.500.50 with the two endpoint readouts, and Food drops from 4.794.79 to 0.950.95 and 2.422.42. This seems structurally implied by the conditioning design rather than evidence that the learned trajectory captures recommendation preference.

  4. The main theoretical explanation is limited and relatively trivial compared with the full method.
    Near a shared item endpoint, several users present similar states to the same least-squares regressor while retaining different displacement targets. A shared prediction therefore approaches a conditional mean, leaving residuals related to deviations such as

    pu−pˉi.p_u-\bar p_i.

    This resembles a direct co-user dispersion regularizer. The effect follows naturally from shared least-squares fitting under limited model sensitivity. It does not explain why flow matching, continuous time, ODE integration, RK4, or trajectory consistency are necessary, and it does not provide a principled justification for the negative branch.

  5. The paper does not establish that the flow formulation is necessary compared with simpler regularizers.
    The theory itself suggests a much simpler regularizer of the form

    ∑i∑u∈Si∥pu−pˉi∥22,\sum_i\sum_{u\in S_i}\|p_u-\bar p_i\|_2^2,

    or related co-user, graph/Laplacian, or shared-regression objectives. However, the paper does not compare against such alternatives.

    The EP+ and EP$\pm$ controls only regularize user–item endpoint distances. They are not equivalent to the co-user regularization implied by the paper's own theory. Thus, showing that PACF outperforms these endpoint penalties does not establish that the flow/ODE machinery itself is necessary.

  6. The motivation understates the cross-user coupling already present in BPR-MF.
    It is true that the direct gradient of a user embedding is computed from that user's observed items and sampled negatives. However, item embeddings are shared parameters: different users interacting with the same item already jointly update that item representation, which subsequently influences their user updates. Standard MF therefore already has item-mediated cross-user information transfer.

    PACF introduces a different and more direct form of co-user smoothing; it does not introduce cross-user constraints into MF for the first time. The paper should more clearly establish why this additional coupling is necessary and how it differs fundamentally from existing MF or graph regularization.

  7. The usual fixed-source/target interpretation of flow matching does not directly apply.
    Both pup_u and qiq_i are trainable. Therefore, the source point, target point, interpolation path, and target velocity all move throughout optimization, and gradients are intentionally propagated through the target displacement itself. PACF is therefore better understood as joint representation learning with a moving, self-constructed regression target than as transport between fixed source and target distributions.

    The theoretical results establish conditional properties for the current embeddings and field, but do not characterize the joint optimization dynamics of these moving endpoints.

  8. The method appears to rely on imperfect fitting by the velocity network.
    For t<1t<1, the complete network input analytically determines the original user embedding,

    pu=ht−tqi1−t,p_u=\frac{h_t-tq_i}{1-t},

    and therefore determines the target displacement. An unrestricted field can in principle fit the target pointwise. If the regression residual becomes zero, the flow term produces no embedding gradient.

    This creates an unusual tension with the flow-matching interpretation: better velocity-field fitting is not necessarily better for the final recommender. The field-capacity experiment in Table 10 is consistent with this issue: increasing network width improves endpoint fitting while ranking changes non-monotonically. The effective mechanism may therefore be structured approximation error from a capacity-limited shared regressor rather than accurate flow modeling.

  9. The sparse-feedback motivation is not directly validated at the user level.
    The paper mainly uses dataset-level sparsity and global interaction-removal experiments. These are not equivalent to showing that users with short histories benefit more. Dataset sparsity also depends strongly on catalog size and does not directly measure user activity.

    Since the motivating claim concerns users with insufficient individual constraints, the paper should report performance segmented by training-history length. Furthermore, because the proposed mechanism depends on users sharing items, an even more direct analysis would segment users by the amount of co-user support available through their interacted items.

  10. The claimed generality is not demonstrated across backbones.
    PACF is implemented only on free MF ID embeddings. LightGCN, SimGCL, LightCCF, and other methods appear only as baselines; the paper does not evaluate LightGCN+PACF or another non-MF backbone. The appendix explicitly leaves graph-encoder extensions to future work because the gradient analysis would need to account for graph propagation.

    The current evidence therefore supports PACF primarily as an MF regularization method, rather than as a generally validated collaborative-filtering framework.

Questions for authors

Please refer to the weakness part.

8 comments

Thank you for your thorough review. We have uploaded a revised manuscript. Each heading notes the revision locations.

W1 & W6. Static MF regularization & pathwise training (Sec. 1, Sec. 4.5, Sec. 6.5, App. C.4, Tables 3–5, 14–18)

Inference is static matrix factorization by design. After the velocity network and ODE solver are discarded, ranking uses pu⊤qip _ u^\top q _ i only. Continuous time and trajectory consistency are therefore training machinery: they add residual gradients beyond BPR's item-mediated route,

−∇puℓ=σ(−Δs)(qi+−qi−)−λ∇puℓf−2βJΦ(pu)⊤ru.-\nabla _ {p _ u}\ell = \sigma(-\Delta s)(q _ {i^+}-q _ {i^-}) - \lambda\nabla _ {p _ u}\ell _ f - 2\beta J _ {\Phi}(p _ u)^\top r _ u.

Eq. (37) sends local flow residuals into pup _ u, and the RK4 endpoint residual returns through JΦ(pu)⊤ruJ _ {\Phi}(p _ u)^\top r _ u as well.

Proposition 2 organizes the shared-predictor residual conflict under a sensitivity bound. Negative regression and LtL _ t complete that same training loop in practice. Sec. 6.5 and #ZXki-W3 confirm their ranking value under PACF's selected hyperparameters: removing LfL _ f or LtL _ t, restricting sampling to t=0t=0, t=0.5t=0.5, or Grid, or replacing the field by EP$^+/EP/EP^\pm$ lowers ranking toward BPR-only. Sec. 6.5.6 then shows that the retained gain is read from the static scorer. The result is a closed residual-coupling loop for static matrix factorization.

W2. Negative branch & μ−\mu _ {-} scale (Sec. 4.2, App. C.4, Tables 18, 19, 21)

Negatives enter only LfL _ f, with reversed target −μ−(qi−−pu)-\mu _ {-}(q _ {i^-}-p _ u) inside one shared residual MLP.

Under BPR's sampler, a negative item i−i^- is drawn from the unobserved set of user uu. When several users draw the same i−i^- in a batch, they share the item and the negative label, but their Const targets still differ: uu−−uv−=μ−(pu−pv)\mathbf{u} _ u^--\mathbf{u} _ v^-=\mu _ {-}(p _ u-p _ v). Near the item end their states approach while those targets remain offset, which is exactly the shared-fitting conflict of Proposition 2 on the repulsive branch. Leftover conflict that remains after the field fit then updates those users through Eq. (37). Negative sampling thus extends residual coupling to the repulsive side of BPR. Ranking under matched controls:

SettingMLFoodBaby
w/o negative regression2.06±0.454.09±0.051.35±0.01
μ−=0\mu _ {-} = 01.98±0.304.09±0.081.39±0.01
Unsigned negative target2.23±0.304.09±0.051.35±0.01
μ−∗/2\mu _ {-}^*/22.87±0.344.46±0.041.44±0.02
min⁡(2μ−∗,1)\min(2\mu _ {-}^*,1)2.96±0.264.46±0.051.43±0.01
PACF3.11±0.154.46±0.041.44±0.01

Negative-to-BPR gradient ratios at t=0.5t=0.5, scaled by 10−310^{-3}:

SettingMLFoodBaby
μ−=0\mu _ {-} = 00.131.800.57
Unsigned negative target158.704.691.66
PACF57.252.031.89

Opposite targets improve ranking. Unsigned targets raise the gradient norm on ML and Food yet hurt accuracy, so the signed Const design contributes structure beyond raw gradient magnitude. Algebraically, the appendix target-route pull still carries μ(y)2\mu(y)^2, which keeps a co-user-mean attraction under the shared signed field. Once the sign is reversed, halving or doubling μ−∗\mu _ {-}^* keeps most of the gain.

W3. Label conditioning vs. static readout (Sec. 4.5, Sec. 6.5.6, Table 5)

Supplying the label yy during training is intentional, because it lets one network fit both signed targets. At test time PACF scores by ordinary inner products a(u,i)=pu⊤qia(u,i)=p _ u^\top q _ i. Table 5 compares three post-training readouts:

ReadoutMLFoodBaby
aΦa _ \Phi0.23±0.080.95±0.010.66±0.01
adista _ {\mathrm{dist}}0.50±0.152.42±0.260.40±0.04
aa3.73±0.034.79±0.051.44±0.01

After LtL _ t trains Φ1\Phi _ 1 toward the conditioning item, endpoint readouts collapse across candidates. The useful signal is therefore amortized into the static embeddings used for inner-product scoring.

W4 & W7. Proposition 2 diagnostics & moving targets (Sec. 5.1, App. C.3–C.4, Figs. 3, 9–10, Tables 8, 9, 20)

Proposition 2 lower-bounds shared residual conflict by co-user dispersion under a sensitivity bound. On trained models the same pathway is visible locally: finite ∥Jh∥2\lVert J _ h \rVert _ 2 at sampled states, positive target–prediction gaps in Fig. 10, and Table 9 directional derivatives in units of 10−610^{-6} with Df,Dt<0D _ f,D _ t<0. These measurements support residual conflict and an auxiliary contraction direction on average.

TargetDetach removes gradients through ut\mathbf{u} _ t for the whole training path while keeping state, item, and trajectory routes. Ranking then falls in #ZXki-W3, so the moving-target route is an active whole-path training signal. Near t→1t\to 1 the state factor 1−t1-t vanishes analytically, so whole-path target gradients stay complementary precisely where the state route thins out.

W5. Comparison with CoUser & graph constraints (Sec. 6.5.1, App. C.4, Tables 2, 3, 14, 15)

EP$^+$ and EP$^\pm$ pull user–item endpoints. We therefore add an explicit co-user Laplacian baseline, and we also report related published graph and contrastive methods.

Two protocols cover the comparison:

  1. Matched substitutes under PACF's selected hyperparameters, see #ZXki-W3 and Tables 3–4 and 14. These rows cover BPR-only, CoUser, EP$^+/EP/EP^\pm$, sampling, loss, and negative-branch controls.
  2. Full HPO comparisons, see #14vB-Q1 and Tables 2 and 15. These rows cover UltraGCN, DirectAU, SCCF, LightCCF, and independently tuned CoUser across seven datasets.

The identity ∑u∈Si∥pu−qi∥22=∑u∈Si∥pu−pˉi∥22+∣Si∣∥pˉi−qi∥22\sum _ {u\in S _ i}\lVert p _ u-q _ i \rVert _ 2^2=\sum _ {u\in S _ i}\lVert p _ u-\bar{p} _ i \rVert _ 2^2+|S _ i|\lVert \bar{p} _ i-q _ i \rVert _ 2^2 shows that CoUser shrinks users toward pˉi\bar{p} _ i and leaves pˉi\bar{p} _ i unanchored at qiq _ i. Fig. 3 then shows stronger normalized within-group contraction for CoUser on ML and Food, while PACF sits between CoUser and Target-detach. Residual coupling is therefore a different collaborative geometry from an unconditional Laplacian: it absorbs common displacements into vθv _ \theta and writes leftovers into the embeddings. Both table protocols above convert that geometry into ranking leadership, under matched PACF substitutes and under full HPO against published baselines.

W8. Capacity dhd _ h, residuals, and ranking (App. C.5, Tables 22–24)

Proposition 2 already says that a finite field leaves residuals when co-users disagree. The width sweep tests endpoint fit and static ranking as separate objectives. We therefore sweep hidden width dh∈{64,128,256,512}d _ h\in\{64,128,256,512\} and report residual norms, Jacobian norms, and the negative endpoint ratio:

Width dhd _ hML E+E^+ML J+J^+Food E+E^+Food E−E^-Food ρ−\rho^-Food J+J^+Baby E+E^+
642.6224.831.231.081.206.932.16
1281.3323.991.071.072.055.601.98
2560.8124.000.971.001.204.761.67
5120.7411.860.990.700.256.401.28

NDCG@10 at the same widths:

Width dhd _ hMLFoodBaby
643.094.391.43
1282.844.471.43
2563.044.591.44
5123.184.481.44

E±E^\pm stay positive at t=0.5t=0.5, so shared-fitting conflict remains available to Eq. (37) across the width sweep. The ranking message is sharper: as dhd _ h grows, endpoint residuals fall, and NDCG peaks at intermediate width on Food (dh=256d _ h=256). Capacity therefore trades shared fitting against residual updates. The peak shows that better endpoint fit need not raise ranking, without identifying which residual components drive that peak.

W9. User history & co-user support (App. C.7, Tables 28, 29)

The sparse-feedback claim is tested at the user level. Table 28 first shows gains on the dominant history groups: all active Food strata, covering 99.7% of users, and ML users with 21+21+ interactions, covering 91% of users. Holding history fixed, Table 29 then sorts users into co-user-support quartiles from lowest-support Q1 to highest-support Q4. The table below reports NDCG@10:

SupportML BPRML PACFFood BPRFood PACFBaby BPRBaby PACF
Q11.9050.9016.1436.0631.5711.558
Q22.3350.4454.4614.6591.5611.580
Q31.7544.9903.7853.8201.3231.315
Q41.1711.8372.5872.5681.5331.536

PACF's edge is therefore neighborhood-dependent: gains concentrate on ML Q3–Q4, Food Q2–Q3, and Baby Q2/Q4 after history is held fixed. Low support measures a smaller co-user neighborhood, and support count measures neighborhood size. The tables therefore show a stratified ranking footprint that tracks residual coupling through collaborative neighborhoods.

W10. Scope of current PACF to MF (App. A, Eqs. 13–15)

Our residual analysis treats free matrix-factorization tables. A graph encoder, by contrast, inserts an extra filter before residuals reach the scored embeddings.

Let EE stack the scored embedding tables PP and QQ, and let G=∇ERG=\nabla _ E R be the pathwise residual gradient. In matrix factorization the scored table is free, so one gradient step is ΔEMF=−ηG\Delta E _ {\mathrm{MF}}=-\eta G. A linear graph encoder instead forms E=ΠE(0)E=\Pi E^{(0)} with Π=1m+1∑ℓ=0mA^ℓ\Pi=\frac{1}{m+1}\sum _ {\ell=0}^{m}\hat{A}^{\ell}, and the induced update becomes ΔEgraph=−ηΠ2G\Delta E _ {\mathrm{graph}}=-\eta\Pi^{2}G. Modes with eigenvalue ∣s∣<1|s|<1 are therefore damped by s2s^{2} before residuals arrive. Current PACF therefore targets free matrix-factorization tables, where residual updates reach the scored embeddings directly; graph-encoder extensions require repeating the residual analysis under that filter.

Reviewer_ATCMOpenReview ↗

Response to the Rebuttal

Thank you for the detailed rebuttal and additional experiments. The new evidence addresses several gaps in the original submission and makes the empirical effectiveness of PACF more convincing.

However, my main concern remains the gap between the theoretical motivation and the actual method. After the clarifications, PACF appears to be an auxiliary regularizer combining signed displacement regression, a shared MLP, and trajectory consistency to improve static MF embeddings. While these components are empirically useful, the paper still lacks a coherent motivation that explains why this particular combination is necessary or theoretically well justified.

In particular, the negative branch uses a reversed velocity target along a user–item path without a corresponding endpoint constraint. The rebuttal demonstrates its empirical benefit but does not establish a convincing flow-matching interpretation or explain how it induces the intended geometric behavior. More broadly, the residual-coupling theory does not sufficiently explain the observed ranking improvements.

As a result, the method still feels more like a combination of effective design choices than a unified, well-motivated research contribution. The rebuttal improves my confidence in the empirical results, but it does not sufficiently strengthen the paper's novelty and conceptual contribution.

I will therefore raise my score from Clear Reject to Weak Reject, but I am not yet convinced that the paper meets the acceptance threshold.

Reviewer_14vBOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PACF, which adds signed flow-matching and trajectory-consistency losses to BPR-based matrix factorization. A shared velocity network regularizes user and item embeddings during training and is discarded at inference. The paper analyzes how bounded state sensitivity induces coupling between users sharing an item and evaluates the method on seven datasets, with ablations, robustness tests, and efficiency measurements.

Critical strengths

An interesting mechanism with simple inference. PACF uses shared pathwise regression to shape static embeddings while retaining ordinary inner-product scoring. The theoretical analysis offers a useful explanation of how this creates additional user coupling, making the approach more compelling than an auxiliary loss supported only by performance gains.

Broad empirical support. PACF achieves the highest mean scores across seven datasets, with five-seed statistics and extensive supplementary analyses. The stronger gains on some sparse datasets, together with the component and robustness experiments, provide encouraging evidence of its practical value.

Critical weaknesses

The proposed mechanism is not fully connected to the empirical results. The coupling bound depends on state sensitivity, but its magnitude is not measured in the trained network. Comparisons with explicit collaborative constraints are also missing, leaving some uncertainty about whether shared-field fitting offers a distinct advantage over simpler regularization.

Some improvements over strong baselines remain uncertain. Although five-seed statistics are provided, several margins are small relative to the reported variability, particularly on Yelp. Without paired tests or confidence intervals for the differences, the evidence for consistent superiority is weaker than the best mean scores alone suggest.

Questions for authors

BPR-MF already shares information across users through item embeddings. The paper explains how PACF adds another form of coupling, but a closer comparison with explicit constraints, such as graph-regularized MF or UltraGCN, would help show what the shared-field approach brings beyond existing methods.

Proposition 2 offers a useful explanation, but its strength depends on the state sensitivity (L_h). Some measurements of local state sensitivity during training, alongside the dispersion of users sharing an item, would help connect the theoretical argument to the learned model.

The negative-target experiments show that label conditioning and negative-pair regression are useful. They do not directly isolate the benefit of reversing the target direction. Given the reparameterization in Eq. (23), please discuss what the sign reversal contributes within the shared network.

Reporting five-seed results is helpful. Still, some margins over the strongest baseline are small, such as the Recall difference on Yelp. Paired tests or confidence intervals for these differences would make it easier to judge which improvements are reliable.

Food and the interaction-removal experiments already provide evidence under sparse feedback. Since the largest relative gains occur on the Amazon datasets, repeating the main loss-component comparison on one of them would further support the explanation across different domains.

The RK4 step-count study is useful, and Remark 5 highlights an interesting role for the final evaluation at (t=1). A comparison with another solver would help clarify how much of the benefit comes from trajectory consistency and how much depends on this particular discretization.

Tables 11–12 provide useful analysis of scoring geometry and popularity. A breakdown of performance on popular and long-tail items would complement these results and show which items benefit most from the proposed regularization.

7 comments

We appreciate your constructive feedback.

We have uploaded a revised manuscript. Each heading notes the corresponding revision locations in the manuscript.

Q1. Explicit constraints & CoUser (Sec. 6.5.1, App. C.4, Tables 2, 3, 14, 15)

BPR already lets users influence each other through shared item embeddings. PACF keeps that path and adds a second training route through one shared velocity field. The field first absorbs common user–item displacements. Leftover residuals then update the embeddings through Eq. (37):

∇pu∥eu(t)∥22=2[(1−t)Jh+μ(y)I]⊤eu(t).\nabla _ {p _ u}\lVert e _ u(t) \rVert _ 2^2 = 2[(1-t)J _ h + \mu(y)I]^\top e _ u(t).

Under the item-end shared-prediction reading of Proposition 2, those residuals create a force that pulls co-user embeddings toward a common mean along the target route of Eq. (37). The measured mean dispersion derivatives in #14vB-Q2 then show that the auxiliary path terms are oriented toward lower relative dispersion, which is the geometric signature of that residual route.

To explore more explicit collaborative constraints, we evaluate UltraGCN [1], DirectAU [2], SCCF [3], LightCCF [4], and graph-Laplacian MF CoUser, with Rc=1∣E∣∑i∑u∈Si∥pu−pˉi∥22R _ c=\frac{1}{|E|}\sum _ i\sum _ {u\in S _ i}\lVert p _ u-\bar{p} _ i \rVert _ 2^2. The table below reports five-seed NDCG@10 in percent as mean ± STD:

MethodMLFoodDanceYelpBabySportsClothing
BPR-MF2.66±0.653.36±0.083.29±0.081.37±0.130.67±0.010.74±0.030.30±0.08
UltraGCN [1]3.03±0.294.53±0.094.87±0.092.10±0.030.74±0.351.72±0.021.31±0.02
DirectAU [2]2.72±0.214.45±0.054.81±0.102.08±0.021.32±0.041.79±0.021.49±0.02
SCCF [3]2.87±0.194.52±0.084.83±0.112.08±0.041.34±0.041.86±0.021.41±0.02
LightCCF [4]3.56±0.284.68±0.065.18±0.042.21±0.011.40±0.031.91±0.041.47±0.02
CoUser3.28±0.224.50±0.054.86±0.101.67±0.051.23±0.041.66±0.051.26±0.06
PACF3.73±0.034.79±0.055.25±0.052.23±0.051.42±0.031.97±0.031.59±0.02

The identity ∑u∈Si∥pu−qi∥22=∑u∈Si∥pu−pˉi∥22+∣Si∣∥pˉi−qi∥22\sum _ {u\in S _ i}\lVert p _ u-q _ i \rVert _ 2^2=\sum _ {u\in S _ i}\lVert p _ u-\bar{p} _ i \rVert _ 2^2+|S _ i|\lVert \bar{p} _ i-q _ i \rVert _ 2^2 then separates two controls behind that ranking lead:

  1. CoUser HPO in Table 15 tunes an unconditional Laplacian on BPR-MF across all seven datasets.
  2. Fixed-hyperparameter CoUser / EP$^+$ / EP$^\pm$ in Table 3 and #ZXki-W3 replace PACF's pathwise terms under PACF's selected settings.

Fig. 3 then reports normalized co-user dispersion. On ML and Food, CoUser contracts groups most strongly, whereas PACF sits between CoUser and Target-detach. That geometric contrast shows residual coupling as a different collaborative mechanism from an unconditional Laplacian, and the tables above convert that mechanism into ranking leadership under both HPO and matched protocols.

[1] Mao et al. UltraGCN: Ultra Simplification of Graph Convolutional Networks for Recommendation. CIKM 2021.

[2] Wang et al. Towards Representation Alignment and Uniformity in Collaborative Filtering. KDD 2022.

[3] Wu et al. Unifying Graph Convolution and Contrastive Learning in Collaborative Filtering. KDD 2024.

[4] Zhang et al. Unveiling Contrastive Learning's Capability of Neighborhood Aggregation for Collaborative Filtering. SIGIR 2025.

Q2. Sensitivity LhL _ h & dispersion VV (Sec. 5.1, App. C.3, Figs. 3, 9–10, Tables 8, 9)

Proposition 2 needs a shared predictor and a sensitivity bound that controls state variation. On trained models we therefore probe that pathway directly at sampled states and t=0.5t=0.5. The state-Jacobian norm ∥Jh∥2\lVert J _ h \rVert _ 2 stays finite. Fig. 10 shows positive target–prediction gaps for co-users, so shared-fitting conflict is visible. Table 9 then reports directional derivatives of mean normalized co-user dispersion in units of 10−610^{-6}: Dr>0D _ r>0 for BPR, while Df,Dt<0D _ f,D _ t<0 for the two auxiliary terms. Near-zero CfC _ f further shows that the flow update has a distinct direction from BPR. Consistent with that directional picture, Fig. 3 shows lower terminal dispersion than BPR.

Taken together, these measurements make the residual pathway concrete: conflict is observed, auxiliary gradients have a distinct geometry, and terminal dispersion moves in the predicted direction. The matched ablations in #ZXki-W3 then show that deleting the same path ingredients lowers ranking, which closes the mechanism-to-performance link.

Q3 & Q5. Negative target sign & Amazon ablations (Sec. 4.2, Sec. 6.5.2–6.5.5, App. C.4, Tables 4, 18–21)

Because one shared network serves both labels, the negative regression target points opposite the positive target. That signed Const design lets a single residual field absorb both branches of BPR's sampler.

Eq. (30) first rescales the network output by 1/μ(y)1/\mu(y). Eq. (37) then maps residuals into embeddings through both state and target routes. In the appendix item-end calculation, the target-route term yields a co-user-mean attraction whose leading coefficient contains μ(y)2\mu(y)^2, so the signed field still writes a structured residual update into the embeddings. The unsigned and μ−=0\mu _ {-} = 0 controls therefore confirm the ranking value of that shared signed-regression design.

The empirical point follows in #ATCM-W2 and #ZXki-W3:

  1. Opposite regression targets improve ranking. Replacing the reversed negative target by an unsigned one more than doubles the negative-branch gradient on ML and Food, yet ranking falls. The signed Const target therefore contributes structure beyond raw gradient magnitude.
  2. Amazon Baby needs both pathwise losses and the negative branch. Under PACF's selected hyperparameters, removing LfL _ f, LtL _ t, or negative regression returns Baby NDCG@10 to the BPR-only level. Local velocity fitting and endpoint consistency therefore close the same residual loop on this sparse e-commerce set.

Q4. Statistical reliability (App. C.1–C.2, C.4, Tables 6, 7)

Tables 6 and 7 place PACF first by five-seed mean on every dataset and metric. The tightest published cell is Yelp Recall@10, PACF 4.49±0.084.49\pm0.08 versus LightCCF 4.46±0.054.46\pm0.05. On the exact Table 6 seeds 2025–2029, the paired gaps are dˉ=+0.030±0.053\bar d=+0.030\pm0.053 for Recall and +0.024±0.044+0.024\pm0.044 for NDCG, with 95% CIs [−0.036,0.096][-0.036,0.096] and [−0.031,0.079][-0.031,0.079]. The main comparison therefore shows mean leadership on every cell, and the Yelp paired gaps remain positive with intervals that cover both signs under this five-seed protocol.

#ZXki-W3 further reports matched-hyperparameter ablations of the residual-coupling loop. Holding splits, seeds, and the PACF HPO setting fixed, removing the shared field, restricting path support to t=0t=0/0.50.5/Grid, replacing signed Const by unsigned or μ−=0\mu _ {-} = 0 targets, or dropping LfL _ f/LtL _ t all fall toward BPR-only on ML, Food, and Baby, while full PACF stays highest. Those gaps show the ranking contribution of each residual-coupling module under the fixed protocol.

Q6. Solver choice (App. C.4, Table 13) The ranking gain comes from supervising the integrated endpoint. We therefore compare four fixed-step solvers under the same budget of 4 function evaluations per trajectory, together with adaptive DOPRI5. The table below reports five-seed NDCG@10 in percent as mean ± STD:

SolverMLFoodBaby
RK43.11±0.154.46±0.041.43±0.01
Explicit midpoint3.05±0.264.48±0.031.43±0.01
Heun3.03±0.284.47±0.041.43±0.01
Euler3.16±0.354.46±0.041.43±0.01
DOPRI53.07±0.244.48±0.071.44±0.01

All five solvers beat the LtL _ t-ablated variant by returning endpoint residuals through JΦ(pu)⊤J _ \Phi(p _ u)^\top. Endpoint supervision is therefore robust across these discretizations.

Q7. Item popularity strata (App. C.6, Tables 26, 27)

Because residual coupling acts through shared items, PACF should concentrate its effect where co-users exist. We therefore partition items into Head, Mid, and Tail by training frequency and report NDCG@10:

StratumML BPRML PACFFood BPRFood PACFBaby BPRBaby PACF
Head5.1964.8257.5957.5752.6772.684
Mid1.6711.6815.1625.1720.5300.499
Tail0.0530.2541.8561.8440.2460.223

On dense ML, PACF lifts Mid and Tail while lowering Head relative to BPR-only. On Food, Mid accuracy rises. On sparse Baby, the Head NDCG edge appears where co-user overlap concentrates. Across strata, the gains therefore concentrate where collaborative support is available, matching the item-side footprint of residual coupling.

AuthorsOpenReview ↗

Follow-up: manuscript revisions and updated references

Thank you again for your review.

In response to Reviewer #ATCM's follow-up, we have further revised the manuscript to connect signed supervision, shared residual coupling and positive endpoint consistency, as detailed in #ATCM-F1.

Several locations cited in our earlier rebuttal have changed. The guide below points to the latest revision.

  • Q1, Q3 and Q5: the residual-to-embedding gradient previously cited as Eq. (37) is now Eq. (42).
  • Q3 and Q5: the output-rescaling identity moves from Eq. (30) to Eq. (35) and applies for μ(y)≠0\mu(y)\ne0. The motivation for the signed negative target is now in Sec. 4.1, previously cited as Sec. 4.2. Fig. 8 illustrates the signed field responses.
  • Q2: Sec. 5.1, Eq. (16) now gives the local contraction of embeddings for users sharing an item and label, produced by a sufficiently small descent step along the target-gradient component, under a fixed state-sensitivity bound and 1−t<∣μ(y)∣/Lh1-t<|\mu(y)|/L_h. The empirical diagnostics remain in Figs. 3, 9 and 10 and Table 9. These measure terminal geometry, local sensitivity, fitting conflict and mean gradient directions.
  • Q6: Eq. (9) now expresses endpoint error as a weighted sum of residuals at generated integration stages. Flow regression fits reference states, while the trajectory loss in Eq. (10) constrains accumulated error on generated states. The solver comparison is in Table 13.
  • Q7: the item-frequency Head/Mid/Tail comparison is in Table 27, correcting the earlier reference to Tables 26–27.

Signed regression directly trains the shared field on observed positives and sampled negatives. Near a shared item, the target-gradient component contracts co-user embeddings under the local conditions in Eq. (16). Trajectory consistency penalizes accumulated residuals at generated states beyond the reference-state fitting objective. Eq. (12) routes both auxiliary gradients into the static embeddings retained for ranking.


With the discussion period nearing its end, we would appreciate your taking a look at our rebuttal and these revisions. If they resolve your concerns, please consider revising your score upward. We would be happy to clarify any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to AC & all reviewers

We thank Reviewers #14vB, #ATCM, and #ZXki for their constructive reviews.

  1. PACF discards the field and solver after training and ranks candidates by static inner products.
  2. Flow and trajectory residual gradients update the user and item embeddings during training.
  3. Proposition 2 lower-bounds shared-field fitting error for separated co-users under bounded state sensitivity.

1. Mechanism

BPR's descent in Eq. (12) combines attraction toward observed items with opposition to sampled negatives. Sec. 4.1 uses these directions and the alignment and uniformity view of DirectAU and contrastive learning to motivate the signed targets in Eq. (4). BPR supplies no field gradient. PACF directly fits both signed targets with one predictor. In Fig. 8, positive trajectory supervision without flow regression produces attraction under both labels, whereas PACF learns opposing responses.

BPR shares item parameters across users. PACF also shares their displacement predictor. For separated co-users of the same item and label, state differences shrink by 1−t1-t, while target differences remain proportional to user separation. For a field with sensitivity bounded by LhL_h, Proposition 2 rules out exact fitting of both targets when Lh(1−t)<∣μ(y)∣L_h(1-t)<|\mu(y)|. Pair-specific constant fields could make both auxiliary losses zero for arbitrary embeddings.

Eq. (42) gives the target-gradient component 2μ(y)eu(t)2\mu(y)e_u(t), where eu(t)e_u(t) is prediction minus target. The state-input component carries the factor 1−t1-t and attenuates near the item. Eq. (16) proves that sufficiently small descent steps along this component reduce co-user separation when 1−t<∣μ(y)∣/Lh1-t<|\mu(y)|/L_h.

Exact positive-field matching makes the reference segment solve the ODE and reach the observed item. With imperfect fitting, generated states can leave that segment. Eq. (9) expresses endpoint error as a weighted sum of generated-stage velocity residuals, with weights summing to one. Stage errors can cancel. Flow regression penalizes local errors at reference states, while trajectory consistency penalizes accumulated error at generated endpoints. In Eq. (12), both residual gradients update the same user embedding whose ranking margin BPR supervises. Table 5 reports higher NDCG@10 for static scoring than for endpoints conditioned on each candidate and label +1+1 on ML, Food and Baby.

2. Theory, evidence, and scope

Fig. 9 measures local state Jacobians. Positive target-minus-prediction gaps in Fig. 10 require nonzero fitting residuals. In Table 9, both weighted auxiliary descent directions have negative directional derivatives of mean normalized co-user dispersion on ML, Food and Baby. Fig. 3 reports lower terminal normalized dispersion than BPR-only.

CoUser has lower normalized co-user dispersion than PACF on ML and Food in Fig. 3 but ranks lower in Table 3. CoUser applies a fixed graph-Laplacian penalty. PACF fits displacement patterns with a shared field and updates embeddings through the remaining errors. Tables 4, 17, 18 and 20 report lower NDCG after restricting path support, removing either loss, changing negative targets or detaching target gradients on ML, Food and Baby under PACF's selected hyperparameters.

PACF leads UltraGCN, DirectAU, SCCF and LightCCF in Table 2 and independently tuned CoUser in Table 15 by five-seed mean on all seven datasets. Eqs. (17)–(19) show that graph propagation filters residual updates, whereas free MF embeddings receive them directly.

3. Concern-to-thread map

ConcernRevised evidence and locationThread
Unified motivationSigned fitting and endpoint errors train the same static embeddings.#ATCM-F1, F2
Flow vs. training residualEq. (9) sums stage errors. Eq. (12) adds both auxiliary gradients to BPR.#ATCM-W1, #ATCM-W6
Theory on trained modelsFig. 10 shows fitting gaps. Table 9 reports negative dispersion derivatives.#14vB-Q2, #ATCM-W4, #ZXki-W1
Co-user / graph constraintsCoUser has lower normalized dispersion but lower NDCG on ML and Food in Fig. 3 and Table 3.#14vB-Q1, #ATCM-W5, #ZXki-W2
Path vs. simpler substitutesRestricting sampling or removing either loss lowers NDCG in Tables 4 and 17.#ATCM-W5, #ZXki-W3
Negative sign and labelsTable 18 changes or removes negative regression. Labels supervise training only.#14vB-Q3, #ATCM-W2, #ATCM-W3
CapacityPositive endpoint fit improves with width, but ranking is nonmonotonic in Tables 22–24.#ATCM-W8
Sparse feedbackTable 29 stratifies co-user support within history groups.#ATCM-W9
Solver / popularity / AmazonTable 13 tests solvers, Table 27 reports item strata, Table 4 includes Baby ablations.#14vB-Q5–Q7
ReliabilityFive-seed statistics in Tables 6–7 and same-seed paired differences.#14vB-Q4, #ZXki-W3
ScopeEqs. (17)–(19) derive direct MF updates and graph-filtered updates.#ATCM-W10

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.