## Reviewer 1 ### Summary and contributions This paper aims to address the data sparsity and negative sampling issue in graph-based CF. To this end, it proposes SCONE, a stochastic sampling framework that generates contrastive views and hard negative samples using score-based generative models. Experiments on six real-world datasets demonstrate the effectiveness of the proposed method. ### Strong points The paper is well-organized and easy to follow, with clear explanations of both the methodology and experiments. This paper conducted extensive experiments on six datasets, as well as a comprehensive sensitivity analysis. The authors provide the code for SCONE, which enhances transparency and reproducibility of the results. ### Weak points The proposed method primarily leverages stochastic generation in the context of contrastive learning, which limits its novelty since both techniques are already well-explored. While the paper aims to address data sparsity in graph-based collaborative filtering (CF), SCONE appears to be a general method without specific design tailored for graph structures The baselines used for comparison are limited and somewhat outdated. Only two baselines are provided for contrastive learning and two for negative sampling, with some dating back to 2013 (DNS). This weakens the impact of the experimental results, making them less convincing in today's context. The results are only reported for top-20 recommendations, without considering other commonly used metrics, such as top-5 or top-10. Including these could give a fuller picture of SCONE's effectiveness across different thresholds. **Detailed evaluation:** Please check strengths and weaknesses **Overall rating:** Weak Reject (Probable reject, unless convinced otherwise) **Reviewer confidence:** Knowledgeable: I have read papers on this topic. ## Reviewer 2 ### Summary and contributions This paper proposes a framework called SCONE, which is a novel framework for addressing data sparsity and negative sampling challenges in graph-based collaborative filtering recommender systems. The key contributins is mainly about the development of an unified stochastic sampling approach which is able to generate both contrastive views for self-supervised learning and hard negative samples. ### Strong points 1. This paper conducted very comprehensive experiments on 6 benchmark datasets , and the results demonstrated that SCONE outperforms SOTA systems, especially on sparse user and items. 2. The SOCNE framework is novel. It combines contrastive learning and hard negative sampling into a single stochastic sampling approach to address key challenges effectively. 3. This paper is well-written, it includes detailed and comprehensive information to help readers get deep into it. ### Weak points 1. The unified framework is novel, but seems like there is not much novelty in every individual compoennt in the framework (e.g. contrastive learning and negative sampling) 2. The computation cost might still be high--- though in table 5 there is a comparison of different approaches' running time --- it still concerns me how the latency or computation cost would be if we apply the approach to real-world large volume traffic 3. The system can be complex --- so I may have concerns on whether it is easy to reproduce the work. ### Detailed evaluation Overall I think this paper proposes an insightful framework to address key challenges in collaborative filtering, with comprehensive experiments results. While there are some concerns like the complexity of the system, the overall contribution appears to be valuable to the community. **Overall rating:** Weak Accept (Probable accept, unless convinced otherwise) **Reviewer confidence:** Familiar: I have passing knowledge on this topic. ## Reviewer 3 ### Summary and contributions This paper introduces a novel approach, SCONE, which leverages stochastic sampling for generating contrastive views and hard negative samples in graph-based collaborative filtering (CF). SCONE considers the generation of contrastive views and hard negative samples as the stochastic sampling tasks based on score-based generation models (SGMs) to overcome the challenges of data sparsity and negative sampling in graph-based CF models. SCONE applies the stochastic process to synthesize the dynamic contrastive views and the stochastic positive injection to generate diverse hard negative samples. Experimental results indicate that SCONE outperforms existing CF models across six benchmark datasets. ### Strong points 1. SCONE effectively tackles the challenges of data sparsity and negative sampling in graph-based collaborative filtering (CF) models through a unified framework. 2. While SCONE introduces additional parameters for its score network, it maintains efficient space and time complexity, with a lower running time per epoch than other negative sampling methods. 3. Experimental results demonstrate that SCONE outperforms six baseline models across benchmark datasets. The results also show SCONE's robustness in addressing user sparsity and item popularity issues. ### Weak points 1. The authors can enhance their work by explaining why they use a single score network for contrastive view generation and hard negative sample generation rather than separate networks for each task. 2. Additionally, the performance of SCONE is influenced by multiple hyperparameters, as discussed in Section 5.5. Identifying an optimal combination of hyperparameters for each dataset is necessary to achieve better results. 3. In Figures 8, 9, and 10, the authors only present performance data on two datasets and conclude the hyperparameters. Expanding the sensitivity analysis to include more datasets would strengthen the results and make the conclusions in Section 5.5 more robust. ### Detailed evaluation This paper presents SCONE, a unified framework to tackle data sparsity and negative sampling challenges in graph-based CF. While some explanations and experimental details are omitted, SCONE makes notable contributions to the graph-based CF by improving recommendation performance, particularly in NDCG@20 and Recall@20, while enhancing user recommendations' diversity and personalization. Despite including an additional score network, space or time complexity does not significantly increase during training. **Overall rating:** Weak Accept (Probable accept, unless convinced otherwise) **Reviewer confidence:** Unfamiliar: This topic is completely new to me. ## Reviewer 5 ### Summary and contributions The SCONE framework introduces a novel method for addressing data sparsity and negative sampling challenges in recommender systems using stochastic sampling based on score-based generative models. ### Strong points 1. SCONE is demonstrated to handle the issues of user sparsity and item popularity effectively. The method seems promising in real-world scenarios, as shown through its performance improvements in both sparse and dense datasets. 2. The manuscript evaluates SCONE against six state-of-the-art models across six diverse benchmark datasets. This thorough experimentation adds weight to the claims of performance superiority. 3. The overall presentation is good, and the paper is easy to follow. ### Weak points 1. While the paper does a good job of detailing the experimental results, the theoretical motivation behind the chosen methods (e.g., the exact benefits of score-based generative models over other generative methods) could be elaborated. A more detailed discussion comparing SGMs to alternative generative approaches (e.g., GANs) may help readers better understand SCONE’s strengths. 2. The paper introduces the idea of “stochastic positive injection” for generating hard negative samples. However, more explanation about the underlying mechanics and intuition behind this technique would benefit readers who may not be familiar with score-based generative models. For example, why was the specific setting of # = 10 chosen as the optimal step for stochastic sampling? 3. While the manuscript acknowledges that SCONE incurs some computational overhead compared to graph-based contrastive learning methods, the runtime analysis could be expanded. A more in-depth discussion on how the stochastic sampling step could be optimized would enhance its practical applicability in large-scale systems. ### Detailed evaluation Please see the Weak Points. There are a few minor typographical errors, such as “adress” instead of “address” in the introduction. Proofreading is recommended to ensure polished writing. **Overall rating:** Weak Accept (Probable accept, unless convinced otherwise) **Reviewer confidence:** Knowledgeable: I have read papers on this topic. ## Meta-review ### Summary The paper introduces SCONE, which is an approach to address challenges of data sparsity and negative sampling in graph-based collaborative filtering. While the individual components (contrastive learning and negative sampling) are established techniques, their unified application within a stochastic sampling framework using score-based generative models is interesting. Experimental evaluation runs across six datasets and shows solid performance, particularly in handling sparse data and improving recommendation accuracy. Some reviewers are showing concerns regarding novelty, computational cost, and hyperparameter sensitivity. While SCONE offers a novel framework, its components, contrastive learning and negative sampling, are established techniques. This limits the originality of the contribution. Furthermore, the computational cost of implementing SCONE, particularly the stochastic sampling step, raises concerns about its scalability and applicability to large-scale recommender systems with high user traffic. Although SCONE demonstrates strong empirical performance, its reliance on multiple hyperparameters introduces complexity in finding the optimal configuration for different datasets. The limited selection of baselines, some of which are outdated, weakens the comparative analysis and may not fully reflect SCONE's advantages against the current state-of-the-art. Also, the evaluation focuses solely on top-20 recommendations, neglecting other common metrics that could provide a more comprehensive understanding of its performance across various recommendation depths. Recommendations for Improvement includes better elaboration on the theoretical motivation for using score-based generative models over alternative generative approaches (e.g., GANs). It is recommended to provide a more intuitive explanation of this technique and justify the specific settings used for stochastic sampling. Also, conduct a more comprehensive sensitivity analysis across all datasets to provide robust conclusions about hyperparameter choices. Expand the set of baselines to include more recent and competitive methods in both contrastive learning and negative sampling. The paper would also benefit from including results for various recommendation thresholds (e.g., top-5, top-10) to provide a more complete picture of SCONE's performance. **Recommendation:** Accept