## Reviewer 1 **Overall evaluation:** 1 (weak accept) ### Overall evaluation (comments) This study aims to address the challenges of data sparsity and the cold-start problem in web service recommendation using Quality of Service. The authors propose a novel approach called QoS-aware graph contrastive learning (QAGCL) for web service recommendation. Their model harnesses the power of graph contrastive learning to handle cold-start problems and improve recommendation accuracy effectively. By constructing contextually augmented graphs with geolocation information and randomness, the model provides diverse views. Through the use of graph convolutional networks and graph contrastive learning techniques, the authors learn user and service embeddings from these augmented graphs. The learned embeddings are then utilized to seamlessly integrate QoS considerations into the recommendation process. Experimental results demonstrate the superiority of the QAGCL model over several existing models, highlighting its effectiveness in addressing data sparsity and the cold-start problem in QoS-aware service recommendations. Overall, this paper is well written, but the Methodology and Experiments chapter needs to be supplemented. * Methodology - It is necessary to supplement the contents of the 'Overall Architecture' chapter and Fig.3. The explanation of the first part of the entire framework is well written, but the explanation of the latter part is insufficient. It seems necessary to supplement the contents of the contrastive learning procedure. * Experiments - There is no explanation in the text of what ratio ρ means. - Ratio κ is a parameter for g_HD, not g_ED. (in equation (6)) - The authors provide information on the hyperparameter candidates used in the experiment, but do not provide information on the final selected best hyperparameter setting. It is necessary to mention which hyperparameter settings the results in Tables II, III, and IV are from. - In chapters D and E, it is necessary to add results for warm-start and cold-start* cases. Only the results for the case of cold-start are covered. - In Table V, it is also necessary to add results for Recall@40 and NDCG@40. It is difficult to assert the effectiveness of HD&ED only with the contents currently written. - Is ED & ED in the third line of Table V a typo? Should it be modified with ED & ND? - The content of Chapter E needs to be rewritten. Why did you do sensitivity analysis for number of graph layers? From the content of this chapter, it is not clear what the author wants to say. Also, it seems that 'L is 2' and 'for Recall@40 and NDCG@40' in the second sentence should be modified to 'L is 3' and 'for NDCG@20, Recall@40 and NDCG@40'. ## Reviewer 2 **Overall evaluation:** 1 (weak accept) ### Overall evaluation (comments) # Summary This study endeavors to recommend web services based on the Quality of Service (QoS). To address data sparsity and the cold-start problem, the authors utilize graph contrastive learning (GCL) integrated with geolocations of users and services. The proposed approach shows a marginal improvement in recommendation quality when compared to existing models. #Pros - The idea of incorporating geographical factors that might affect responses and throughput to improve the accuracy of web service recommendations seems justifiable. - This idea, though to a limited extent, has contributed to the enhancement of recommendation quality. # Cons - The degree of improvement in recommendation quality is minimal. Additionally, the absence of an ablation study means it's unclear which components of the proposed method were most effective. - It's a common practice for users to consider the geographical distance of services. Presumably, users would naturally prefer services closer to them. Thus, for users, distance might be the final criterion when choosing from a set of suitable services. A thorough discussion should be made regarding the benefits to users of focusing on proximity at the recommendation stage. - Major cloud service providers offer equivalent web services at multiple locations. From the perspective of application developers, the diversity of web service provision locations should be a major criterion for web service selection. However, this method seems limited to services provided at a single location, raising doubts about its relevance to developers' needs. A detailed discussion on this aspect seems warranted. - Setting aside the above points, even if we accept the premise that geolocation is vital for users and services, it's unclear why geolocation should be handled within the GCL framework. Intuitively, vertex attributes should be treated as such and processed using a Graph Neural Network (GNN). The advantages of constructing a "Geolocation-aware Graph" with GCL, as opposed to treating it simply as a vertex attribute, should be demonstrated both in terms of design intent and experimental results. # Comments - The pairing here is convoluted, making it unclear how "respectively" applies. > Models M1 and M2 predict the QoS grades of s1 and s2 as > 0.3, 0.6 and 0.5, 0.45, respectively. - The definition of Recall@k in Equation (19) seems off. The denominator should simply be relrel, not min(K,rel)min(K,rel). Also, while relrel is defined as the total number of related items for the user, it's unclear what constitutes "related items" and how they're measured. ## Reviewer 3 **Overall evaluation:** 1 (weak accept) ### Overall evaluation (comments) 1. The data set of the whole paper uses WSDream, in which there are 57,727 interactions in the warm-start environment and two data sets in the cold-start environment, each with 8,490 and 1,036 interaction matrices, to verify whether the conclusion data set of the paper is insufficient. Would you consider supplementing some other data sets for cross-validation? 2. The total amount of data is mentioned throughout the paper, but the relevant data structure of the selected data set is not explained. It is recommended to consider giving an appropriate explanation. 3. It is mentioned in this paper that context (such as geolocation) needs to be used. When the experiment was explained, the processing flow of the context was not analyzed and explained. It is recommended to supplement the process relationship. 4. The process of the model training algorithm proposed in this paper is too rough, and it is suggested to properly supplement and improve the core algorithm involved. 5. The paper mentions the importance of using context (such as geolocation) and suggests adding some literature support, which makes it more convincing. The importance of using context (such as geolocation) mentioned in the paper suggests adding some literature support, so as to make it more convincing. 6. The conclusion obtained at the end of the paper uses graph contrastive learning. The model effectively handles the cold start problem and the data sparse problem and provides more accurate suggestions. However, the ratio of superiority in the experiment is not very obvious. So it is suggested to add more dimensions to the experimental demonstration.