On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders

Thanks for your constructive suggestions. We have detailed our responses point-by-point to address your concerns.

Weakness 1:

The novelty of this paper is questionable.

Thanks for your question. However, we respectfully argue that our primary contribution is not the invention of these base concepts, but rather the novel task, technologies designed for recommenders, and theoretical analysis.

For tasks: We are the first to identify and address the specific efficiency-effectiveness trade-off in diffusion-based recommenders. This problem arises from the discretization errors inherent in the multi-step generation process.

For technologies: Our methods are not direct applications of existing methods but are novelly adapted for the recommendation context.

• Novelty in TCR for Recommendation: Unlike standard Consistency Models which operate on deterministic ODEs, our TCR is designed for the stochastic SDE/DDPM framework, which is crucial for the stochastic nature of user behavior. Furthermore, TCR loss is trained jointly with the diffusion reconstruction loss $\mathcal{L}_{diff}$, circumventing the need for a trained teacher diffusion model, which is often impractical in the recommendation domain.

• Novelty in APA for Diffusion Alignment: Similarly, our APA is more than a simple application of DPO in diffusion models. Its novelty lies in the introduction of the adaptive coefficient, $\lambda_\beta(t,d)$, which dynamically modulates the optimization strength based on both the timestep (noise level) and preference pair similarity. This can avoid diffusion recommenders' overfitting to ambiguous preferences or high noise levels.

For theoretical analysis: Our theoretical analysis provides a rigorous foundation for our framework. This includes the analysis of the bounded error of our one-step generation (Theorem 1) and the impact of our adaptive alignment coefficient (Theorem 2), distinguishing our work from a simple application of existing concepts.

Weakness 2:

Analyzing the impact of the scale of $\lambda_{base}$ without the time step and similarity terms.

Thanks for your comment. We have analyzed the sensitivity of TA-Rec to the $\lambda_\beta$ ($\lambda_{base}$) without the time step and similarity terms in Figure 4 of our paper. The methods of fix_one and fix_mul denote fine-tuning TA-Rec with fixed $\lambda_\beta$ without adapting to timestep or similarity in one or multiple steps-generation. The curves of fix-one and fix-mul fluctuate as $\lambda_\beta$ ($\lambda_{base}$) changes, highlighting the significance of $\lambda_{\beta}$ in performance optimization. The curves of Ours_one and Ours_mul, which consider timestep and similarity terms, lie above those of fix_one and fix_mul, validating the superiority of our design of adaptive coefficient $\lambda_\beta(t,d)$ in aligning user preference.

Weakness 3:

The premise of Theorem 1, which assumes the pretraining loss as zero, is highly unrealistic.

Thanks for your comment. We agree that the premise of zero pretraining loss in the original Theorem 1 is an idealization for the sake of clarity, and is unattainable in practice. To address this and enhance the theorem's practical applicability, we have reformulated the error bound to depend on the non-zero pretraining loss:

||f_\theta(\mathbf{x}\_{s/\Delta t}, \mathbf{g}, s/\Delta t) - \mathbf{x}||\_2 &\le ||f\_\theta(\mathbf{x}\_{s/\Delta t}, \mathbf{g}, s/\Delta t) - f\_\theta(\mathbf{x}\_{t},\mathbf{g},{t})||\_2 + ||f\_\theta(\mathbf{x}\_{t},\mathbf{g},{t}) - \mathbf{x}||\_2\\\\ & \leq ||f\_\theta(\mathbf{x}\_{s/\Delta t}, \mathbf{g}, s/\Delta t) - f\_\theta(\mathbf{x}\_{t},\mathbf{g},{t})||\_2+\delta. \end{aligned}

The second term $||f\_\theta(\mathbf{x}\_{t},\mathbf{g},{t}) - \mathbf{x}||\_2$ represents denoising model's reconstruction error. This error is inherent to denoising model, regardless of whether a multi-step or our accelerated one-step inference is used. Instead of assuming it is zero, we now bound it by the pretraining loss $\mathcal{L}\_{pre}$. We can have $||f\_\theta(\mathbf{x}\_{t}, \mathbf{g}, t) - \mathbf{x}||\_2<\delta$. Since we minimize the $L\_{pre}$ loss, the error of one-step generation is still bounded.

Questions:

1. The number of reverse steps used by other methods (e.g., DiffRec, DiffuRec) in Table 1?

The reverse steps for DiffuRec and DiffRec are chosen in [10,100,500,1000]. They are configured as [100,500,200] for DiffuRec and [100,500,100] for DiffRec, respectively, in the Yoochoose, KuaiRec, and Zhihu datasets.

2. The reported training time for TA-Rec

The reported training times in Table 3 reflect ​the time cost​​ for each epoch in the independent phase (pretraining/fine-tuning), not the cumulative time for the two stages. The per-epoch training time for each stage is comparable to baselines (e.g., SASRec), as the main computational cost in these methods comes from the Transformer encoder. The extra computation required in the pre-training stage is negligible, as it only involves a second pass through the lightweight MLP denoising model.

3. Explanation of the design of the frozen Transformer model during the fine-tuning stage.

For stability: Freezing the guidance encoder provides a stable target for the denoising model to align with during preference optimization, reducing fluctuations in performance. For efficiency: Fine-tuning only the smaller denoising model (MLP) is significantly faster and less computationally expensive.

4. Additional analysis to assess the diversity of the recommended items.

We recognize that embedding space visualization techniques like t-SNE are ideal for illustrating the diversity of generated items. However, since we are unable to submit images during the rebuttal, we conducted additional experiments using the coverage@20 metric to analyze the diversity of our recommendations, which measures the proportion of unique items recommended across multiple users. The experimental results presented in Table R1 indicate that both TA-Rec and DreamRec, which leverage diffusion models for generating recommendations, exhibit greater diversity compared to traditional recommenders like SASRec. This underscores the advantages of diffusion models in achieving diverse recommendations through stochastic noise. Furthermore, the recommendations from TA-Rec demonstrate comparable diversity to those from DreamRec, suggesting that the inherent diversity benefits of diffusion models are preserved even with a one-step reverse process. This is attributed to the diffusion mechanism of random noise injection at each step, along with the noise-to-item mapping in TA-Rec.

Table R1: Experiments on the diversity of diffusion recommenders.

5. Typos: the second-best baseline in Table 1 and typo "training" in Table 3

Thanks again for your thorough review and feedback on our paper. We will correct these typos in the paper and carefully examine the context to improve our clarity.

We sincerely hope that our response has adequately addressed your concerns. If so, we would greatly appreciate your consideration in raising the score. If there are any remaining concerns, please let us know, and we will continue to actively address your comments and improve our work.

Thanks for your constructive suggestions. We have detailed our responses point-by-point to address your concerns.

Weakness 1

Computational cost of executing the denoising model twice per training iteration for adjacent steps due to TCR loss.

Training Complexity

• Previous Diffusion Recommender (e.g., DreamRec): The reconstruction loss $\mathcal{L}_{diff}$ requires guidance encoding from the Transformer and a denoising process through the MLP denoising model. The training complexity is $O (N \cdot L^2 \cdot D + N \cdot D^2)$.
• TA-Rec: The ​​TCR loss​​ adds one extra MLP forward pass for an adjacent timestep. The training complexity of TA-Rec is $O(N \cdot L^2 \cdot D +2* N \cdot D^2)$, which is comparable to previous diffusion recommender since the complexity of MLP is negligible vs. Transformer.

Table 3 in our paper provides a comparative analysis of training time costs, demonstrating the efficiency of TA-Rec. Importantly, the integration of TCR loss during training facilitates one-step generation during inference, thereby significantly enhancing the efficiency of diffusion-based generative recommendation in real-world deployment.

Weakness 2

Experiments on more diverse datasets.

Here, we conduct experiments to validate TA-rec on more diverse datasets (Steam, Beauty, and Toys), varying in sizes and domains. The statistics of these datasets are shown in Table R1. Experimental results are presented in Table R2.

Table R1: Statistics of the Steam, Beauty, and Toys datasets.

Table R2: Overall performance of different methods for the sequential recommendation on diverse datasets.

Our method consistently outperforms various baselines on larger dataset (Steam) and diverse datasets (Amazon-beauty and Amazon-toys), further highlighting the effectiveness of TA-Rec.

We sincerely hope that our response has adequately addressed your concerns. If so, we would greatly appreciate your consideration in raising the score. If there are any remaining concerns, please let us know, and we will continue to actively address your comments and improve our work.

Hi reviewer dntH, This is a gentle reminder that the authors added their response on your question. Can you provide your feedback on their response? Please keep in mind that the deadline of August 6th approaching, and your additional timely feedback greatly enhance further discussions if needed.

Thanks, Area Chair.

Dear Reviewer dntH,

Thank you once again for your constructive comments. We have thoroughly addressed your concerns by incorporating new datasets (Steam, Beauty, and Toys) and clarifying the computational costs associated with training during the pretraining stage.

We look forward to further discussion with you and would greatly appreciate your positive feedback on our rebuttal.

Best regards,

The Authors

Thanks for your constructive suggestions. We have detailed our responses point-by-point to address your concerns.

Weakness 1 and Question 1:

Analysis of the diversity associated with diffusion-based models

For diffusion models, the mechanism of adding random noise to target items and generating "oracle items" from stochastic noise inherently enhances the diversity of recommendation results. Our model, TA-Rec, preserves this fundamental mechanism and accelerates generation by smoothing the denoising function. The TCR loss is designed to ensure the discretization error remains bounded during this accelerated generation. It does not eliminate the model's dependence on the random noise, so the potential for diverse generation is maintained in TA-Rec.

To validate the diversity of diffusion-based recommender and our TA-Rec, we conduct additional experiments on the coverage metric, which measures the proportion of unique items recommended across multiple users The experimental results presented in Table R1 indicate that both TA-Rec and DreamRec, which leverage diffusion models for generating recommendations, exhibit greater diversity compared to traditional recommenders like SASRec. This underscores the advantages of diffusion models in achieving diverse recommendations. Furthermore, the recommendations from TA-Rec show comparable diversity to those from DreamRec, suggesting that the inherent diversity benefits of diffusion models are maintained even with the faster inference speed.

Table R1: Experiments on the diversity of diffusion recommenders.

Weakness 2 and Question 2:

Removing inter-step differences to enable 1-step generation raises the question of whether using a diffusion model is necessary.

(1) Diffusion is a superior training mechanism: The core mechanism of diffusion models involves adding random noise to items and then training a model to generate "oracle items" from stochastic noise. We consider this a powerful training mechanism that enriches item distribution, facilitates the learning of item representations, and generates high-quality items to recommend. While effective, the multi-step reverse process often leads to significant inference overhead. Our proposed model, TA-Rec, accelerates the reverse process to a single denoising step with TCR while adhering to the fundamental diffusion mechanism. The value of the diffusion mechanism is empirically demonstrated by comparing TA-Rec to a strong baseline like SASRec (directly predicts next items without diffusion processing). As shown in Table 1 of our paper, TA-Rec significantly outperforms SASRec, even though both utilize similar core modules (e.g., Transformer, MLP). This performance gap highlights the necessity of the diffusion process for learning higher-quality representations and generating oracle items.

(2) Whether the noise and generation are indistinguishable in TA-Rec: The one-step generation of TA-Rec is achieved by smoothing the denoising function $f_\theta$ via Temporal Consistency Regularization (TCR), which enforces $f_\theta(x_t,t,g)\approx f_\theta(x_{t-1},t-1,g)$ during the noise-to-output mapping. The TCR does not eliminate inter-step differences between $x_t$, $x_{t-1}$, and $x_0$, but strategically smooths the $f_\theta$ to ensure consistent denoising output $x_0$ across different steps $t$. This aligns with the concepts of "Consistency Models".

(3) Why our one-step generation outperforms multi-step samplers DDPM/DDIM: DDPM/DDIM approximate a continuous reverse-SDE process with finite steps $T$ (typically 1000). This introduces the truncation errors that increase as T decreases (as discussed in Definition 1 of our paper). Even with $T=1000$, DDPM/DDIM converges to a ​​lossy approximation​​ of the true data distribution $P(x_0)$. For TA-Rec, the Temporal Consistency Regularization (TCR) enforces $f_\theta(x_t,t,g) - f_\theta(x_{t-1},t-1,g)\leq \delta$ to smooth the denoising trajectory. By maintaining output consistency across all adjacent steps, TCR ensures that $f_\theta(x_T, T,g)\approx f_\theta(x_t,t,g) \approx x_0$. Thus, w/ APA with one-step generation ​​implicitly integrates the infinite-step SDE (infinite $T$)​​, alleviating truncation errors and achieving better performance.

Limitation:

Discussion of potential negative societal impacts.

Over-personalization can lead to filter bubbles or echo chambers. Immediate feedback mechanisms may encourage users to engage compulsively, which could negatively impact their mental health.

We sincerely hope that our response has adequately addressed your concerns. If so, we would greatly appreciate your consideration in raising the score. If there are any remaining concerns, please let us know, and we will continue to actively address your comments and improve our work.

Hi reviewer J9pE, This is a gentle reminder that the authors added their response on your question. Can you provide your feedback on their response? Please keep in mind that the deadline of August 6th approaching, and your additional timely feedback greatly enhance further discussions if needed.

Thanks, Area Chair.

Dear Reviewer J9pE,

Thank you again for your constructive comments. We have thoroughly addressed your concerns by analyzing the diversity of diffusion models and our one-step TA-Rec with additional experiments, and clarifying the advantage and necessity of diffusion training mechanisms.

We look forward to further discussion with you and would greatly appreciate your positive feedback on our rebuttal.

Best regards,

The Authors

Thanks for your constructive suggestions. We have detailed our responses point-by-point to address your concerns.

Weakness 1

Experiments on more diverse datasets.

Here, we conduct experiments to validate TA-rec on more diverse datasets (Steam, Beauty, and Toys), varying in sizes and domains. The statistics of these datasets are shown in Table R1. Experimental results are presented in Table R2.

Table R1: Statistics of the Steam, Beauty, and Toys datasets.

Table R2: Overall performance of different methods for the sequential recommendation on diverse datasets.

Our method consistently outperforms various baselines on larger dataset (Steam) and diverse datasets (Amazon-beauty and Amazon-toys), further highlighting the effectiveness of TA-Rec.

Weakness 2

Theoretical computational costs of TA-Rec.

Let $N$ be the batch size, $L$ be the sequence length, and $D$ be the embedding dimension. The primary components of our model are the Transformer encoder and the MLP-based denoising model.

Inference Complexity

The primary contribution of TA-Rec is the significant reduction in inference complexity compared to prior diffusion models.

• Traditional Recommender (with Transformer and MLP in model design): The complexity of self-attention in Transformer is $O(N \cdot L^2 \cdot D )$, and the complexity of MLP is $O(N \cdot D^2)$, so the total inference complexity is $O(N \cdot L^2 \cdot D + N \cdot D^2)$.

• Previous Diffusion Recommender (e.g., DreamRec): The complexity is dependent on the number of denoising steps ($T$), which is typically large (~1000).

  • Complexity Formula: $O(T\cdot(N \cdot L^2 \cdot D + \cdot N \cdot D^2))$

• TA-Rec: By enabling one-step generation ($T=1$), the inference complexity is no longer dependent on a large number of steps and becomes comparable to traditional Transformer-based models.

  • Complexity Formula: $O(N \cdot L^2 \cdot D + N \cdot D^2)$

Thus, the inference complexity of TA-Rec is comparable to traditional transformer-based recommenders (e.g., SASRec).

Training Complexity:

• Traditional Recommender (e.g., SASRec): Due to Transformer and MLP modules, the training complexity is $O(N \cdot L^2 \cdot D + N \cdot D^2)$.

• Previous Diffusion Recommender (e.g., DreamRec): The reconstruction loss $\mathcal{L}_{diff}$ requires guidance encoding from the Transformer and a denoising process through the MLP denoising model. The training complexity is $O( N \cdot L^2 \cdot D + N \cdot D^2)$.

• TA-Rec: The ​​TCR loss​​ adds one extra MLP forward pass for an adjacent timestep. The training complexity of TA-Rec is $O(N \cdot L^2 \cdot D +2* N \cdot D^2)$, which is comparable to traditional recommender since the complexity of MLP is negligible vs. Transformer.

The time cost comparison in Table 3 of our paper validates the efficiency of TA-Rec for real-world deployment.

Weakness 3 & Question 2

Implementation details and ​sensitivity analysis​ for the negative sampling strategy for APA.

• As presented in Sec 4.4 of our paper, we randomly sample one item from the item corpus that the user has not interacted with as the negative to construct a preference pair for each user. Since our main contribution is the adaptive alignment strength which can mitigate overfitting to ambiguous preferences and noise, the random sampling strategy is employed for simplicity.

• To conduct ​sensitivity analysis on the negative sampling strategy, we perform additional experiments to explore different negative sampling strategies in Table R3. The ours-hard represents selecting hard negatives based on high cosine similarity to the positive item. The ours-popular denotes selecting negative items according to item popularity. The slightly superior performance of ours-popular and ours-hard over ours suggests that adopting different negative sampling strategies can be further considered.

Table R3: Experiments on hard negative sampling strategy.

Questions

Heuristic guidance for selecting $\lambda_{base}$.

• In DPO loss, the hyperparameter $\lambda_\beta$ typically ranges from 0.01 to 0.9. As shown in Figure 3 of our paper, the alignment strength $\lambda_\beta=1/2$ can achieve optimal performance across diverse datasets. As a result, we set $\lambda_{base}=1/2$ and vary $\lambda_\beta(t,d)$ from 0 to 1 with the adaptive mechanism.

Limited improvement of APA over other diffusion recommenders in Table 7.

• After further calculations and validations, the performance improvement of APA over other diffusion recommenders ranges from 1% to 4%. All results are the average of 5 independent runs, and these improvements are statistically significant (p-value < 0.05). The improvements observed in TCR and DDIM exceed those in DDPM, indicating that APA may be more effective in enhancing the accuracy of methods designed to accelerate generation.

We sincerely hope that our response has adequately addressed your concerns. If so, we would greatly appreciate your consideration in raising the score. If there are any remaining concerns, please let us know, and we will continue to actively address your comments and improve our work.

Hi reviewer NQjp, This is a gentle reminder that the authors added their response on your question. Can you provide your feedback on their response? Please keep in mind that the deadline of August 6th approaching, and your additional timely feedback greatly enhance further discussions if needed.

Thanks, Area Chair.

Dear Reviewer NQjp,

Thank you again for your constructive comments. We have thoroughly addressed your concerns by incorporating new datasets (steam, beauty, and toys), providing theoretical computational costs in addition to experimental time costs in our paper, exploring additional negative sampling strategies for APA, and clarifying the selection guidance for $\lambda_base$.

We look forward to further discussion with you and would greatly appreciate your positive feedback on our rebuttal.

Best regards,

The Authors

Thanks for your constructive suggestions. We have detailed our responses point-by-point to address your concerns.

Weakness 1:

The computational resources for the dot product between $x_0$ and candidate items

• Calculating similarity between $x_0$ and item corpus then recommending the top-K is a standard final step in recommendation systems (e.g., SASRec). It is not unique to our framework. Crucially, the computational cost of dot products is negligible compared to model inference as it involves only simple matrix multiplication after embeddings are generated.

• Our key innovation lies in accelerating the generation of embedding ($x_0$). As shown in Table 3, TA-Rec's inference time is lower than other diffusion recommenders and comparable to traditional recommenders like SASRec.

Weakness 2

Random negative sampling of APA may fail to capture hard negative samples for preference alignment.

• The negative sampling strategy and our primary innovation of the adaptive alignment coefficient $\lambda_\beta(t,d)$ are independent effective solutions. The coefficient $\lambda_\beta(t,d)$ dynamically adjusts the optimization strength to mitigate overfitting to ambiguous preferences and noise. The effectiveness of APA is validated in the Sec 5.3 of our paper. The random negative sampling we employed is indeed a simplification, and we have acknowledged it as a limitation in Appendix E.3 of our paper.

• We consider your suggestion a valuable direction for future research, and we conduct additional experiments to explore hard negative sampling strategy in Table R1. The w/ hard samples 5 random negatives, selects the most similar to the target via cosine similarity as a hard negative, and conducts preference alignment without adaptive coefficients. ours-hard incorporates APA with hard negative sampling for adaptive alignment. The superiority of w/ hard over w/o APA and ours-hard over ours demonstrates the potential of hard negative sampling.

Table R1: Experiments on hard negative sampling strategy.

Weakness 3

Performance in data-sparse or diverse scenarios

• While cold-start and diversity are important challenges in recommendation systems, they are orthogonal to our core contribution: ​​resolving the fundamental efficiency-effectiveness trade-off in diffusion-based recommenders​​ deployment.

• Extending our approach to more challenging scenarios, such as cold-start problems or diversity, could be a focus for future research. Here, we construct datasets in data-sparse setting for exploration, by removing 30% interactions. The experimental results are presented in Table R2. The superior performance of our method compared to the baseline indicates its robustness in data-sparse scenarios. Additionally, to validate the diversity of our TA-Rec, we conduct additional experiments on the coverage metrics, which measure the proportion of unique items recommended across multiple users. The experimental results presented in Table R3 indicate that both TA-Rec and DreamRec, which leverage diffusion models for generating recommendations, exhibit greater diversity compared to traditional recommenders like SASRec. This underscores the advantages of TA-Rec in achieving diverse recommendations.

Table R2: Experiments on data-sparse scenarios.

Table R3: Experiments on the diversity of diffusion recommenders.

Questions:

Online Deployment and Inference Time & Lipschitz Continuity Assumptions & Online Learning

(1) By achieving one-step generation with TCR in the pretraining stage, TA-Rec has similar inference time to traditional recommenders (e.g., SASRec), as shown in Table 3 of our paper. Thus, our method can be effectively deployed in production environments. (2) The Lipschitz continuity assumption is not an inherent assumption applied to data, but rather a desirable property that we realize on the denoising function $f_\theta$ through TCR objective. By constraining TCR loss, we enforce local smoothness between adjacent denoising steps and ensure $f_\theta$ satisfies the Lipschitz continuity assumption. Then, $f_\theta$ can achieve one-step generation with bounded error, mapping noise directly to oracle items as shown in Theorem 1. (3) Our algorithm primarily focuses on offline training setting, which is consistent with comparable methods like DreamRec and PreferDiff. We agree that it can support real-time incremental and streaming learning, which is a potential direction for future work.

We sincerely hope that our response has adequately addressed your concerns. If so, we would greatly appreciate your consideration in raising the score. If there are any remaining concerns, please let us know, and we will continue to actively address your comments and improve our work.

Hi reviewer BncC, This is a gentle reminder that the authors added their response on your question. Can you provide your feedback on their response? Please keep in mind that the deadline of August 6th approaching, and your additional timely feedback greatly enhance further discussions if needed.

Thanks, Area Chair.

Dear Reviewer BncC,

Thank you again for your constructive comments. We have thoroughly addressed your concerns by further investigating a hard negative sampling strategy, evaluating the performance of our TA-Rec in data-sparse and diverse scenarios through additional experiments, and clarifying its efficiency and potential for online learning.

We look forward to further discussion with you and would greatly appreciate your positive feedback on our rebuttal.

Best regards,

The Authors