------------------------- METAREVIEW ------------------------ My final recommendation for this manuscript is weak accept / accept. The final reviewer recommendations are positive overall: R1 remains at **weak accept** with medium confidence, R2 moved from **weak reject** to **weak accept** with high confidence after the rebuttal, and R3 remains at **accept** with high confidence. Thus, all three final reviews are on the positive side, although the paper remains somewhat borderline due to evaluation and scalability concerns. The manuscript makes a clear and relevant contribution to graph spectral collaborative filtering. Its main idea—augmenting the user-item graph with item-item proximity derived from interaction ordering and then applying a Gaussian bandpass filter to capture intermediate-frequency preference signals—is well motivated and technically coherent. The reviewers appreciated the strong empirical results across GSP-based and GCN-based baselines, the clear ablation evidence that the bandpass component is important, and the useful insight that graph construction and filter design are coupled. In particular, the reported results and ablations support the claim that the method improves over existing GSP-based collaborative filtering approaches, and the runtime comparison suggests that the method remains close to GF-CF in efficiency while achieving better accuracy. At the same time, several limitations remain. The most important concern is that GSPRec uses interaction ordering, while the main comparisons are against general CF GCN/GSP baselines rather than temporal or sequential recommenders such as SASRec or BERT4Rec. This makes it harder to fully separate the benefit of the spectral framework from the benefit of using ordering information. Reviewers also raised concerns about inference-time scalability, the need to compute or store full user-item scores, limited robustness analysis for some hyperparameters, the fact that key ablations are mainly shown on ML1M, and some missing reproducibility details. These issues do not, in my view, invalidate the core contribution, but they do limit the strength of the empirical claim and should be clearly addressed in a camera-ready version. The rebuttal was helpful. The authors clarified that GSPRec-NS removes the ordering-derived graph and still performs strongly on ML1M, suggesting that the bandpass framework itself contributes substantially; they further argued that the ordering graph adds an additional gain rather than being the sole source of improvement. They also clarified why direction is discarded, namely to obtain a symmetric Laplacian, and why the user-user block is left empty, since user-user similarity is incorporated in the low-pass branch. The rebuttal did not fully resolve the missing sequential-baseline and deployment-efficiency concerns, but it addressed enough of the framing and methodological concerns that R2 revised the score from weak reject to weak accept. Overall, I recommend acceptance because the paper presents a focused, technically sound, and empirically strong contribution within graph spectral collaborative filtering. The remaining concerns are important limitations, especially regarding comparison to sequential methods and scalability, but they are better viewed as qualifications and directions for strengthening the work rather than reasons for rejection. I encourage the authors to make these limitations explicit, clarify the evaluation framing, discuss inference-time implications more concretely, and, where possible, add further details on reproducibility and robustness. I believe the paper will be of interest to the RecSys community and can stimulate useful discussion on spectral filtering and graph construction for collaborative filtering. ----------------------- REVIEW 1 --------------------- SUBMISSION: 357 TITLE: GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering ----------- Strong Points ----------- + The proposed method utilizes user interaction ordering (sequences) to define an item-item proximity matrix that is injected into the unified graph adjacency matrix. + The Gaussian bandpass filter is designed to selectively amplify the intermediate-frequency components that encode community-level user preferences. + The experimental comparisons with the state-of-the-art GSP-based and GCN-based CF baselines show that the proposed method outperforms them. ----------- Weak Points ----------- - In the experimental comparisons, only the proposed method utilizes the user interaction ordering (sequences) as additional information given to the input. This may be rather unfair, while it is true that standard datasets like MovieLens and Netflix include timestamps for every interaction. - The optimal parameters selected for the final model (Table 3) do not match the visual peaks found in the individual sensitivity plots (Figure 5) for the Netflix dataset. The authors should explain the reason. - The diffusion depth d = 2 and the decay \alpha = 0.4 are fixed across all datasets as they are stable, but without further justifications. ----------- Relevance to the Track ----------- SCORE: 4 (Fits) ----------- Originality ----------- SCORE: 3 (Moderate) ----------- Coverage of Related Literature ----------- SCORE: 4 (Strong) ----------- Technical Soundness ----------- SCORE: 3 (Adequate Technical Soundness) ----------- Overall Presentation ----------- SCORE: 3 (Adequate) ----------- Impact ----------- SCORE: 3 (Moderate) ----------- Reproducibility of Methods (N/A for Past, Present and Future papers) ----------- SCORE: 3 (Moderate) ----------- Review ----------- GSPRec achieves state-of-the-art results by integrating user interaction ordering into a unified adjacency matrix and employing a Gaussian bandpass filter to target community-level user preferences. Fairness concerns remain in the experimental evaluation, because only GSPRec utilizes sequential data. Additionally, the choice of $\hat{S}$ as the item-item proxy matrix may not be optimal. A simpler definition - such as setting $s_{ij} = 1$ when at least $k \geq 2$ users share a direct transition between items $i$ and $j$ - might perform better and even bypass multi-hop diffusion, as the graph Laplacian already accounts for higher-order proximity. Lastly, while this work focuses on interaction ordering to derive item-item proximity, the proposed framework is flexible enough to incorporate other extra information, like item descriptions, to measure item-item similarity. ----------- Overall Evaluation ----------- SCORE: 1 (Weak accept) ----------------------- REVIEW 2 --------------------- SUBMISSION: 357 TITLE: GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering ----------- Strong Points ----------- - The paper presents a well-designed method to addresses the over-emphasis on the low-frequency signals. - The proposed idea of incorporating item-item proximity for user interaction ordering is intuitive and potentially useful. - Experimental results show that GSPRec consistently improves over both GCN-based and GSP-based baselines across multiple datasets. - The ablation study is informative and shows that the bandpass component is a key contributor to performance. ----------- Weak Points ----------- - The paper heavily relies on the claim that existing graph-based CF methods mostly behave as low-pass filters, but it does not provide sufficient empirical evidence for this claim. - The proposed sequence-derived item-item graph is not clearly distinguished from item-item relations already induced by multi-hop user-item propagation. - Since the method uses interaction ordering, copmarisons with temporal/sequential methods (e.g., SASRec [1], BERT4Rec [2]) would be valuable. - The unified graph adds item-item edges but keeps the user-user block empty. The paper should justify this and discuss whether it can be extended to user-user edges as well. - The sensitivity of the fusion weight \phi is only partially studied, so it would be better to support the robustness more comprehensively. - Scalability remains unclear, given the eigendecomposition, item-item diffusion graph, and full user-item score matrix computation. - The efficiency analysis focuses on offline runeimt, but real deployment requires inference-time efficiency. I am concerned that GSPRec may require computing or storing a full user-item score matrix, unlike embedding-based methods that can support efficient top-K retrieval via ANN search. - The method uses temporal order to build item proximity, but symmetrizes transitions (undirected edges), so it may capture local co-occurrence rather than true temporal direction. - Key ablations are mainly reported on ML1M, but the main claims should be validated across all datasets to show that the observed effects are not dataset-specific. [1] Self-Attentive Sequential Recommendation (ICDM 2018) [2] BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer (CIKM 2019) ----------- Relevance to the Track ----------- SCORE: 5 (Strong fit) ----------- Originality ----------- SCORE: 3 (Moderate) ----------- Coverage of Related Literature ----------- SCORE: 4 (Strong) ----------- Technical Soundness ----------- SCORE: 4 (Strong Technical Soundness) ----------- Overall Presentation ----------- SCORE: 4 (Well-Presented) ----------- Impact ----------- SCORE: 3 (Moderate) ----------- Reproducibility of Methods (N/A for Past, Present and Future papers) ----------- SCORE: 2 (Weak) ----------- Review ----------- This paper proposes GSPRec, a graph signal processing framework for collaborative filtering that augments the user-item graph with sequence-derived item-item proximity and applies a combination of bandpass and low-pass spectral filters. The main idea is to address the over-emphasis on low-frequency signals in existing graph-based recommenders and better capture intermediate-frequency performance patterns. The paper is generally well motivated and presents an intuitive approach. The proposed bandpass filtering component is well aligned with the paper's spectral perspective. The empirical results show consistent improvements over both GCN-based and GSP-based baselines across multiple datasets. However, I have several concerns. First, the paper heavily relies on the claim that existing graph-based CF methods, but provides limited direct empirical evidence. Second, it does not clearly distinguish the proposed sequence-derived item-item graph from item-item relations already capturable from multi-hop user-item propagation. Third, comparisons with temporal/sequential recommenders would be valuable. Fourth, the choice of adding item-item edges but not user-user edges needs more justification. Finally, robustness, scalability, and reproducibility remain unclear. Overall, the paper presents a promising and intuitive idea with strong empirical results, but there is room for improvement and clarification. ----------- Overall Evaluation ----------- SCORE: 1 (Weak accept) ----------- Key Clarification ----------- Please refer to the weaknesses above. Among them, I would especially appreciate clarification on two issues: - the inference-time efficiency of GSPRec, particularly whether it requires computing or storing a full user-item score matrix and how it compares with embedding-based recommendation - comparisons with temporal/sequential recommendation models such as SASRec and BERT4Rec, since the proposed method explicitly uses interaction ordering ----------------------- REVIEW 3 --------------------- SUBMISSION: 357 TITLE: GSPRec: On Improving Item Representations in Graph Signal Processing for Collaborative Filtering ----------- Strong Points ----------- 1. The paper identifies a genuine gap: existing GSP-based CF methods all derive item representations from the user-item interaction matrix alone, systematically discarding intermediate-frequency components. The spectral energy analysis in Figure 4 provides empirical grounding for why a bandpass filter is appropriate, not just a design choice. 2. The coupling insight between graph construction and filter design is well-validated. GSPRec-NB falling below all GSP baselines is a strong result because it shows the item-item proximity structure actively hurts without the bandpass filter, which is a non-obvious and informative finding. 3. Achieving near-GF-CF runtime (1.27 vs. 1.18 min on ML1M) while substantially outperforming it on accuracy is a meaningful practical contribution. The gap over GCN-based methods (100x faster than LightGCN) is especially notable for deployment contexts. 4. Lemma 1 (diffusion convergence) and Proposition 2 (Laplacian validity) are well-placed and directly relevant to the method. They provide formal guarantees that the constructed Laplacian is well-behaved for spectral filtering, which is necessary for the bandpass filter to target a well-defined spectral region. ----------- Weak Points ----------- 1. The paper does not compare against sequential recommendation methods such as SASRec or BERT4Rec, which also exploit user interaction ordering. Since the proposed graph construction explicitly leverages interaction ordering as a key signal, these are natural and important comparisons. Without them, it is unclear whether the gains come from the spectral framework itself or simply from incorporating sequential co-occurrence information that sequential models already capture more directly. 2. The paper shows empirically that bandpass filtering on the enriched graph outperforms low-pass competitors, but provides no theoretical analysis of why intermediate-frequency components carry more personalization signal than low-frequency ones in the general case. The spectral energy distribution in Figure 4 is dataset-specific evidence, not a proof. A bound on approximation error or an information-theoretic argument for why the enriched Laplacian exposes more useful structure would substantially strengthen the contribution. 3. The bandpass center c and width w are tuned per dataset on a fixed validation split. When the underlying user preference distribution shifts (cold-start users, evolving item catalogs, concept drift over time), the optimal spectral band likely shifts too. The paper does not evaluate performance under such conditions, and the fixed hyperparameters across time may be a significant limitation in production settings. 4. The paper acknowledges that polynomial filter approximations could reduce complexity from O(r(m+n)^2) to O(K|D|), but this is deferred to future work. For the Netflix dataset (20K users, 17K items), a truncated decomposition with r=256 is already non-trivial. The scalability gap to industrial-scale graphs is large and the conclusion's suggestion does not constitute a solution. ----------- Relevance to the Track ----------- SCORE: 4 (Fits) ----------- Originality ----------- SCORE: 4 (High) ----------- Coverage of Related Literature ----------- SCORE: 3 (Moderate) ----------- Technical Soundness ----------- SCORE: 4 (Strong Technical Soundness) ----------- Overall Presentation ----------- SCORE: 4 (Well-Presented) ----------- Impact ----------- SCORE: 3 (Moderate) ----------- Reproducibility of Methods (N/A for Past, Present and Future papers) ----------- SCORE: 4 (Highly Reproducible) ----------- Review ----------- GSPRec makes a focused and well-motivated contribution to graph spectral collaborative filtering. The core insight, that existing GSP-based CF methods derive item representations from the user-item interaction matrix alone and thereby discard the intermediate-frequency band where community-level preferences reside, is clearly articulated and consistently supported throughout the paper. The proposed solution, enriching the graph topology with item-item proximity derived from interaction ordering before applying a Gaussian bandpass filter, is technically coherent and the coupling between graph construction and filter design is convincingly established through ablation. ----------- Overall Evaluation ----------- SCORE: 2 (Accept)