acceptodds
Under review as a conference paper at ICLR 2027

Conditional Gaussian Mixture Denoising for Cross-Domain Sequential Recommendation

Abstract

Cross-domain sequential recommendation is widely studied for alleviating the cold-start problem via knowledge transfer. Existing methods usually derive the user's noisy domain-shared multi-interest (NDI) from multi-domain interactions, with diffusion as an effective denoising strategy for adapting NDI to the target domain. However, this diffusion denoising process is initialized with a fixed, user-agnostic Gaussian prior, which may be far from the user's NDI, resulting in an excessively long denoising path. To address this, we propose a personalized prior initialization strategy, which enables efficient user adaptation by learning a user-specific NDI distribution as the denoising starting point. Specifically, a domain-shared preference modeling strategy employs a codebook-based Gaussian mixture distribution to capture the user's NDI, with a pairwise proximity regularizer preventing the mixture components from being excessively separated. Building on this, a target-domain adaptation mechanism reformulates the noising and denoising processes under the learned NDI prior, with a domain-specific log-density condition steering the denoising trajectory toward the target domain. Experiments on three real-world datasets show that our model converges in substantially fewer denoising steps than a variant using the standard Gaussian prior, without relying on step-skipping samplers, while also improving recommendation accuracy.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Sequential Recommendation, Cross-Domain, Gaussian Mixture Denoising

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_8FfzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GMD learns user-specific Gaussian mixture priors via a common codebook in the framework of cross-domain sequential recommendation. It employs domain-specific log density difference-based denoising. Three pairs of Amazon domains are utilized in experiments on users with scarce or zero interactions in the target domain.

Critical strengths

Personalized initialization is a useful direction for cross-domain adaptation, and the prior comparisons, ablations, and runtime measurements provide evidence of gains in both accuracy and efficiency.

Critical weaknesses
  1. The stated mixture prior does not correspond with the described sampling method. Equation (4) describes a Gaussian mixture, but Equations (8)–(9) weight each component's mean and variance by the weight of the component and concatenate parallel trajectories from Gaussian distributions. During inference, the output of each component is averaged. These processes are typically different from sampling the mixture in Equation (4). The authors need to state which distribution the sampling method employs, since the use of a mixture prior is key to the paper.

  2. The reduction in denoising steps requires better proof. The number of steps is compared in Figure 3, but the noise schedule and final noise level are not described well enough to confirm that the comparison is apples-to-apples. Comparison to faster standard-prior samplers at similar accuracy would also demonstrate whether the learned prior provides an advantage over other sampling methods.

Questions for authors

Would you mind explaining how the weight-scaled Gaussian trajectory Eqs. (8)–(9) relate to the mixture prior of Eq. (4)? Also, does Fig. 3 use consistent noise schedules and terminal noise for both priors?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8Ffz(1)

W1 & Q The stated mixture prior does not correspond with the described sampling method. Equation (4) describes a Gaussian mixture, but Equations (8)–(9) weight each component's mean and variance by the weight of the component and concatenate parallel trajectories from Gaussian distributions. During inference, the output of each component is averaged. These processes are typically different from sampling the mixture in Equation (4). Would you mind explaining how the weight-scaled Gaussian trajectory Eqs. (8)–(9) relate to the mixture prior of Eq. (4)? Also, does Fig. 3 use consistent noise schedules and terminal noise for both priors?

Response (1/2):

We thank the reviewer for this careful reading. The process uses the Gaussian mixture in Eq. (4). The parallel trajectories in Eqs. (8)--(9) carry the parameters of each mixture component, i.e., the mixing weights π1:NX\pi^{X} _ {1:N}, means μ1:NGX\mu^{G _ X} _ {1:N} and standard deviations σ1:NGX\sigma^{G _ X} _ {1:N} of all NN components, which preserves the mixture structure (Appendix B.5, line 939). Gumbel-Softmax (GS) then selects one component according to the mixing weights.

On the per-component expressions in Appendix B.5. The per-component expressions in Appendix B.5 (lines 939--943), N(πnμn,(πnσn)2I)\mathcal{N}(\pi _ n\mu _ n,(\pi _ n\sigma _ n)^2I), are another way of writing the Gaussian mixture. By the linear transformation property of the Gaussian distribution, if xn∼N(μn,σn2I)x _ n\sim\mathcal{N}(\mu _ n,\sigma _ n^2I), then πnxn∼N(πnμn,(πnσn)2I)\pi _ n x _ n\sim\mathcal{N}(\pi _ n\mu _ n,(\pi _ n\sigma _ n)^2I); this expression therefore represents each component together with its mixing weight in a parallel form. When the weights are realized by the Gumbel-Softmax selection wnw _ n, which assigns 1 to the selected component and 0 to the others, we have ∑nwnxn=xn∗\sum _ n w _ n x _ n=x _ {n^{\ast}} with xn∗∼N(μn∗,σn∗2I)x _ {n^{\ast}}\sim\mathcal{N}(\mu _ {n^{\ast}},\sigma _ {n^{\ast}}^2I), i.e., exactly a sample from the Gaussian mixture in Eq. (68). We have revised these expressions together with Eq. (8) and Algorithm 2 to make this correspondence explicit.

On the term "aggregation" in Appendix B.5 and C.1. We apologize that the word "aggregation" in Appendix B.5 (lines 944--950) and C.1 (lines 1046--1048) is imprecise and may have caused confusion. It refers to the Gumbel-Softmax combination in Eq. (12) (Eq. (72) in Appendix B.5), whose weights π^n\hat\pi _ n are relaxed one-hot sampling weights rather than fixed averaging coefficients. It is therefore a differentiable relaxation of selecting one component, not a fixed average of the components, and we have revised the wording accordingly.

Setting gn=gm=0g _ n=g _ m=0 is applied only at inference. During training, Gumbel noise gn∼Gumbel(0,1)g _ n\sim\mathrm{Gumbel}(0,1) is always drawn, so the model is learned by sampling components according to the mixing weights π1:NX\pi^{X} _ {1:N}, exactly as in Eq. (68). At inference, the Gumbel noise is set to zero solely to avoid randomness in the test results and make them reproducible. This does not change what the model has learned; the core of GMD remains sampling from the Gaussian mixture.

To verify that this inference-only choice does not affect the results, we compare it with keeping the default Gumbel noise at inference, i.e., stochastic sampling, averaged over five runs (NDCG@10):

Inference settingKF-FRKF-NREB-FREB-NRBM-FRBM-NR
gn=gm=0g _ n=g _ m=0 (reported)0.20520.12870.16920.11010.16960.0897
Default Gumbel sampling (mean)0.20410.12780.16810.10930.16870.0889
Default Gumbel sampling (std)0.00210.00160.00190.00140.00180.0012

The two settings yield nearly identical results: the differences (0.0008--0.0011) are all within one standard deviation of stochastic sampling. This confirms that setting gn=gm=0g _ n=g _ m=0 at inference only removes evaluation randomness and does not change the behavior of the model.

Specifically, the model prediction based on the forward (noising) process in Eq. (12),

x^pre=∑n=1N[exp⁡(log⁡πnX+gn)⋅x^0,n]∑m=1Nexp⁡(log⁡πmX+gm),gn,gm∼Gumbel(0,1),\hat{x} _ {pre}=\frac{\sum _ {n=1}^{N}\left[\exp\left(\log\pi^{X} _ {n}+g _ n\right)\cdot\hat{x} _ {0,n}\right]}{\sum _ {m=1}^{N}\exp\left(\log\pi^{X} _ {m}+g _ m\right)},\qquad g _ n,g _ m\sim\mathrm{Gumbel}(0,1),

implements the sampling process of Eqs. (8)--(9). The formulation of Eq. (8) in the paper,

q(xt∣xt−1):=N(αtxt−1+(1−αt)π1:NX⋅μ1:NGX, βt(π1:NX⋅σ1:NGX)2I),q(x _ t\mid x _ {t-1}):=\mathcal{N}\left(\sqrt{\alpha _ t} x _ {t-1}+(1-\sqrt{\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \beta _ t\left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}I\right),

may have been misleading. Here π1:NX⋅μ1:NGX\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N} and π1:NX⋅σ1:NGX\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} only denote a parallel structure that represents the Gaussian mixture; they are not computed as weighted sums. Only the single component obtained by sampling participates in the computation.

AuthorsOpenReview ↗

Response to Reviewer 8Ffz(2)

W1 & Q The stated mixture prior does not correspond with the described sampling method. Equation (4) describes a Gaussian mixture, but Equations (8)–(9) weight each component's mean and variance by the weight of the component and concatenate parallel trajectories from Gaussian distributions. During inference, the output of each component is averaged. These processes are typically different from sampling the mixture in Equation (4). The authors need to state which distribution the sampling method employs, since the use of a mixture prior is key to the paper. Would you mind explaining how the weight-scaled Gaussian trajectory Eqs. (8)–(9) relate to the mixture prior of Eq. (4)? Also, does Fig. 3 use consistent noise schedules and terminal noise for both priors?

Response (2/2):

To make this clear, we have rewritten Eq. (8) as

q(xt∣xt−1):=N(αtxt−1+Sampling((1−αt)π1:NX⋅μ1:NGX), Sampling(βt(π1:NX⋅σ1:NGX)2)I),q(x _ t\mid x _ {t-1}):=\mathcal{N}\Big(\sqrt{\alpha _ t} x _ {t-1}+\mathrm{Sampling}\big((1-\sqrt{\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N}\big),\ \mathrm{Sampling}\big(\beta _ t\left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}\big)I\Big),

where both Sampling(⋅)\mathrm{Sampling}(\cdot) operations refer to the same component selected by GS according to π1:NX\pi^{X} _ {1:N}, so the mean and variance always come from one sampled sub-distribution. Its expanded form (Eq. (9)) is

xt=αˉtx0+Sampling(N((1−αˉt)π1:NX⋅μ1:NGX, 1−αˉtπ1:NX⋅σ1:NGXϵ)),x _ t=\sqrt{\bar\alpha _ t} x _ 0+\mathrm{Sampling}\left(\mathcal{N}\left((1-\sqrt{\bar\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \sqrt{1-\bar\alpha _ t} \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right),

where ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I) is the Gaussian noise.

For the denoising sampling process (Eq. (17)), the starting point xTx _ T is drawn from the Gaussian mixture given in Appendix Eq. (42):

q(xT∣x0)≈N(π1:NX⋅μ1:NGX, (π1:NX⋅σ1:NGX)2I).q(x _ T\mid x _ 0)\approx\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}I\right).

This expression likewise denotes GS sampling of the distribution of xTx _ T. Both the noising and the denoising processes therefore use GS sampling of the Gaussian mixture.

The same applies to the initialization in Algorithm 2. Its original form,

xT=π1:NX⋅μ1:NGX+π1:NX⋅σ1:NGXϵ,x _ T=\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N}+\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon,

is likewise intended to denote sampling from the Gaussian mixture rather than scaling all components by the mixture weights. We have corrected it to

xT=Sampling(N(π1:NX⋅μ1:NGX, π1:NX⋅σ1:NGXϵ)).x _ T=\mathrm{Sampling}\left(\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right).

The core properties of the diffusion process are fully preserved under the revised formulation. The timestep tt is uniformly sampled from {1,…,T}\lbrace 1,\dots,T \rbrace. When t=1t=1, αˉt≈1\bar\alpha _ t\approx1 and thus xt≈x0x _ t\approx x _ 0. When t=Tt=T, αˉT=0\bar\alpha _ T=0 exactly (see W2 & Q), and thus

xT=Sampling(N(π1:NX⋅μ1:NGX, π1:NX⋅σ1:NGXϵ)),x _ T=\mathrm{Sampling}\left(\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right),

which is consistent with both the starting point of the denoising process in Eq. (17) and the corrected initialization in Algorithm 2.

Finally, all formulas involving the Gaussian mixture (e.g., Eqs. (4), (8)--(9) and Algorithm 2) follow the sampling procedure given in Eq. (68) of Appendix B.5. We have clarified this in the revised paper.

On Fig. 3. Yes, both priors use exactly the same noise schedule and terminal noise level (αˉT=0\bar\alpha _ T=0) at every TT; details are given in our response to W2 & Q.

Revision. Sec. 4.2.1--4.2.2 (Eqs. (8)--(9)), Appendix A (Alg. 1--2), Appendix B and B.5 (Table 4), in blue in the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer 8Ffz(3)

W2 The reduction in denoising steps requires better proof. The number of steps is compared in Figure 3, but the noise schedule and final noise level are not described well enough to confirm that the comparison is apples-to-apples. Comparison to faster standard-prior samplers at similar accuracy would also demonstrate whether the learned prior provides an advantage over other sampling methods.

Response:

We thank the reviewer for this constructive comment. Yes, Figure 3 uses matched noise schedules and terminal noise levels for both priors.

Noise schedule. We use a linear schedule that increases βt\beta _ t from 0 to 1 over TT steps:

βt=t−1T−1,t=1,…,T,αˉt=∏s=1t(1−βs).\beta _ t=\frac{t-1}{T-1},\quad t=1,\dots,T,\qquad \bar\alpha _ t=\prod _ {s=1}^{t}(1-\beta _ s).

The interval between consecutive steps depends on the total number of steps TT: the smaller TT is, the more noise is added at each step.

Matched terminal noise. Since βT=1\beta _ T=1 for every TT, αˉT=0\bar\alpha _ T=0 holds exactly for all step budgets. In Figure 3, the number of diffusion steps TT is a hyperparameter of each setting: for each value of TT, a separate model is trained with only TT changed and all other hyperparameters kept fixed. DP and SP use exactly the same schedule at every TT, and differ only in the prior (Sec. 5.4, lines 456--458). Therefore, the gap between the two curves comes from the learned prior itself.

Comparison with accelerated samplers. The purpose of Figure 3 is to verify the effect of the prior itself, i.e., that a learned prior alone already accelerates denoising. Accelerated samplers are complementary to the prior: GMD can also adopt DDIM for further acceleration. For a fair comparison, we therefore equip both priors with DDIM (KF-FR, NDCG@10):

TT51015203050
Mixture prior (DP)0.1720.1900.1940.2050.2010.207
Mixture prior + DDIM0.1860.2040.2050.2060.2060.207
Standard prior (SP)0.0790.0890.0960.1030.1200.145
Standard prior + DDIM0.1050.1160.1260.1330.1410.143

DDIM improves both priors at small step counts, and the mixture prior with DDIM performs best at every step budget, reaching 0.204 with only 10 steps. Under the same sampler, the mixture prior consistently outperforms the standard prior by a large margin: even with DDIM and 50 steps, the standard prior (0.143) remains below the mixture prior with only 5 steps (0.172). This shows that the advantage comes from the prior itself rather than from the sampling procedure, and that the two are complementary: combining the mixture prior with DDIM further accelerates GMD. We have added these details to the revised paper.

Revision. New Appendix C.3 (Table 7), Sec. 5.4 and new Appendix C.8 (Table 13, DDIM), in blue in the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer 8Ffz(4)

Summary of revisions for Reviewer 8Ffz (changes are marked in blue in the revised PDF)

IssueLocationRevision
W1 & QSec. 4.2.1–4.2.2, App. A, B, B.5Eqs. (8)–(9) rewritten as sampling from the mixture in Eq. (4); "aggregation" revised; Table 4
W2Sec. 5.4, App. C.3, C.8Table 7 (matched schedule and terminal noise); Table 13 (DDIM for both priors)

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

This paper proposes GMD for cross-domain sequential recommendation with few or no target-domain interactions. It constructs personalized priors from a shared Gaussian codebook and performs denoising conditioned on domain log-density differences. Experiments on three Amazon domain pairs report improved recommendation accuracy and fewer denoising steps.

Critical strengths
  1. Personalized denoising initialization is reasonably motivated, and the shared codebook provides an explicit mechanism for transferring knowledge to sparse users.
  2. Tables 1–2 show consistent improvements across three domain pairs and both FR and NR settings.
  3. The ablations and prior comparisons are informative. Table 5 and Figure 3 provide useful evidence for personalized initialization.
Critical weaknesses
  1. The stated mixture prior does not match the actual initialization.

Equation (68) describes categorical selection of a Gaussian component, whereas Algorithm 2 initializes all components and scales their means and standard deviations by mixture weights. This is actually not equivalent to the stated mixture. Please reconcile the probability model and explain the impact of mixture weights used for both parameter scaling and output aggregation.

  1. The cold-start evaluation protocol lacks essential details.

The paper does not specify data splits, FR history budgets, exclusion of NR target-domain interactions, or whether ranking uses full-catalog or sampled candidates. Please also clarify how codebook pretraining and cross-domain histories exclude held-out targets and future interactions. These details can be necessary to assess fairness and reproducibility.

  1. The few-step denoising claim requires a clearer experimental specification.

Figure 3 shows that GMD achieves good performance with fewer steps, but does not explain how the noise added at each step is adjusted or whether a separate model is trained for each step setting. These choices would affect the results, making it difficult to determine how much of the advantage comes from the learned prior. In addition, it would also be informative to clarify the sampling steps, inference batch sizes, and operations included in the timing measurements in Table 6.

Questions for authors
  1. Does the implementation use a Gaussian mixture or parallel scaled Gaussian components, and how are these interpretations reconciled?
  2. Can the authors provide the complete FR/NR partitioning, information-exclusion, and ranking-candidate protocols?
  3. Does the few-step advantage persist under consistent terminal noise conditions and an explicit timing protocol?
4 comments
AuthorsOpenReview ↗

Response to Reviewer LbYS(1)

W1 & Q1 The stated mixture prior does not match the actual initialization. Equation (68) describes categorical selection of a Gaussian component, whereas Algorithm 2 initializes all components and scales their means and standard deviations by mixture weights. This is actually not equivalent to the stated mixture. Please reconcile the probability model and explain the impact of mixture weights used for both parameter scaling and output aggregation. Does the implementation use a Gaussian mixture or parallel scaled Gaussian components, and how are these interpretations reconciled?

Response (1/2):

We thank the reviewer for this careful reading. The implementation uses a Gaussian mixture. The parallel components in Algorithm 2 carry the parameters of each mixture component, i.e., the mixing weights π1:NX\pi^{X}_{1:N}, means μ1:NGX\mu^{G_X}_{1:N} and standard deviations σ1:NGX\sigma^{G_X}_{1:N} of all NN components, which preserves the mixture structure (Appendix B.5, line 939). Gumbel-Softmax (GS) then selects one component according to the mixing weights.

The mixing weights π1:NX\pi^{X}_{1:N} are the selection probabilities of the mixture components.

On the per-component expressions in Appendix B.5. The per-component expressions in Appendix B.5 (lines 939--943), N(πnμn,(πnσn)2I)\mathcal{N}(\pi_n\mu_n,(\pi_n\sigma_n)^2I), are another way of writing the Gaussian mixture. By the linear transformation property of the Gaussian distribution, if xn∼N(μn,σn2I)x_n\sim\mathcal{N}(\mu_n,\sigma_n^2I), then πnxn∼N(πnμn,(πnσn)2I)\pi_n x_n\sim\mathcal{N}(\pi_n\mu_n,(\pi_n\sigma_n)^2I); this expression therefore represents each component together with its mixing weight in a parallel form. When the weights are realized by the Gumbel-Softmax selection wnw_n, which assigns 1 to the selected component and 0 to the others, we have ∑nwnxn=xn∗\sum_n w_n x_n=x_{n^{\ast}} with xn∗∼N(μn∗,σn∗2I)x_{n^{\ast}}\sim\mathcal{N}(\mu_{n^{\ast}},\sigma_{n^{\ast}}^2I), i.e., exactly a sample from the Gaussian mixture in Eq. (68). We have revised these expressions together with Eq. (8) and Algorithm 2 to make this correspondence explicit.

On the term "aggregation" in Appendix B.5 and C.1. We apologize that the word "aggregation" in Appendix B.5 (lines 944--950) and C.1 (lines 1046--1048) is imprecise and may have caused confusion. It refers to the Gumbel-Softmax combination in Eq. (12) (Eq. (72) in Appendix B.5), whose weights π^n\hat\pi_n are relaxed one-hot sampling weights rather than fixed averaging coefficients. It is therefore a differentiable relaxation of selecting one component, not a fixed average of the components, and we have revised the wording accordingly.

Setting gn=gm=0g_n=g_m=0 is applied only at inference. During training, Gumbel noise gn∼Gumbel(0,1)g_n\sim\mathrm{Gumbel}(0,1) is always drawn, so the model is learned by sampling components according to the mixing weights π1:NX\pi^{X}_{1:N}, exactly as in Eq. (68). At inference, the Gumbel noise is set to zero solely to avoid randomness in the test results and make them reproducible. This does not change what the model has learned; the core of GMD remains sampling from the Gaussian mixture.

To verify that this inference-only choice does not affect the results, we compare it with keeping the default Gumbel noise at inference, i.e., stochastic sampling, averaged over five runs (NDCG@10):

Inference settingKF-FRKF-NREB-FREB-NRBM-FRBM-NR
gn=gm=0g_n=g_m=0 (reported)0.20520.12870.16920.11010.16960.0897
Default Gumbel sampling (mean)0.20410.12780.16810.10930.16870.0889
Default Gumbel sampling (std)0.00210.00160.00190.00140.00180.0012

The two settings yield nearly identical results: the differences (0.0008--0.0011) are all within one standard deviation of stochastic sampling. This confirms that setting gn=gm=0g_n=g_m=0 at inference only removes evaluation randomness and does not change the behavior of the model.

AuthorsOpenReview ↗

Response to Reviewer LbYS(2)

W1 & Q1 The stated mixture prior does not match the actual initialization. Equation (68) describes categorical selection of a Gaussian component, whereas Algorithm 2 initializes all components and scales their means and standard deviations by mixture weights. This is actually not equivalent to the stated mixture. Please reconcile the probability model and explain the impact of mixture weights used for both parameter scaling and output aggregation. Does the implementation use a Gaussian mixture or parallel scaled Gaussian components, and how are these interpretations reconciled?

Response (2/2):

Specifically, the model prediction based on the forward (noising) process in Eq. (12),

x^pre=∑n=1N[exp⁡(log⁡πnX+gn)⋅x^0,n]∑m=1Nexp⁡(log⁡πmX+gm),gn,gm∼Gumbel(0,1),\hat{x} _ {pre}=\frac{\sum _ {n=1}^{N}\left[\exp\left(\log\pi^{X} _ {n}+g _ n\right)\cdot\hat{x} _ {0,n}\right]}{\sum _ {m=1}^{N}\exp\left(\log\pi^{X} _ {m}+g _ m\right)},\qquad g _ n,g _ m\sim\mathrm{Gumbel}(0,1),

implements the sampling process of Eq. (8). The formulation of Eq. (8) in the paper,

q(xt∣xt−1):=N(αtxt−1+(1−αt)π1:NX⋅μ1:NGX, βt(π1:NX⋅σ1:NGX)2I),q(x _ t\mid x _ {t-1}):=\mathcal{N}\left(\sqrt{\alpha _ t} x _ {t-1}+(1-\sqrt{\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \beta _ t\left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}I\right),

may have been misleading. Here π1:NX⋅μ1:NGX\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N} and π1:NX⋅σ1:NGX\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} only denote a parallel structure that represents the Gaussian mixture; they are not computed as weighted sums. Only the single component obtained by sampling participates in the computation. To make this clear, we have rewritten Eq. (8) as

q(xt∣xt−1):=N(αtxt−1+Sampling((1−αt)π1:NX⋅μ1:NGX), Sampling(βt(π1:NX⋅σ1:NGX)2)I),q(x _ t\mid x _ {t-1}):=\mathcal{N}\Big(\sqrt{\alpha _ t} x _ {t-1}+\mathrm{Sampling}\big((1-\sqrt{\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N}\big),\ \mathrm{Sampling}\big(\beta _ t\left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}\big)I\Big),

where both Sampling(⋅)\mathrm{Sampling}(\cdot) operations refer to the same component selected by Gumbel-Softmax according to π1:NX\pi^{X} _ {1:N}, so the mean and variance always come from one sampled sub-distribution. Its expanded form (Eq. (9)) is

xt=αˉtx0+Sampling(N((1−αˉt)π1:NX⋅μ1:NGX, 1−αˉtπ1:NX⋅σ1:NGXϵ)),x _ t=\sqrt{\bar\alpha _ t} x _ 0+\mathrm{Sampling}\left(\mathcal{N}\left((1-\sqrt{\bar\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \sqrt{1-\bar\alpha _ t} \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right),

where Sampling(⋅)\mathrm{Sampling}(\cdot) denotes GS sampling of a mixture component, and ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I) is the Gaussian noise.

For the denoising sampling process (Eq. (17)), the starting point xTx _ T is drawn from the Gaussian mixture given in Appendix Eq. (42):

q(xT∣x0)≈N(π1:NX⋅μ1:NGX, (π1:NX⋅σ1:NGX)2I).q(x _ T\mid x _ 0)\approx\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}I\right).

This expression likewise denotes GS sampling of the distribution of xTx _ T. Both the noising and the denoising processes therefore use GS sampling of the Gaussian mixture.

The same applies to the initialization in Algorithm 2. Its original form,

xT=π1:NX⋅μ1:NGX+π1:NX⋅σ1:NGXϵ,x _ T=\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N}+\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon,

is likewise intended to denote sampling from the Gaussian mixture rather than scaling all components by the mixture weights. We have corrected it to

xT=Sampling(N(π1:NX⋅μ1:NGX, π1:NX⋅σ1:NGXϵ)).x _ T=\mathrm{Sampling}\left(\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right).

The core properties of the diffusion process are fully preserved under the revised formulation. The timestep tt is uniformly sampled from {1,…,T}\lbrace 1,\dots,T \rbrace. When t=1t=1, αˉt≈1\bar\alpha _ t\approx1 and thus xt≈x0x _ t\approx x _ 0. When t=Tt=T, αˉT=0\bar\alpha _ T=0 exactly (see W3 & Q3), and thus

xT=Sampling(N(π1:NX⋅μ1:NGX, π1:NX⋅σ1:NGXϵ)),x _ T=\mathrm{Sampling}\left(\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right),

which is consistent with both the starting point of the denoising process in Eq. (17) and the corrected initialization in Algorithm 2.

Finally, all formulas involving the Gaussian mixture (e.g., Eqs. (4) and (8) and Algorithm 2) follow the sampling procedure given in Eq. (68) of Appendix B.5. We have clarified this in the revised paper.

Revision. Sec. 4.2.1: the Sampling(⋅)\mathrm{Sampling}(\cdot) notation is introduced, and Eqs. (8)--(9) are rewritten with it. Sec. 4.2.2: xTx _ T is written in sampling form, with αˉT=0\bar\alpha _ T=0 exactly. Appendix A: the initialization of Algorithm 2 is corrected, and the forward sampling step is added to Algorithm 1. Appendix B (Notation paragraph, B.1, B.2): the parallel form is defined as component-wise parameters, not weighted sums. Appendix B.5: new text on the per-component expressions, the revised "aggregation" wording, inference-only Gumbel noise and Table 4. Appendix C.1: inference setting. Section, table and appendix numbers in this Revision paragraph refer to the revised PDF; changes are marked in blue there.

AuthorsOpenReview ↗

Response to Reviewer LbYS(3)

W2 & Q2 The cold-start evaluation protocol lacks essential details. The paper does not specify data splits, FR history budgets, exclusion of NR target-domain interactions, or whether ranking uses full-catalog or sampled candidates. Please also clarify how codebook pretraining and cross-domain histories exclude held-out targets and future interactions. These details can be necessary to assess fairness and reproducibility. Can the authors provide the complete FR/NR partitioning, information-exclusion, and ranking-candidate protocols?

Response:

We thank the reviewer for this comment. The split unit is the user. Users are split into training, validation and test sets with an 8:1:1 ratio, and the three sets share no users. Training users contribute their full cross-domain sequences. For validation and test users, we sort target-domain interactions by time. Since FR test users have only a few target-domain interactions, we hold out only the last one as the prediction target and keep the rest as history, so each FR test user has 1--3 target-domain interactions in the history. FR users thus have a non-empty target-domain history (SuB≠∅S _ u^B \neq \varnothing). NR users have no target-domain history (SuB=∅S _ u^B = \varnothing) and use only their source-domain sequences as input. All cross-domain sequences are strictly ordered by timestamp (Appendix C.1), and only interactions that occur before the prediction target, including source-domain interactions, are used as input, so no future information is introduced. Validation users are used only for model selection. Test ranking covers all target-domain items, and each user's history interactions are filtered out.

The codebook is pretrained only on training users' sequences (see Appendix C.1 for the two-stage training) and then frozen, so it never sees interactions of validation or test users. All baselines use the same user split and ranking protocol. We have added these details to Sec. 5.1 and the appendix of the revised paper.

Revision. Sec. 5.1 and new Appendix C.2 (Evaluation Protocol), in blue in the revised PDF.

W3 & Q3 The few-step denoising claim requires a clearer experimental specification. Figure 3 shows that GMD achieves good performance with fewer steps, but does not explain how the noise added at each step is adjusted or whether a separate model is trained for each step setting. These choices would affect the results, making it difficult to determine how much of the advantage comes from the learned prior. In addition, it would also be informative to clarify the sampling steps, inference batch sizes, and operations included in the timing measurements in Table 6. Does the few-step advantage persist under consistent terminal noise conditions and an explicit timing protocol?

Response:

We thank the reviewer for this comment. Yes, the few-step advantage persists under consistent terminal noise conditions and an explicit timing protocol.

Noise adjustment at each step. We use a linear schedule that increases βt\beta _ t from 0 to 1 over TT steps:

βt=t−1T−1,t=1,…,T,αˉt=∏s=1t(1−βs).\beta _ t=\frac{t-1}{T-1},\quad t=1,\dots,T,\qquad \bar\alpha _ t=\prod _ {s=1}^{t}(1-\beta _ s).

The interval between consecutive steps depends on the total number of steps TT: the smaller TT is, the more noise is added at each step.

Consistent terminal noise. Since βT=1\beta _ T=1 for every TT, αˉT=0\bar\alpha _ T=0 holds exactly for all step budgets, and the terminal distribution is exactly the mixture prior in Eq. (42).

Separate training for each step budget. In Figure 3, the number of diffusion steps TT is a hyperparameter of each setting: for each value of TT, a separate model is trained with only TT changed and all other hyperparameters kept fixed. DP and SP use exactly the same schedule at every TT, and differ only in the prior (Sec. 5.4, lines 456--458). Therefore, the gap between the two curves comes from the learned prior itself.

Few-step advantage. Under the same conditions, DP converges within about 20 steps, whereas SP requires about 40 steps (Sec. 5.4, Fig. 3).

Timing protocol. In Table 6, GMD uses 20 denoising steps, and each baseline uses the step count beyond which its performance no longer improves significantly (DM = 40, CD = 30, HzRec = 30). All methods are timed with the same batch size of 1024. The timing includes codebook lookup (MDR retrieval for HzRec), all denoising steps and full ranking. On KF-FR, the per-item inference time of GMD is 1.45×10−31.45\times10^{-3} s, compared with 2.24×10−32.24\times10^{-3} s for HzRec, 2.30×10−32.30\times10^{-3} s for CD and 2.75×10−32.75\times10^{-3} s for DM. We have added these details to the revised paper.

Revision. New Appendix C.3 (Table 7), Sec. 5.4 and Appendix C.10 (Table 15), in blue in the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer LbYS(4)

Summary of revisions for Reviewer LbYS (changes are marked in blue in the revised PDF)

IssueLocationRevision
W1 & Q1Sec. 4.2.1–4.2.2, App. A, B, B.5Alg. 2 initialization and Eqs. (8)–(9) rewritten as mixture sampling; "aggregation" revised; Table 4
W2 & Q2Sec. 5.1, App. C.2FR/NR partition, information exclusion, full-catalog ranking
W3 & Q3Sec. 5.4, App. C.3, C.10Table 7 (matched schedule); separate model per TT; timing protocol; Table 15

Reviewer_dpa1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Conditional Gaussian Mixture Denoising (GMD) for cross-domain sequential recommendation (CDSR). It targets the limitation of prior diffusion-based CDSR methods (e.g., CD, DM, HzRec) that start denoising from a fixed, user-agnostic Gaussian N(0,I)\mathcal{N}(0,I). This starting point can be far from a user’s noisy domain-shared multi-interest (NDI), producing long denoising trajectories.GMD instead learns a user-specific NDI prior as a codebook-based Gaussian mixture. A domain-shared preference module encodes multi-domain sequences, indexes a shared codebook of Gaussian components via Wasserstein distance, and forms mixture weights; a pairwise proximity regularizer keeps components from becoming overly separated. The forward and reverse processes are then redefined so that trajectories start from (and terminate at) this mixture prior. Denoising is further conditioned on a domain-specific log-density difference between source- and target-domain features (JiT-style clean-data prediction is used). For users with no target-domain history the target term is zeroed while the shared codebook still carries cross-domain signal.Experiments on three Amazon pairs (Kitchen→Food, Electronics→Beauty, Book→Movie) under both few-interaction (FR) and no-interaction (NR) settings show consistent gains over strong CDSR/CDR baselines in HR@K and NDCG@K, together with substantially fewer diffusion steps to reach peak accuracy than a standard-Gaussian-prior ablation.

Critical strengths
  1. Clear, well-motivated technical contribution. Replacing a generic isotropic Gaussian with a learned, user-specific multi-modal prior and deriving the corresponding controlled forward/reverse processes (closed-form marginal, posterior, and sampling formulas) is a coherent and non-trivial extension of diffusion for recommendation. The pairwise proximity regularizer and log-density conditioning are sensible engineering responses to the multi-interest and sparse-target challenges the authors identify.

  2. Strong empirical evidence on the core claim. Tables 1–2 and the diffusion-step curves (Fig. 3) demonstrate both higher ranking metrics and markedly faster convergence (≈15–20 steps vs. ≈40 for the standard prior). Ablations (Table 3) isolate the contribution of the prior, the codebook, the regularizer, and the two masking strategies, and the NR results show that cold-start users still benefit.

Critical weaknesses
  1. Model complexity and design opacity. The pipeline combines a VQ-style Gaussian codebook, Wasserstein indexing, Gumbel-Softmax mixture selection, a custom pairwise proximity regularizer, a redesigned diffusion process with non-standard mean/variance coefficients, and a quadratic log-density condition. Many of these pieces are justified only empirically or by reference to other papers; limited ablation or sensitivity analysis is provided for the precise coefficient choices (e.g., 1−αt1-\sqrt{\alpha_t}) or the functional form of the proximity penalty. This makes the method hard to re-implement or diagnose.

  2. Heavy reliance on overlapping users and a shared codebook. The codebook is pre-trained on multi-domain interactions of overlapping users and then frozen. While the paper shows that cold-start (NR) users still improve, the absolute gains and the quality of the prior are likely to degrade when overlap ratios become smaller or domains are more heterogeneous. The paper does not quantify performance as a function of overlap ratio or explore non-overlapping regimes.

  3. Limited experimental scope. Evaluation is confined to three Amazon category pairs that are relatively close (Kitchen–Food, Electronics–Beauty, Book–Movie). No results are reported on more distant domain pairs, industrial-scale logs, or non-Amazon datasets. Hyper-parameter choices (codebook size scaling ζ\zeta, number of components NN, diffusion steps) are tuned per dataset; it is unclear how robust the method is under a single fixed configuration.

  4. Incomplete comparison and fairness concerns. Several recent or concurrent diffusion-based recommenders are missing or only briefly discussed. The strongest diffusion baselines (CD, HzRec) still start from a standard Gaussian; a stronger test would be to equip them with the same mixture prior or the same x0x_0-prediction objective and measure the residual gain. Inference-time and memory comparisons (Appendix C.5) are only briefly mentioned; wall-clock training cost relative to lighter CDSR models is not reported.

  5. Theoretical and analysis gaps. The paper supplies detailed derivations of the controlled diffusion marginal and posterior, yet offers almost no analysis of the approximation quality of the mixture prior, the effect of the proximity regularizer on the Wasserstein geometry of the codebook, or the bias introduced by Gumbel-Softmax temperature annealing. Failure cases (e.g., users whose interests are poorly captured by a small fixed number of components) are not examined, and the claim that “learning where denoising starts is as important as learning how to denoise” remains largely empirical.

Questions for authors

Please refer to the weaknesses above.

7 comments
AuthorsOpenReview ↗

Response to Reviewer dpa1(1)

W1 Model complexity and design opacity. The pipeline combines a VQ-style Gaussian codebook, Wasserstein indexing, Gumbel-Softmax mixture selection, a custom pairwise proximity regularizer, a redesigned diffusion process with non-standard mean/variance coefficients, and a quadratic log-density condition. Many of these pieces are justified only empirically or by reference to other papers; limited ablation or sensitivity analysis is provided for the precise coefficient choices (e.g., ζ\zeta) or the functional form of the proximity penalty. This makes the method hard to re-implement or diagnose.

Response:

We thank the reviewer for this thorough analysis.

Source of complexity. The computational complexity of GMD comes mainly from diffusion. Codebook retrieval, i.e., the VQ-style Gaussian codebook and Wasserstein indexing, only serves to construct the starting point of diffusion, i.e., the Gaussian mixture prior. Its cost is far lower than that of diffusion, yet it effectively improves diffusion efficiency: on KF-FR, GMD takes 4.05 min per epoch for training and 1.45×10−31.45\times10^{-3} s per item for inference, compared with 6.35 min and 2.24×10−32.24\times10^{-3} s for the diffusion-based HzRec (Table 6). The time of each stage is reported below (minutes per epoch):

StageKF-FREB-FRBM-FRKF-NREB-NRBM-NR
Pretraining: codebook lookup0.090.070.210.080.060.21
Pretraining total (GMD-D)0.970.732.810.950.712.78
Main training: codebook lookup0.090.070.220.080.070.19
Main training: full ranking0.180.150.450.180.130.40
Total training (GMD)4.053.1011.843.983.0211.65

Across all six settings, codebook lookup in both stages takes less than 5% of the total training time. Gumbel-Softmax selection is the sampling step of the mixture and takes only 0.34--0.38 ms per batch across the six settings.

Design transparency. Every component of GMD has solid theoretical support, and the design is not opaque. The redesigned forward kernel is derived from the telescoping structure and boundary conditions in Appendix B.2. Proposition 1 gives the closed-form marginal (Appendix B.1), and Appendix B.4 derives the posterior. The domain-specific condition is derived from the Bayes log-density ratio in Appendix B.3 (Eqs. (45)--(47)). The key components are therefore supported by both mathematical derivation and ablation experiments, not only by empirical evidence.

Ablation and sensitivity analysis. The coefficient choices, such as ζ\zeta, are already validated in the paper (Appendix C.2, Fig. 4(a), lines 1055--1072). The paper includes the following ablation and sensitivity analyses: component-wise ablations of the masking strategies (w/o M1, w/o M2), the pairwise proximity regularizer (w/o R), the Gaussian codebook (w/o G), the domain-specific condition (w/o CA, w/o CB, w/o C) and the mixture prior (w/o P) in Table 3; a comparison of prior types (Appendix C.3, Table 5); sensitivity analyses of the codebook size ζ\zeta and the number of components NN (Appendix C.2, Fig. 4); and the impact of the number of diffusion steps (Sec. 5.4, Fig. 3). We are glad to provide further analyses if the reviewer finds them helpful.

Revision. Appendix C.10: new paragraph Timing protocol and Table 15 (stage-wise cost; codebook lookup takes less than 5% of training time, and Gumbel-Softmax takes 0.34--0.38 ms per batch). Section, table and appendix numbers in this Revision paragraph refer to the revised PDF; changes are marked in blue there.

AuthorsOpenReview ↗

Response to Reviewer dpa1(2)

W2 Heavy reliance on overlapping users and a shared codebook. The codebook is pre-trained on multi-domain interactions of overlapping users and then frozen. While the paper shows that cold-start (NR) users still improve, the absolute gains and the quality of the prior are likely to degrade when overlap ratios become smaller or domains are more heterogeneous. The paper does not quantify performance as a function of overlap ratio or explore non-overlapping regimes.

Response:

We thank the reviewer for this insightful comment.

Role of the shared codebook. The shared codebook is the core design of GMD. It enables GMD to effectively transfer the cross-domain interests learned from overlapping users to all users, including those without target-domain history. Its usefulness is confirmed by the ablation in Table 3: removing the codebook (w/o G) reduces NDCG@10 on all six settings, e.g., from 0.2052 to 0.1846 on KF-FR and from 0.1692 to 0.1458 on EB-FR.

Performance under smaller overlap ratios. We agree that, in theory, the performance of all cross-domain methods degrades as the overlap ratio decreases, since overlapping users are the only bridge between the two domains. When the overlap ratio drops to 0, the two domains share neither users nor items, so the model can only rely on the limited data of a single domain. Following the reviewer's suggestion, we reduce the overlap ratio of the training set on EB from its original 19.64% to 15%, 10%, 5% and 0%, and compare GMD with the strongest baseline (I2VAE), the most recent diffusion-based baseline (HzRec) and the variant without the codebook (w/o G). Results on EB-FR (NDCG@10) are shown below.

Overlap ratio19.64%15%10%5%0%
I2VAE0.15090.10210.05880.02530.0154
HzRec0.14480.09680.05410.02280.0138
w/o G0.14580.11410.07790.03780.0176
GMD0.16920.13630.09680.05210.0213

As expected, all methods degrade as the overlap ratio decreases, and all of them drop to a very low level at 0%, which is consistent with the analysis above. GMD consistently performs best at every overlap ratio, and the codebook-based GMD outperforms w/o G throughout, showing that the shared codebook makes better use of limited overlapping users.

Non-overlapping regimes. The paper already covers this regime at test time. In the NR setting, cold-start test users have no target-domain interactions at all (Sec. 3, lines 140--141), so NR test users are not overlapping users. GMD still outperforms all baselines in this setting (Table 2).

Revision. New Appendix C.6 Robustness to the User Overlap Ratio with Table 11; Appendix E (Limitations) updated accordingly. Section, table and appendix numbers in this Revision paragraph refer to the revised PDF; changes are marked in blue there.

AuthorsOpenReview ↗

Response to Reviewer dpa1(3)

W3 Limited experimental scope. Evaluation is confined to three Amazon category pairs that are relatively close (Kitchen--Food, Electronics--Beauty, Book--Movie). No results are reported on more distant domain pairs, industrial-scale logs, or non-Amazon datasets. Hyper-parameter choices (codebook size scaling ζ\zeta, number of components NN, diffusion steps) are tuned per dataset; it is unclear how robust the method is under a single fixed configuration.

Response:

We thank the reviewer for pointing out this problem.

Datasets. Following existing CDSR baselines, we adopt the same three Amazon domain pairs as ABXI [1], with similar configurations (Appendix C.1). Considering the reviewer's concern, we further evaluate on a non-Amazon dataset, Douban (Book$\rightarrow$Music):

MethodDouban-FR HR@10Douban-FR NDCG@10Douban-NR HR@10Douban-NR NDCG@10
ABXI0.22130.12860.13050.0724
I2VAE0.24510.14520.14920.0831
HzRec0.24020.14180.14210.0793
GMD0.26870.15980.17380.0972

Hyperparameters. Different datasets have different sensitivities to the hyperparameters, such as the codebook size scaling ζ\zeta, the number of components NN and the number of diffusion steps. Following common practice in prior CDSR work, we select the best-performing values for each dataset. Nevertheless, the existing experiments in the paper show that GMD is robust under a single fixed configuration. In Fig. 4(a), performance on all six settings plateaus once ζ≥0.05\zeta\ge0.05, with little change from 0.05 to 0.09. In Fig. 4(b), performance changes little once N≥3N\ge3 on all datasets, with EB peaking at N=4N=4. In Fig. 3, performance converges within about 20 diffusion steps on all three datasets and stays stable up to 50 steps. Thus a single configuration, e.g., ζ=0.05\zeta=0.05, N=3N=3 and T=20T=20, lies in the stable region of every dataset. The shaded bands, which show the standard deviation over five random seeds, are also narrow, indicating low variance. We have added these analyses to the revised paper.

Revision. New Appendix C.7 Results on a Non-Amazon Dataset with Table 12 (Douban); new paragraph Hyperparameter robustness in Appendix C.1; Conclusion and Appendix E updated to mention Douban. Section, table and appendix numbers in this Revision paragraph refer to the revised PDF; changes are marked in blue there.

[1] Qingtian Bian, Marcus de Carvalho, Tieying Li, Jiaxing Xu, Hui Fang, and Yiping Ke. ABXI: Invariant interest adaptation for task-guided cross-domain sequential recommendation. In WWW, pp. 3183--3192, 2025.

Reviewer_VGH9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses sparse target-domain interactions in cross-domain sequential recommendation and proposes Conditional Gaussian Mixture Denoising (GMD). Masked pretraining learns a shared Gaussian codebook, from which multiple interest components are selected for each user. The model then runs component-wise diffusion trajectories, conditioned on the difference between target- and source-domain log densities, to reconstruct the target-item representation. Component predictions are combined through Gumbel-Softmax during training and a mixture-weighted average at inference.

Experiments cover three Amazon domain pairs, Kitchen-to-Food, Electronics-to-Beauty, and Book-to-Movie, under settings with few target-domain interactions (FR) or no target-domain history (NR). GMD is compared with nine baselines and achieves the best results on all reported metrics in Tables 1 and 2. The paper also includes component ablations, comparisons of prior types, codebook sensitivity analyses, and an efficiency study.

Critical strengths
  1. The gains are consistent across six sparse cross-domain settings.

    In Table 1, NDCG@10 reaches 0.2052 on KF-FR and 0.1692 on EB-FR, compared with 0.1845 and 0.1509 for the strongest respective baselines. In Table 2, KF-NR reaches 0.1287 versus 0.1047 for the strongest baseline. Improvements across different domain pairs and both target-history conditions make the approach worth investigating.

  2. The controls go beyond removing individual modules.

    Table 5 compares a standard Gaussian, a user-dependent single Gaussian, a single-codebook Gaussian, and a mixture without the codebook while retaining the other main design choices. These controls are more informative about the prior's contribution than a comparison against standard DDPM alone. Figure 3 and Table 6 also provide step-budget curves and measured runtimes, rather than treating the number of diffusion steps as sufficient evidence of efficiency.

Critical weaknesses
  1. The mixture prior is not clearly reconciled with the actual diffusion and aggregation procedure.

    Equation (4) and Eq. (68) in Appendix B.5 describe a Gaussian mixture in which a component is selected with probability πn\pi_n. Equation (8), however, constructs NN trajectories terminating at N(πnμn,πn2σn2I)\mathcal N(\pi_n\mu_n,\pi_n^2\sigma_n^2I), whose concatenation has a product distribution. Their predictions are then averaged at inference. Scaling component parameters by πn\pi_n is not equivalent to assigning that component mixture probability πn\pi_n, and a weighted average is not a discrete component sample. The appendix explains the parallel trajectories and soft selection, but does not resolve how these distributions correspond. This affects the central claim about preserving multimodal interests, rather than just the notation.

  2. The cold-start evaluation protocol is underspecified.

    Section 5.1 and Appendix C.1 provide dataset statistics, chronological ordering, and model hyperparameters, but do not specify how FR histories are truncated, how NR users are constructed, how training/validation/test data are split, or whether test ranking uses all items or sampled candidates. Appendix B.5 describes training negatives for InfoNCE, not the test ranking protocol. The users and time ranges available to codebook pretraining also need to be stated. These omissions make it difficult to reproduce the HR/NDCG results or determine whether all baselines face the same cold-start conditions. They are not, by themselves, evidence that leakage occurred.

  3. The distinction from the closest prior work is based on an inaccurate description.

    HorizonRec (HzRec) is included as a baseline, but Section 2 and Appendices C.1 and D characterize it as using a user-agnostic standard Gaussian. Its MDR module already retrieves user-dependent noise distributions from historical behavior, as described in Sections 4.4 and 4.5 of the original paper. Its public predict implementation also initializes reverse diffusion with samples from that distribution. The relevant distinction is therefore between a retrieval-based personalized Gaussian and a multi-component codebook prior. The comparison should retain the full MDR module and establish the value of the proposed codebook and conditioning mechanism against it. The contribution cannot rest on the premise that the closest method lacks a personalized distribution.

Questions for authors
  1. Probabilistic model: How does the discrete mixture in Eq. (68) correspond to the parallel trajectories with parameters scaled by πn\pi_n? Why are the mixture weights πn\pi_n used both to scale component distributions and to aggregate predictions? Please give a consistent joint distribution and sampling procedure. A control without component-parameter scaling would help separate mixture selection from changes in noise scale. Table 5 already includes single-Gaussian controls; a Gaussian matched to the mixture's mean and covariance would further help isolate the value of multimodality.

  2. Evaluation protocol: Please specify how FR/NR users and histories are constructed, the split unit and temporal boundaries, the use of validation data, the test candidate set, and the filtering of previously interacted items. Does codebook pretraining use only records available in the corresponding training split? How does each baseline handle missing target-domain history in NR?

  3. Closest baseline and contribution: Does the HorizonRec implementation retain MDR retrieval and its user-dependent noise? Does its retrieval database follow the same training-data boundaries as GMD? Against the full implementation, which new mechanisms account for GMD's gains?

  4. Few-step diffusion and efficiency: Please report the noise schedule {βt}\{\beta_t\} and the resulting αˉT\bar\alpha_T at different step budgets. The derivation assumes αˉT≈0\bar\alpha_T\approx0, while the default uses 20 steps. Is this assumption satisfied, and how are schedules controlled between DP and SP in Figure 3? Table 6 reports measured runtimes, but the baseline step counts, timing batch sizes, inclusion of retrieval/codebook lookup and full ranking, and accounting of total pretraining and main-training costs still need to be specified.

6 comments
AuthorsOpenReview ↗

Response to Reviewer VGH9 (1)

W1 The mixture prior is not clearly reconciled with the actual diffusion and aggregation procedure.

Response (1/2):

We thank the reviewer for this careful reading. Both the prediction and the sampling process rely on Gumbel-Softmax (GS) sampling rather than direct averaging.

Specifically, the model prediction based on the forward (noising) process in Eq. (12),

x^pre=∑n=1N[exp⁡(log⁡πnX+gn)⋅x^0,n]∑m=1Nexp⁡(log⁡πmX+gm),gn,gm∼Gumbel(0,1),\hat{x} _ {pre}=\frac{\sum _ {n=1}^{N}\left[\exp\left(\log\pi^{X} _ {n}+g _ n\right)\cdot\hat{x} _ {0,n}\right]}{\sum _ {m=1}^{N}\exp\left(\log\pi^{X} _ {m}+g _ m\right)},\qquad g _ n,g _ m\sim\mathrm{Gumbel}(0,1),

implements the sampling process of Eq. (8). The formulation of Eq. (8) in the paper,

q(xt∣xt−1):=N(αtxt−1+(1−αt)π1:NX⋅μ1:NGX, βt(π1:NX⋅σ1:NGX)2I),q(x _ t\mid x _ {t-1}):=\mathcal{N}\left(\sqrt{\alpha _ t} x _ {t-1}+(1-\sqrt{\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \beta _ t\left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}I\right),

may have been misleading. Here π1:NX⋅μ1:NGX\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N} and π1:NX⋅σ1:NGX\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} only denote a parallel structure that represents the Gaussian mixture; they are not computed as weighted sums. Only the single component obtained by sampling participates in the computation. To make this clear, we have rewritten Eq. (8) as

q(xt∣xt−1):=N(αtxt−1+Sampling((1−αt)π1:NX⋅μ1:NGX), Sampling(βt(π1:NX⋅σ1:NGX)2)I),q(x _ t\mid x _ {t-1}):=\mathcal{N}\Big(\sqrt{\alpha _ t} x _ {t-1}+\mathrm{Sampling}\big((1-\sqrt{\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N}\big),\ \mathrm{Sampling}\big(\beta _ t\left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}\big)I\Big),

where both Sampling(⋅)\mathrm{Sampling}(\cdot) operations refer to the same component selected by Gumbel-Softmax according to π1:NX\pi^{X} _ {1:N}, so the mean and variance always come from one sampled sub-distribution. Its expanded form (Eq. (9)) is

xt=αˉtx0+Sampling(N((1−αˉt)π1:NX⋅μ1:NGX, 1−αˉtπ1:NX⋅σ1:NGXϵ)),x _ t=\sqrt{\bar\alpha _ t} x _ 0+\mathrm{Sampling}\left(\mathcal{N}\left((1-\sqrt{\bar\alpha _ t}) \pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \sqrt{1-\bar\alpha _ t} \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right),

where Sampling(⋅)\mathrm{Sampling}(\cdot) denotes GS sampling of a mixture component, and ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I) is the Gaussian noise.

For the denoising sampling process (Eq. (17)), the starting point xTx _ T is drawn from the Gaussian mixture given in Appendix Eq. (42):

q(xT∣x0)≈N(π1:NX⋅μ1:NGX, (π1:NX⋅σ1:NGX)2I).q(x _ T\mid x _ 0)\approx\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \left(\pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N}\right)^{2}I\right).

This expression likewise denotes GS sampling of the distribution of xTx _ T. Both the noising and the denoising processes therefore use GS sampling of the Gaussian mixture.

The core properties of the diffusion process are fully preserved under this formulation. The timestep tt is uniformly sampled from {1,…,T}\lbrace 1,\dots,T \rbrace. When t=1t=1, αˉt≈1\bar\alpha _ t\approx 1 and thus xt≈x0x _ t\approx x _ 0. When t=Tt=T, αˉT=0\bar\alpha _ T=0 exactly (see Q4), and thus

xT=Sampling(N(π1:NX⋅μ1:NGX, π1:NX⋅σ1:NGXϵ)),x _ T=\mathrm{Sampling}\left(\mathcal{N}\left(\pi^{X} _ {1:N}\cdot\mu^{G _ X} _ {1:N},\ \pi^{X} _ {1:N}\cdot\sigma^{G _ X} _ {1:N} \epsilon\right)\right),

which is consistent with the starting point of the denoising process in Eq. (17).

Finally, all formulas involving the Gaussian mixture (e.g., Eqs. (4) and (8)) follow the sampling procedure given in Eq. (68) of Appendix B.5. We have clarified this in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer VGH9 (2)

W1 The mixture prior is not clearly reconciled with the actual diffusion and aggregation procedure.

Response (2/2):

On the per-component expressions in Appendix B.5. The per-component expressions in Appendix B.5 (lines 939--943), N(πnμn,(πnσn)2I)\mathcal{N}(\pi _ n\mu _ n,(\pi _ n\sigma _ n)^2I), are another way of writing the Gaussian mixture. By the linear transformation property of the Gaussian distribution, if xn∼N(μn,σn2I)x _ n\sim\mathcal{N}(\mu _ n,\sigma _ n^2I), then πnxn∼N(πnμn,(πnσn)2I)\pi _ n x _ n\sim\mathcal{N}(\pi _ n\mu _ n,(\pi _ n\sigma _ n)^2I); this expression therefore represents each component together with its mixing weight in a parallel form. When the weights are realized by the Gumbel-Softmax selection wnw _ n, which assigns 1 to the selected component and 0 to the others, we have ∑nwnxn=xn∗\sum _ n w _ n x _ n=x _ {n^{\ast}} with xn∗∼N(μn∗,σn∗2I)x _ {n^{\ast}}\sim\mathcal{N}(\mu _ {n^{\ast}},\sigma _ {n^{\ast}}^2I), i.e., exactly a sample from the Gaussian mixture in Eq. (68). We have revised these expressions together with Eq. (8) and Algorithm 2 to make this correspondence explicit.

On the term "aggregation" in Appendix B.5 and C.1. We apologize that the word "aggregation" in Appendix B.5 (lines 944--950) and C.1 (lines 1046--1048) is imprecise and may have caused confusion. It refers to the Gumbel-Softmax combination in Eq. (12) (Eq. (72) in Appendix B.5), whose weights π^n\hat\pi _ n are relaxed one-hot sampling weights rather than fixed averaging coefficients. It is therefore a differentiable relaxation of selecting one component, not a fixed average of the components, and we have revised the wording accordingly.

Setting gn=gm=0g _ n=g _ m=0 is applied only at inference. During training, Gumbel noise gn∼Gumbel(0,1)g _ n\sim\mathrm{Gumbel}(0,1) is always drawn, so the model is learned by sampling components according to the mixing weights π1:NX\pi^{X} _ {1:N}, exactly as in Eq. (68). At inference, the Gumbel noise is set to zero solely to avoid randomness in the test results and make them reproducible. This does not change what the model has learned; the core of GMD remains sampling from the Gaussian mixture.

To verify that this inference-only choice does not affect the results, we compare it with keeping the default Gumbel noise at inference, i.e., stochastic sampling, averaged over five runs (NDCG@10):

Inference settingKF-FRKF-NREB-FREB-NRBM-FRBM-NR
gn=gm=0g _ n=g _ m=0 (reported)0.20520.12870.16920.11010.16960.0897
Default Gumbel sampling (mean)0.20410.12780.16810.10930.16870.0889
Default Gumbel sampling (std)0.00210.00160.00190.00140.00180.0012

The two settings yield nearly identical results: the differences (0.0008--0.0011) are all within one standard deviation of stochastic sampling.

W2 & Q2 The cold-start evaluation protocol is underspecified. Evaluation protocol: Please specify how FR/NR users and histories are constructed, the split unit and temporal boundaries, the use of validation data, the test candidate set, and the filtering of previously interacted items. Does codebook pretraining use only records available in the corresponding training split? How does each baseline handle missing target-domain history in NR?

Response:

We thank the reviewer for this careful remark. The split unit is the user. Users are split into training, validation and test sets with an 8:1:1 ratio, and the three sets share no users. Training users contribute their full cross-domain sequences. For validation and test users, we sort target-domain interactions by time. Since FR test users have only a few target-domain interactions, we hold out only the last one as the prediction target and keep the rest as history, so each FR test user has 1--3 target-domain interactions in the history. FR users thus have a non-empty target-domain history (SuB≠∅S _ u^B \neq \varnothing). NR users have no target-domain history (SuB=∅S _ u^B = \varnothing) and use only their source-domain sequences as input. Validation users are used only for model selection. Test ranking covers all target-domain items, and each user's history interactions are filtered out.

The InfoNCE negatives in Appendix B.5 are used only in training. The codebook is pretrained only on training users' sequences (see Appendix C.1 for the two-stage training) and then frozen, so it never sees interactions of validation or test users. All baselines use the same user split and ranking protocol. For NR users, ASR and ABXI take the source-domain sequence as input, HzRec retrieves with MDR from the source-domain history, DM and CD follow their original cold-start protocols, and single-domain baselines (VSAN, CVAE) are run on the merged sequence. We have added these details to Sec. 5.1 and the appendix of the revised paper.

Revision. Sec. 5.1 and new Appendix C.2 (Evaluation Protocol), in blue in the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer VGH9 (3)

W3 & Q3 The distinction from the closest prior work is based on an inaccurate description. Closest baseline and contribution: Does the HorizonRec implementation retain MDR retrieval and its user-dependent noise? Does its retrieval database follow the same training-data boundaries as GMD? Against the full implementation, which new mechanisms account for GMD's gains?

Response:

We thank the reviewer for this careful reading. Our original description is that HzRec starts "denoising from a user-agnostic Gaussian prior"; we did not describe it as a user-agnostic standard Gaussian. These two statements are fundamentally different. Specifically, our statement concerns the prior of the diffusion model, i.e., the initial distribution assumed by the DDPM derivation. It does not concern the noise injected during diffusion, which in HzRec does carry user-specific features.

HzRec fully follows the DDPM derivation under the N(0,I)\mathcal{N}(0,I) assumption. Its Eq. (12),

hˉtD=αˉt hD+1−αˉt zˉD,\bar h^D _ t=\sqrt{\bar\alpha _ t}\ h^D+\sqrt{1-\bar\alpha _ t}\ \bar z^D,

is the DDPM closed form with the noise ϵ\epsilon replaced by the retrieved sample zˉD\bar z^D. HzRec does not redefine the DDPM prior. It injects user-specific sampled features into the noise rather than changing the initial prior. If the prior were N(μD,(σD)2I)\mathcal{N}(\mu^D,(\sigma^D)^2 I), the forward kernel would need to be redefined as

q(hˉtD∣hˉt−1D)=N(αt hˉt−1D+(1−αt) μD, βt(σD)2I),q(\bar h^D _ t\mid \bar h^D _ {t-1})=\mathcal{N}\big(\sqrt{\alpha _ t}\ \bar h^D _ {t-1}+(1-\sqrt{\alpha _ t})\ \mu^D,\ \beta _ t(\sigma^D)^2 I\big),

with the closed form

hˉtD=αˉt hD+(1−αˉt) μD+1−αˉt σDϵ,\bar h^D _ t=\sqrt{\bar\alpha _ t}\ h^D+(1-\sqrt{\bar\alpha _ t})\ \mu^D+\sqrt{1-\bar\alpha _ t}\ \sigma^D\epsilon,

and the reverse process re-derived accordingly. HzRec does not do this, whereas GMD redefines the forward kernel at the prior level (Eq. (8)). We agree that HzRec's injected noise is user-dependent, and we have clarified this distinction in Sec. 2, Sec. 5.2, Appendix C.2 and Appendix D.

Motivation of the Gaussian mixture prior. User interests in cold-start cross-domain recommendation are inherently multimodal. A single Gaussian, even a user-dependent one, is unimodal and can only represent one interest center. GMD therefore defines the prior itself as a codebook-based Gaussian mixture (Eq. (4)), where each component captures one interest mode and the mixing coefficients determine how likely each mode is. This is the core difference from HzRec: a multi-component prior defined at the prior level, rather than a single retrieved noise injected into the standard DDPM process. The matched single Gaussian control in Q1.3 (E2) supports this motivation: with the same mean and covariance, the Gaussian mixture still outperforms a single Gaussian.

Implementation and gains. We use the public HzRec code with MDR retrieval and its user-dependent noise fully enabled, and its retrieval database is built only from the same training users as GMD. Against this full implementation, GMD improves NDCG@10 from 0.1845 to 0.2052 on KF-FR and from 0.1033 to 0.1287 on KF-NR (Tables 1--2). Table 3 attributes these gains to two mechanisms that HzRec does not have: the codebook-based Gaussian mixture prior (w/o G: $-$0.8 to $-$2.7 points) and the domain-specific log-density condition (w/o C: $-$1.8 to $-$6.2 points).

Q1.1 Probabilistic model: How does the discrete mixture in Eq. (68) correspond to the parallel trajectories with parameters scaled by π\pi?

Response:

We thank the reviewer for this question. π1:NX\pi^X _ {1:N} are the mixing coefficients of the standard Gaussian mixture definition. Eq. (4),

NXMix={πnX⋅NnGX}n=1N,\mathcal{N}^{Mix} _ X=\left\lbrace \pi^X _ n\cdot\mathcal{N}^{G _ X} _ n\right\rbrace _ {n=1}^{N},

denotes a Gaussian mixture whose nn-th component NnGX=N(μnGX,(σnGX)2I)\mathcal{N}^{G _ X} _ n=\mathcal{N}\big(\mu^{G _ X} _ n,(\sigma^{G _ X} _ n)^2 I\big) has mixing coefficient πnX\pi^X _ n, i.e.,

p(x)=∑n=1NπnX NnGX(x),πnX≥0,  ∑n=1NπnX=1.p(x)=\sum _ {n=1}^{N}\pi^X _ n\ \mathcal{N}^{G _ X} _ n(x),\qquad \pi^X _ n\ge 0,\ \ \sum _ {n=1}^{N}\pi^X _ n=1 .

Here πnX⋅NnGX\pi^X _ n\cdot\mathcal{N}^{G _ X} _ n denotes a component together with its mixing coefficient; its correspondence to mixture sampling is explained in W1. Eq. (68) is the standard sampling procedure of this mixture:

n∗∼Categorical(π1X,…,πNX),x∼Nn∗GX.n^{\ast}\sim\mathrm{Categorical}(\pi^X _ 1,\dots,\pi^X _ N),\qquad x\sim\mathcal{N}^{G _ X} _ {n^{\ast}} .

Since categorical sampling of n∗n^{\ast} is not differentiable, we approximate it with the commonly used Gumbel-Softmax (GS) trick (Appendix B.5). The parallel trajectories are the GS relaxation of Eq. (68). Instead of a hard, non-differentiable choice of n∗n^{\ast}, GS assigns the components a relaxed one-hot weight, which approximates selecting a single component in a differentiable way. Thus the parallel trajectories follow the sampling semantics of Eq. (68) rather than a scaled or averaged distribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.