Reviewer #1
Questions
1. Relevance to WSDM
Highly relevant
2. Strong points. Three (or more) strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
1. The paper studied session-based recommendation by exploiting both explicit and implicit item relations and proposed a dual-model framework to decouple the explicit and implicit relations separately.
2. Extensive experiments are conducted and the ablation study is comprehensive. The experiment result shows the effectiveness of the proposed model.
3. Weak points. Three (or more) weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
1.To capture the implicit relation, the author constructed a global graph that includes all the pair-wise correlations. It seems to construct a huge graph. The efficiency of investigating such a graph is questionable.
2. It is desired to include the time complexity analysis of A-GNN and runtime efficiency experiments compared to other models.
3. The technical contribution should be strengthened.
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. Please be polite, specific, and constructive.
This work studied the session-based recommendation problem by investigating implicit and explicit item relations separately via a dual model. The implicit relation is learned from the sparse adjacent matrix constructed by pair-wise occurrence in a session. And the explicit relation is learned from the pair-wise correlation of all nodes.

The intuition to cover the multi-scaled relation is reasonable. However, learning all the pair-wise correlations seems exhausting. It is desired to see the complexity analysis of A-GNN and some experiments on the efficiency.

The explicit relation graph that was constructed by pair-wise occurrence bypassed the intra-session and inter-session relation and lack of semantic information.

The methodologies that are adopted in this paper are not new. The technical contributions are limited.
5. Paper clarity. Is the paper well-written and well-organized?
Average: the main points of the paper were understandable, but some parts were not clear
6. Experiments: Do the experiments support the claims made in the paper and justify the proposed techniques?
OK, but certain claims are not supported by the experiments
7. Overall Rating
Reject
8. Reviewer confidence
Knowledgeable in the subarea
Reviewer #2
Questions
1. Relevance to WSDM
Highly relevant
2. Strong points. Three (or more) strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
1. The idea of decoupling explicit and implicit item relationships is intuitive and well motivated.
2. The proposed approach is easy to follow.
3. A significant improvement is obtained across five benchmark datasets in the experiment.
3. Weak points. Three (or more) weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
1. The technical part can better motivate the use of two different GNNs for explicit and implicit graphs.
2. The case study and visualization in the experiments can be presented in a clearer way.
3. Some GNN-based sequential recommenders that model cross-session/sequence information need to be discussed.
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. Please be polite, specific, and constructive.
This paper studies the session-based recommendation problem, where the authors point out the lack of separation between implicit and explicit item relations is a defect within existing solutions. As a solution, this paper separates the inter-session and intra-session correlations into implicit and explicit graphs, which are respectively modeled by an adaptive GNN and gated GNN. The method is clearly presented, and the experimental results can back the key claims made in this paper.

My detailed comments are provided below:
1. For implicit graph, an adaptive GNN is used for learning node embeddings; while for the explicit graph, a slightly more complex SG-GNN is used. These empirical designs need to be justified further, i.e., why gating is necessary in the explicit graph while the implicit graph only needs a more straightforward node aggregation.
2. It would be beneficial to couple the theoretical time/space complexity comparison with the real running time and memory consumption results in Table 3. Some experimental results are expected to have more discussions, e.g., the hyperparameter analysis in Section 5.6 where different datasets' sensitivity towards SG-GNN's layer size and A-GNN blocks vary a lot.
4. The visualization used in Fig 4, especially (a) and (b) are less intuitive in showing the session embedding similarity - some metrics like cosine or Jaccard might make more sense in this scenario. It is also not clear whether these embedding figures share the same scale as the attention figure (c).
5. There are GNN-based sequential recommenders that are missed in the Related Work:
- Discovering Collaborative Signals for Next POI Recommendation with Iterative Seq2Graph Augmentation, IJCAI'21
- Exploiting cross-session information for session-based recommendation with graph neural networks, TOIS'20
- Graph-enhanced multi-task learning of multi-level transition dynamics for session-based recommendation, AAAI'21
5. Paper clarity. Is the paper well-written and well-organized?
Good: most of the paper was understandable, with minor typos or details that could be improved
6. Experiments: Do the experiments support the claims made in the paper and justify the proposed techniques?
The experiments nicely support the claims made in the paper
7. Overall Rating
Weak Accept
8. Reviewer confidence
Expert in the problem
Reviewer #3
Questions
1. Relevance to WSDM
Highly relevant
2. Strong points. Three (or more) strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
1. The session-based recommendation is an active topic within RS research community. Different from existing approaches focused on defining group structures inside graph neural networks, this research looks through the implicit/explicit relations amongst items and decouples their inter-item relations, which is novel
2. A dual graph neural network is proposed that leverages the feature representations extracted by two GNNs: a graph neural network with a single gate (SG-GNN) and an adaptive graph neural network (A-GNN). SG-GNN is able to capture the explicit dependencies among items in sessions and aggregates neighbors’ information to represent items, while A-GNN employs a self-learning strategy to capture implicit correlations among items in sessions.
3. Comprehensive experiments conducted on four real world datasets are comprehensive and convincing, which have shown the significant improvement in comparisons with a good number of baselines.
3. Weak points. Three (or more) weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
1. A tabulated summary of symbals used in the paper for better brevity.
2. The empirical study results in Table 2 vary largely in different datasets, some more discussion behind the results helps
3. Toy examples/explanations for implicit/explicit relations
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. Please be polite, specific, and constructive.
Different from existing approaches focused on defining group structures inside graph neural networks, this research looks through the implicit/explicit relations amongst items and decouples their inter-item relations, which is novel.
5. Paper clarity. Is the paper well-written and well-organized?
Good: most of the paper was understandable, with minor typos or details that could be improved
6. Experiments: Do the experiments support the claims made in the paper and justify the proposed techniques?
The experiments nicely support the claims made in the paper
7. Overall Rating
Accept
8. Reviewer confidence
Expert in the problem
Reviewer #4
Questions
1. Relevance to WSDM
Highly relevant
2. Strong points. Three (or more) strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
* Strong experimental results on multiple benchmark dataset.
* The proposed method is technically sound and well motivated.
* Writing is mostly clear and easy to follow.
3. Weak points. Three (or more) weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
* Method is not scalable to real-world datasets.
* Lack of time complexity analysis for the A-GNN module.
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. Please be polite, specific, and constructive.
Overall, the experimental results are solid -- the authors compare against several strong baselines and the performance of DGNN is significantly better than all of them. The method is also sound and well-motivated.

However, one major disadvantage of the proposed method is its time complexity. In particular, the A-GNN module captures the correlations between all pairs of items. This means the time complexity of the proposed algorithm is quadratic in the number of items, which will not scale in real-world applications with hundred of millions items.

Accordingly, the authors should clarify the time complexity of the proposed A-GNN module, and discuss mitigations to make DGNN work on larger real-world datasets -- the largest dataset used in the paper only has ~43k items.
5. Paper clarity. Is the paper well-written and well-organized?
Good: most of the paper was understandable, with minor typos or details that could be improved
6. Experiments: Do the experiments support the claims made in the paper and justify the proposed techniques?
OK, but certain claims are not supported by the experiments
7. Overall Rating
Reject
8. Reviewer confidence
Knowledgeable in the subarea
Reviewer #5
Questions
1. Relevance to WSDM
Highly relevant
2. Strong points. Three (or more) strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
1. The paper is well-motivated and novel, highlighting the limitations of sequential models and existing graph modelling approaches.
2. The proposed approach is novel compared with previous approaches. It is simple yet effective and can be easily replicated for application.
3. Experiments are solid and comprehensive. The authors choose 11 baselines from five categories and the results show the proposed approach outperforms the SOTA by a larger margin, especially in LastFM dataset, which is very impressive.
3. Weak points. Three (or more) weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
1. While the merits of this paper seem to be apparent, the possible limitations of the proposed approach can be discussed.
2. Inconsistent y-axis view limits and tick locations in Figure 3.

4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. Please be polite, specific, and constructive.
This work designs a dual graph neural network to capture the prior knowledge encapsulated in explicit dependencies and learned implicit correlations among items simultaneously. Thus, the item dependencies and correlations can be modeled in a flexible and more interpretable manner for effective recommendations. The paper is generally clearly presented, and the methodology and techniques are plausible. The ablation study and case analyses further discuss and illustrate the effectiveness of each part in this model compared with other representative modules, e.g., SR-GNN, self-attention module. Figure 5 also demonstrates that the A-GNN could optimize the item representation from two key properties in contrastive learning which explains the superiority of this work from visual and theoretical aspects.

Strengths:
1. The paper is well-motivated and novel, highlighting the limitations of sequential models and existing graph modelling approaches.
2. The proposed approach is novel compared with previous approaches. It is simple yet effective and can be easily replicated for application.
3. Experiments are solid and comprehensive. The authors choose 11 baselines from five categories and the results show the proposed approach outperforms the SOTA by a larger margin, especially in LastFM dataset, which is very impressive.

Weaknesses:
1. While the merits of this paper seem to be apparent, the possible limitations of the proposed approach can be discussed.
2. Inconsistent y-axis view limits and tick locations in Figure 3.
5. Paper clarity. Is the paper well-written and well-organized?
Excellent: the paper was clear and easy to follow
6. Experiments: Do the experiments support the claims made in the paper and justify the proposed techniques?
The experiments nicely support the claims made in the paper
7. Overall Rating
Accept
8. Reviewer confidence
Knowledgeable in the subarea
Reviewer #6
Questions
1. Relevance to WSDM
Highly relevant
2. Strong points. Three (or more) strong points about the paper. Please be precise and explicit. Clearly explain the value and nature of the contribution.
1. The authors present a novel GNN structure, which can encode higher-order interactions between items in sessions.
2. The authors perform thorough experimentation. Comparing their algorithm to many of the state-of-the-art methods, and also perform an ablation study, as well as parameter sensitivity.
3. The results show strong improvements over baselines.
3. Weak points. Three (or more) weak points about the paper. Please clearly indicate whether the paper has any mistakes, missing related work, or results that cannot be considered as a contribution. Please be polite, specific, and constructive.
1. I have my doubts regarding the usability of the model in practice, given the increased time and space complexity it requires.
2. The authors do not compare to the best performing baseline from recent works comparing methods on sequential recommendation. Such as ​​Latifi, Sara, and Dietmar Jannach. "Streaming Session-Based Recommendation: When Graph Neural Networks meet the Neighborhood." Proceedings of the 16th ACM Conference on Recommender Systems. 2022.
3. The paper's use of the terms explicit and implicit information, is a bit confusing, regarding they are commonly used to distinguish users' explicit or implicit feedback.
4. Adding an example of what explicit and implicit information is in the introduction would help solve this issue I think.
4. Detailed Evaluation. Please provide detailed feedback about the strengths and the weaknesses of the paper and support your overall rating. Please be polite, specific, and constructive.
The paper describes a novel Graph Neural Network architecture, which is able to outperform other models, by taking into account both the direct relations between items visited one after the other, as well as second-order interactions using the AGNN module.
My main concern with the paper, is the feasibility of training the model. It constitutes various complex layers, and training it on larger datasets will be a challenge. Especially given that the users decided to not use the full yoochoose dataset.
For practical use, this method seems a bit too computationally complex to be relevant.
A second point of critique, is that the baselines does not include the VS-KNN.
Latifi et. al. recently published a paper at Recsys about comparing neighbourhood methods to GNN models. They confirmed that VS-KNN is also competitive with graph neural networks, previously introduced in Ludewig et al. (Ludewig, Malte, and Dietmar Jannach. "Evaluation of session-based recommendation algorithms." User Modeling and User-Adapted Interaction 28.4 (2018): 331-390.).
It would be of great benefit to the paper, to add this algorithm to the baselines used.
The paper is well written, and the results are promising enough, that I would not count these negatives strongly against the paper.

Typos and unclarities:
In section 4.2 the authors reuse the variable name `'m' to signify the amount of blocks, when it was previously used as the session length in section 3.
If the authors can early on clarify what constitutes implicit and explicit information, this could help the user intuitively understand the components of the network.
5. Paper clarity. Is the paper well-written and well-organized?
Good: most of the paper was understandable, with minor typos or details that could be improved
6. Experiments: Do the experiments support the claims made in the paper and justify the proposed techniques?
The experiments nicely support the claims made in the paper
7. Overall Rating
Accept
8. Reviewer confidence
Knowledgeable in the subarea