acceptodds
Under review as a conference paper at ICLR 2027

PRGCL: Path Reconstruction for Graph Contrastive Learning in Collaborative Filtering

Abstract

Graph contrastive learning (GCL) has recently emerged as an effective self-supervised enhancement paradigm for alleviating data sparsity. However, existing methods still face three key challenges: (1) Stochastic structural perturbations may compromise local structure, making the constructed views unreliable; (2) The mixing of frequency domain signals can cause noise interference, leading to ambiguous optimization objectives; (3) Independently constructed views may induce different connectivity in the signal propagation path. To address these challenges, we propose a coupled reconstruction principle that leverages path reconstruction to identify task relevant structural subsets, guiding self-supervised view enhancement for graph collaborative filtering. Building on this, we propose Path Reconstruction for Graph Contrastive Learning in collaborative filtering (PRGCL), which improves graph contrastive recommendation by constructing reliable self-supervised views, refining frequency domain signals, and enforcing connectivity along signal propagation paths. Extensive experiments on three highly sparse real-world datasets demonstrate that PRGCL consistently outperforms state-of-the-art baselines in recommendation accuracy.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
recommender system, graph contrastive learning, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_ucgqOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors claimed that in previous contrastive learning methods for recommender system, the data augmentation will destroy the topology of the interaction graph. PRGCL introduces learnable auxiliary prototype nodes and establishes new connections based on their associations with users and items, providing additional channels for information propagation. It constructs two augmented graph views and jointly optimizes a recommendation ranking loss, a contrastive learning loss, and auxiliary regularization terms to learn more effective user and item representations. The authors report state-of-the-art performance on all three evaluated datasets: Douban-book, Tmall, and Alibaba-iFashion.

Critical strengths
  1. The motivation of this paper is pretty interesting. It is an interesting idea to find an augmentation methods that will not destroy the topology.
  2. The authors give some theoretical proofs for this paper, which makes the paper more convincing.
Critical weaknesses

The main weakness of this paper lies in its experimental evaluation. I would be willingly to reconsider my score if the revised version provides a more comprehensive and convincing experimental assessment.

  1. The evaluation is limited to three datasets and omits Amazon-Book, an established benchmark also used by MixRec. Could the authors justify their dataset selection and consider including Amazon-Book to broaden the evaluation? Furthermore, the MixRec results on the two shared datasets differ from those reported in its original paper: Recall@20 is 0.1668 versus 0.1778 on Douban-Book, and 0.0875 versus 0.0900 on Tmall. Could the authors explain these discrepancies, including any differences in preprocessing, data splits, evaluation protocols, or hyperparameter tuning, and disclose any difficulties reproducing the original results? If it is because some of the methods cannot be reproduced (It is pretty common in the field of recommender system), I can fully understand and we can discuss a way to make the results more convincing.
  2. The current ablation study mainly evaluates individual components within PRGCL. Controlled comparisons with related methods such as SimGCL would be informative—for example, swapping view-construction strategies while keeping the backbone, contrastive objective, and training settings fixed. Such experiments would help isolate which design choices account for the improvements over existing approaches.
  3. Although the appendix includes a case visualization, a more detailed analysis of specific propagation paths would make the mechanism clearer. Could the authors trace representative user–item pairs and illustrate how reconstructed connections change their propagation paths and recommendation rankings compared with SimGCL? This would provide a more concrete connection between the proposed mechanisms and the observed gains.
  4. (Not weakness but a question...I just want to put them together) Can this methods be applied on more general graph tasks?Or it is specially designed for recommender system?
Questions for authors

See weaknesses

3 comments

We thank reviewer ucgq and provide our responses below.

W1

Dataset Selection. We selected Douban-book, Alibaba-iFashion, and Tmall to evaluate the generality of the coupling reconstruction principle across user-item graphs with different domains, scales, and sparsity levels. Amazon-book and Douban-book both belong to the book recommendation domain, while Alibaba-iFashion provides a different recommendation scenario with 300,000 users and an interaction density of 0.01%. Moreover, previous studies have identified false-positive interactions in Amazon-book[1], which may affect the performance of methods designed to data augmentation. We therefore prioritized Alibaba-iFashion to broaden the evaluation settings and will include Amazon-book in the revision. We will include Amazon-book dataset in the revised version.

MixRec Reproduction. We evaluated all baselines under a unified experimental framework to ensure a fair comparison and minimize the influence of other factors. Following the SimGCL implementation, we filter duplicate negative samples within each batch when computing the InfoNCE loss, while MixRec does not apply this filtering. This difference may explain the performance variations on Douban-book. Furthermore, MixRec removed some interactions during preprocessing, leaving 2,357,450 interactions in its Tmall dataset, approximately 90% of the 2,619,389 in our dataset, which may explain the performance differences. We appreciate the reviewer's constructive feedback and will further clarify the final results in a subsequent update.

Experimental Settings. To ensure a fair comparison, we evaluated all baselines under a unified PyTorch framework with a fixed random seed of 2026. Following previous classical studies [2, 3], we randomly split the datasets into training and test sets at a ratio of 8:2. For data preprocessing, we used the existing integer user and item IDs to construct a user-item bipartite graph from the training interactions and applied symmetric degree normalization. We adopted full-ranking evaluation with observed interactions masked and reported Recall@K and NDCG@K (K = 10, 20). All GCN-based models use 64-dimensional embeddings, 3 layers, and Xavier initialization. We used Adam with a learning rate of 0.001, an L2 coefficient of 0.0001, a batch size of 4096, and early stopping with a patience of 20. Experiments were conducted on an RTX 4090 GPU, except for DimCL on Alibaba-iFashion due to memory requirements. Some details are provided in Section 4.1, 4.4, and Appendix A.7.1, A.7.6.

W2

We conducted an experiment by replacing PRGCL's reconstructed views with SimGCL's reconstructioned views(+SimGCL_views), while keeping the GCN backbone, contrastive objective, and training settings unchanged :

DatasetModelRecall@10NDCG@10Recall@20NDCG@20
Douban-bookBase0.12470.17210.18110.1757
+SimGCL_views0.12050.16900.17960.1542
Alibaba-iFashionBase0.08370.04830.12230.0586
+SimGCL_views0.08030.04590.11660.0557
TmallBase0.05600.05370.08900.0677
+SimGCL_views0.05390.05140.08570.0597

+SimGCL_views does not perform as well as PRGCL. These results further support the contribution of learned path reconstruction to recommendation performance.

W3

Appendix A.7.2(Table 5) examines the survival and recovery of three-hop propagation paths, showing that affinity shortcuts can restore disrupted connections. The shortest-path view ablation further demonstrates the contribution of learned reconstructed views to recommendation performance. In addition, Appendix A.7.3(Figure 4) compares the local graph structures and Top-K predictions of PRGCL and SimGCL.

To further clarify the mechanism, we traced three user–item pairs from Douban-book, Tmall, and Alibaba-iFashion. The results show that PRGCL introduces additional three-hop propagation paths through learned affinity shortcuts in both reconstructed views. For the Douban-book pair (u10689, i5), the replacement view introduces the path u10689 → Q202 → u12079 → i5, while the addition view introduces u10689 → i725 → P509 → i5. The target item ranks 1st in PRGCL and 22,319th in SimGCL. Similar reconstructed paths are observed on Tmall and Alibaba-iFashion, where the corresponding rankings are 7th versus 39,788th and 10th versus 105th, respectively. These cases illustrate the changes in propagation paths and their association with improved recommendation rankings. We will include the detailed results in Appendix A.7.3.

W4

To further examine the applicability of PRGCL beyond recommendation, we extended it to the Link Prediction task and compared it with GAE[4] and AICoG[5]. When PuP_u = 8 and QiQ_i = 8, PRGCL achieves the best performance:

DatasetModelAUCAP
EnzymeGAE0.95580.9613
AICoG0.97960.9696
PRGCL0.98810.9919
IonChannelGAE0.93220.9191
AICoG0.97470.9735
PRGCL0.97390.9810

PRGCL achieves the best performance on Enzyme and remains competitive with AICoG on IonChannel, outperforming GAE on both datasets. These results suggest that the proposed coupled reconstruction principle can also benefit general graph learning tasks.

Reference

[1] Wang W, et al. Denoising implicit feedback for recommendation. [C]// WSDM 2021.

[2] Wang X, et al. Neural graph collaborative filtering[C]// SIGIR 2019.

[3] He X, et al. Lightgcn: Simplifying and powering graph convolution network for recommendation[C]// SIGIR 2020.

[4] Kipf, T. N., & Welling, M. (2016). Variational graph auto-encoders. arXiv preprint arXiv:1611.07308.

[5] N. Nakis et al. Aitchison Embeddings for Learning Compositional Graph Representations. [C]// ICML 2026.

Thanks for responses. After reading all the responses and other reviewers' comments, I plan to maintain the score.

Reviewer_jj9bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper argues that random structural augmentation in graph contrastive learning for recommendation can break key shortest propagation paths, mix signals from different propagation orders, and produce views with inconsistent connectivity. It formalizes these issues with a truncated path complex and three propositions, then proposes PRGCL. PRGCL builds two views on a LightGCN backbone: one with a few edges removed and one with the full graph. Both views receive top-q weighted shortcut edges to learnable user/item affinity matrices. Training combines BPR, an InfoNCE loss across views, and an entropy regularizer on the affinities. Experiments on Douban-book, Alibaba-iFashion, and Tmall show gains over eight baselines in Recall/NDCG. These are supported by ablations, path diagnostics, sparsity, robustness, and sensitivity studies, plus a complexity analysis.

Critical strengths

1.The paper offers an intuitive, path-centric view of how random augmentation can harm self-supervised signal propagation in GCL. Fig. 1 illustrates the three failure modes with clear concrete examples. 2. The method is practical and scalable. It avoids explicit path enumeration through interaction-level sampling and top-q shortcuts, and the appendix gives a reasonably complete complexity analysis. 3. The empirical study is broad. Beyond the main comparison, it includes ablations, sparsity and robustness analyses, and a path-level diagnostic that tries to directly measure shortest-path survival and recovery, linking the motivation to empirical evidence.

Critical weaknesses
  1. Prop. 1 is trivial and assumes views add no new paths, which contradicts PRGCL’s shortcut edges and full-graph addition view. Props. 2–3 are essentially definitions. The boundary operator, the ρ-tailed and dominated-subpath criteria, and the D/C/I terms of Eq. 3 are never computed or optimized. The shortcuts connect nodes to affinity prototypes rather than to user–item chords, so they do not implement the contraction operator T.
  2. There is no validation set, the best epochs are marked on the test set (Fig. 3), and all hyperparameters are selected using test results, which suggests test-set leakage.
  3. Frequency domain refers to learnable prototypes and hop order, with no spectral analysis. Homotopic consistency actually means connected-component consistency and is never defined.
Questions for authors
  1. The paper uses an 80/20 train/test split together with early stopping. Which data split is monitored for early stopping and hyperparameter selection? Is there an independent validation set? If not, please re-report all methods under a train/validation/test protocol.
  2. PRGCL uses uniform random sampling to determine deleted edges and shortcut anchors. Could you compare it against a structure-preservation strategy guided by path importance, for example one that protects edges lying on shortest paths between critical pairs? This would show whether explicit path information is actually necessary.
  3. Which terms in the actual training objective (Eq. 11) correspond to D, C, and I in Eq. 3? Can you prove any upper-bound, lower-bound, or other optimization relationship between the implemented losses and these terms?
3 comments

We thank reviewer jj9b and provide our responses below.

W1

Prop.1 quantifies the loss of effective self-supervised signals caused by shortest path disruption under local structural collapse (Appendix A.3). Guided by this result, we preserve the full interaction backbone in the addition view and use affinity shortcuts to restore disrupted connectivity in the replacement view (Eq.4 and Appendix A.7.2). Props.2–3 further establish the contraction operator TT and the reliable path family PKrelP_K^{rel}, based on the truncated propagation path complex, the boundary operator, and the strictly dominated-subpath and ρ\rho-tailed criteria. Following these path-level principles, we aggregate the affinity assignments of sampled opposite endpoints (Eq.6) and construct weighted top-qq shortcuts through affinity prototypes (Eq.7), compressing high-order propagation signals into compact connections. We further learn representations from reconstructed views to encourage reliable propagation paths. Since explicit path enumeration is computationally prohibitive, we instantiate the D\mathcal D, C\mathcal C, and I\mathcal I terms in Eq.3 through core signal path structure encoding, frequency domain shortcuts, and contrastive learning (Eqs.4–8), as detailed in Appendix A.4.1 and A.6.

W2 and Q1

As shown in Section 4.1, we apply the 8:2 training and testing set division following the classical previous works [1, 2]. To rule out test-informed model selection, we also use an 7:1:2 train-validation-test split. Parameters are learned only from training data. Validation controls hyperparameter selection and early stopping and the test set is evaluated once after all settings are fixed. If we are able to complete all the experiments before the rebuttal deadline, we will report the full results.

DatasetModelRecall@10NDCG@10Recall@20NDCG@20
Douban-bookLightGCN0.10460.13640.15010.1434
SimGCL0.12480.16480.17740.1580
PRGCL0.12590.17300.18290.1762
Alibaba-iFashionLightGCN0.06990.03860.10510.0498
SimGCL0.08100.04690.11550.0565
PRGCL0.08430.04890.12270.0591
TmallLightGCN0.04360.04150.07020.0510
SimGCL0.05620.05330.08460.0649
PRGCL0.05670.05400.08940.0678

W3

In PRGCL, frequency domain signal separation is related to the propagation paths in GCN. Different GCN layers aggregate signals along paths of different lengths, leading to different frequency responses (Eq.1 and Appendix A.6). The learned affinity assignments (Eqs.5–6) guide the construction of top-q frequency domain shortcuts (Eq.7), which change signal propagation across GCN layers and help reduce cross-order frequency domain interference, as motivated by Prop. 2. For homotopic consistency, we consider the consistency of reliable propagation paths across contrastive views. Fig.1(d) illustrates this issue through changes in connected components, while Prop. 3 defines the relativized path family PKrelP_K^{rel}, and Eq.3 describes the incompatibility between the two path families. We further learn representations from reconstructed views through contrastive learning (Eq.8), whose connection to multihop path consistency is analyzed in Appendix A.6. We will clarify these connections in Sections 3.1–3.2.

Q2

To examine whether explicit path importance is necessary for view construction, we compare PRGCL with an importance sampling strategy that protects edges on shortest paths between critical pairs. As shown in table, PRGCL consistently achieves better performance across all three datasets. In PRGCL, the addition view preserves the original shortest propagation paths, while the replacement view uses learned affinity shortcuts to recover disrupted connectivity. These results support our coupled reconstruction principle, showing that reliable propagation paths can be preserved and reconstructed without explicitly computing shortest-path importance during sampling.

DatasetModelRecall@10NDCG@10Recall@20NDCG@20
Douban-bookImportance Sampling0.12000.16870.16790.1515
PRGCL0.12470.17210.18110.1757
Alibaba-iFashionImportance Sampling0.08170.04560.11990.0554
PRGCL0.08370.04830.12230.0586
TmallImportance Sampling0.05360.04920.08390.0600
PRGCL0.05600.05370.08900.0677

Q3

In Eq.3, we formulate D\mathcal D, C\mathcal C, and I\mathcal I as three criteria for path reconstruction. We control D\mathcal D through core signal path structure encoding (Eq.4) and C\mathcal C through weighted top-qq shortcuts (Eqs.5–7), with Lcar\mathcal L_{\mathrm{car}} regularizing the learnable affinity matrices (Eq.9). We further optimize I\mathcal I through Lssl\mathcal L_{\mathrm{ssl}} (Eq.8), while Lbpr\mathcal L_{\mathrm{bpr}} guides recommendation learning. For the optimization relationship, Appendix A.6 establishes an upper bound on representation discrepancy in terms of multihop path signal differences (Eq.43). The cosine similarity identity in Eq.44 further connects the positive contrastive term to normalized representation distance. These results establish a mathematical connection between the implemented contrastive objective and path-induced representation consistency. We will clarify these relationships in Section 3.2 and Appendix A.6.

Reference

[1] Wang X, et al. Neural graph collaborative filtering[C]// SIGIR 2019.

[2] He X, et al. Lightgcn: Simplifying and powering graph convolution network for recommendation[C]// SIGIR 2020.

Thank you for your response. I choose to maintain my score.

Reviewer_jdiROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

PRGCL identifies a practical challenge: in traditional GCL, small random perturbations may lead to an excessively large semantic gap.

Critical strengths
  1. The idea of using highly reliable results obtained from the complete graph to constrain the perturbation process is interesting and reasonable.
Critical weaknesses
  1. The paper raises a valid problem, but there is insufficient evidence connecting the core method to its theoretical claims. The paper aims to appropriately preserve and reconstruct propagation paths; however, it does not demonstrate that the proposed method achieves this. This edge-addition strategy may also introduce a semantic gap, for example by artificially creating interactions where none exist. The paper does not analyze the distribution of the edges added in the first stage or whether this distribution supports its claims.

  2. Related methods that also involve graph perturbation, such as AdaGCL and LMACL, are not mentioned in either the related work or the experimental comparisons.

Questions for authors
  1. The experimental protocol is unclear. The added paths are obtained from the first-stage model, but was the test set used to evaluate that model?
2 comments

We thank reviewer jdiR and provide our responses below.

W1

Path Reconstruction Validation

The proposed reconstruction operations are motivated by the path analysis in Section 3.1. As shown in Eq.4, the addition view retains all original interaction edges to preserve key propagation paths (Prop.1). The replacement view removes sampled edges and introduces weighted affinity shortcuts through Eq.7 to recover disrupted connectivity, following the path contraction motivation of Prop.2. These shortcuts connect users and items to learnable affinity nodes, with scores derived from observed interactions. The original interactions are still used for BPR supervision. Contrastive learning further encourages consistency between the two reconstructed views (Prop.3).

Appendix A.7.2 provides direct evidence of path recovery. On Tmall and Alibaba-iFashion, the recovery rates reach 99.70% and 90.67%, compared with strict survival rates of 78.86% and 29.79% after edge removal. Table 2 further shows that replacing learned paths with random paths reduces recommendation performance. We will clarify these connections in the revised manuscript.

Semantic Reliability of Added Edges

We construct weighted affinity shortcuts based on observed interactions to connect user/item nodes with learnable affinity nodes, while keeping the original interactions unchanged for BPR supervision. Table 2 shows that replacing learned reconstructed paths with random paths reduces recommendation performance, demonstrating the effectiveness of the learned connections.

Distribution of Added Affinity Shortcuts

To examine the distribution of added affinity shortcuts, we compare the historical interaction overlap of connected user and item pairs with that of random connections. The overlap is measured using Jaccard similarity based on historical interactions. As shown in table, the learned shortcuts consistently achieve higher overlap than random connections in both reconstructed views across all three datasets. This indicates that the learned shortcut distribution favors nodes with similar interaction histories. The path diagnostics in Appendix A.7.2 further show that these shortcuts recover disrupted connectivity, supporting the preservation and reconstruction of propagation paths.

DatasetViewUser Overlap RatioItem Overlap Ratio
Douban-bookReplacement151.8%219.2%
Addition126.1%154.8%
TmallReplacement469.6%292.3%
Addition255.9%171.2%
Alibaba-iFashionReplacement1010.0%590.6%
Addition574.1%317.6%

W2

AdaGCL constructs adaptive contrastive views through graph generation and denoising, while LMACL employs GCN and multi-head GAT for learnable augmentation. PRGCL focuses on reconstructing signal propagation paths through core graph structures and affinity shortcuts. We will include these two methods in Section 2 and have further evaluated them on three datasets. As shown in table, PRGCL consistently achieves better recommendation performance than both methods.

DatasetModelRecall@10NDCG@10Recall@20NDCG@20
Douban-bookAdaGCL0.12460.15730.17260.1500
LMACL0.12110.15980.17190.1582
PRGCL0.12470.17210.18110.1757
Alibaba-iFashionAdaGCL0.07050.03500.10120.0434
LMACL0.07410.03910.10490.0448
PRGCL0.08370.04830.12230.0586
TmallAdaGCL0.04960.04820.07430.0535
LMACL0.04760.04750.07640.0531
PRGCL0.05600.05370.08900.0677

Regarding jdiR-W1, dsFN-W1, and jj9b-W1, all three reviewers raised concerns about the lack of a sufficient connection between the core method and the theoretical claims. Although the authors provided further explanations and experimental analyses, the evidence remains insufficient to demonstrate that the proposed algorithm achieves the effects claimed by the theory. The authors' current explanations and the theoretical propositions in the paper are not sufficient to change my assessment. I will also follow the other reviewers' further discussions on this issue.

Regarding jdiR-W2, I consider graph augmentation to be the core of the proposed method. Therefore, the distinction between PRGCL and existing learnable or adaptive graph augmentation methods is crucial for assessing the novelty of this work. Although the authors added experimental comparisons with AdaGCL and LMACL in the rebuttal, their discussion of these methods remains relatively limited. I hope the authors can further consider the fundamental differences between PRGCL and such methods and provide a more thorough explanation, rather than merely describing the methods and comparing their performance.

Regarding jdiR-Q1, the authors have not directly answered my question. According to their response to jj9b-W2/Q1, the original experiments used an 8:2 training/test split. Therefore, I believe the first-stage model may have used the test set for evaluation or model selection.

Although dsFN and jj9b gave weak accept ratings, I believe the above concerns, particularly those regarding the novelty of the core method and its connection to the theoretical claims, have not been sufficiently addressed. Therefore, I have decided to maintain my original rating.

Reviewer_dsFNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRGCL, a graph contrastive learning framework for collaborative filtering based on path reconstruction. The method identifies three limitations of existing graph augmentation strategies: structural reliability degradation, cross-order signal interference, and propagation topology inconsistency. Experiments on real-world datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper provides an interesting perspective on graph contrastive learning by analyzing view construction from the perspective of signal propagation paths.

  2. The proposed path reconstruction strategy is well motivated and integrates structural preservation, shortcut construction, and contrastive learning into a unified framework.

  3. Experimental results on three datasets show consistent improvements over several competitive baselines.

Critical weaknesses
  1. The theoretical analysis is somewhat disconnected from the actual algorithm. In particular, the relationship between the proposed path-based propositions and the specific reconstruction operations could be better justified.

  2. The concept of frequency-domain signal separation is not sufficiently clear. The method mainly relies on learnable affinity assignments and shortcut connections, but it remains unclear how these operations explicitly separate different frequency components.

  3. Experiments are conducted on only three datasets. More evaluations on datasets with different characteristics would help demonstrate the generalizability of the proposed method.

  4. Although the paper discusses computational complexity, the additional computational overhead introduced by path reconstruction and shortcut construction is not sufficiently evaluated against a broader range of baselines.

  5. Figure 2 is overly crowded, with too many elements and cluttered lines, making it difficult to read.

  6. The experimental setup is not described in sufficient detail, including the GPU model, hyperparameter configurations, optimizer selection, and other relevant implementation details.

Questions for authors
  1. How do the proposed theoretical propositions directly justify the specific path reconstruction operations in PRGCL? Can the authors provide a more explicit connection?

  2. How does PRGCL perform on larger-scale recommendation datasets, particularly in terms of training efficiency and memory consumption?

  3. Could the authors provide more details about the experimental setup, such as the GPU model, hyperparameter configurations, optimizer selection, and other implementation details?

1 comment

We thank reviewer dsFN and provide our responses below.

W1 and Q1

The three propositions provide the theoretical guidance for the reconstruction operations in PRGCL. Specifically, Prop.1 shows that losing shortest propagation paths weakens effective signals. Therefore the addition view preserves the original interaction backbone, while the replacement view uses learned shortcuts to recover disrupted connectivity (Eqs.(4) and (7)). Prop.2 motivates shortening high-order propagation paths, which is achieved through Top-q affinity shortcuts (Eqs.(5)–(7)). Prop.3 characterizes reliable path families, guiding selective shortcut construction and contrastive representation (Eqs.(7)–(9)) to encourage consistency across reconstructed views. These reconstruction steps follow the coupled reconstruction principle in Eq.(3), as described in Appendix A.4.1. We will further explain this connection in Section 3.2.

W2

In PRGCL, frequency domain signal separation is related to the propagation paths in GCN. Different GCN layers aggregate signals along paths of different lengths, leading to different frequency responses. The learned affinity assignments (Eqs.(5)–(6)) guide the construction of Top-q frequency domain shortcuts (Eq.(7)), which shorten selected high order propagation paths. This changes how signals propagate across GCN layers and helps reduce cross order frequency domain interference, as motivated by Prop.2. We will clarify this connection in Section 3.2.

W3

To further examine the applicability of coupled reconstruction principle, we extended it to the Link Prediction task and compared it with GAE[1] and AICoG[2]. When PuP_u = 8 and QiQ_i = 8, PRGCL achieves the best performance:

DatasetModelAUCAP
EnzymeGAE0.95580.9613
AICoG0.97960.9696
PRGCL0.98810.9919
IonChannelGAE0.93220.9191
AICoG0.97470.9735
PRGCL0.97390.9810

PRGCL achieves the best performance on Enzyme and remains competitive with AICoG on IonChannel, outperforming GAE on both datasets. These results suggest that the proposed coupled reconstruction principle can also benefit general graph learning tasks.

W4 and Q2

We analyze the computational complexity of PRGCL in Appendix A.4.2 and compare its training efficiency across three datasets in Section 4.2. To provide a broader comparison, we summarize the main computational complexity of four methods below:

ModelComputational Complexity
LightGCNO(K∣E∣d)\mathcal{O}(K\lvert E\rvert d)
SimGCLO(3K∣E∣d+∣B∣2d)\mathcal{O}(3K\lvert E\rvert d+\lvert B\rvert^2d)
DimCLO(3K∣E∣d+∣B∣2d)\mathcal{O}(3K\lvert E\rvert d+\lvert B\rvert^2d)
PRGCLO(K(∣E∣+∣E~(1)∣+∣E~(2)∣)d+3∣B∣2d)\mathcal{O}\left(K(\lvert E\rvert+\lvert\widetilde{E}^{(1)}\rvert+\lvert\widetilde{E}^{(2)}\rvert)d+3\lvert B\rvert^2d\right)

In PRGCL, view reconstruction and affinity shortcut construction are performed once per epoch, with their complexity analyzed in Eqs. (28)–(34). We further evaluate training efficiency and GPU memory consumption on Alibaba-iFashion, a large-scale dataset containing approximately 300K users, 81.6K items, and 1.61M interactions:

ModelAverage Time per Epoch (min)GPU Memory (GiB)
LightGCN0.6931.736
SimGCL1.0191.809
PRGCL1.0494.386

The results show that PRGCL achieves a training time close to SimGCL, while requiring more GPU memory for its reconstructed views. This indicates that PRGCL maintains competitive training efficiency on large-scale datasets despite the additional reconstruction overhead.

W5

We will revise Fig.2 by simplifying its layout, reducing unnecessary elements and line crossings, and optimizing the spacing to improve its clarity and readability.

W6 and Q3

To ensure a fair comparison, we evaluated all baselines under a unified PyTorch framework with a fixed random seed of 2026. Following previous classical studies [2, 3], we randomly split the datasets into training and test sets at a ratio of 8:2. For data preprocessing, we used the existing integer user and item IDs to construct a user-item bipartite graph from the training interactions and applied symmetric degree normalization. We adopted full-ranking evaluation with observed interactions masked and reported Recall@K and NDCG@K (K = 10, 20). All GCN-based models use 64-dimensional embeddings, 3 layers, and Xavier initialization. We used Adam with a learning rate of 0.001, an L2 coefficient of 0.0001, a batch size of 4096, and early stopping with a patience of 20. Experiments were conducted on an RTX 4090 GPU, except for DimCL on Alibaba-iFashion due to memory requirements. Some details are provided in Section 4.1, 4.4, and Appendix A.7.1, A.7.6.

Reference

[1] Kipf, T. N., & Welling, M. (2016). Variational graph auto-encoders. arXiv preprint arXiv:1611.07308.

[2] N. Nakis et al. Aitchison Embeddings for Learning Compositional Graph Representations. [C]// ICML 2026.

Reviewer_LoH3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To overcome the limitations of stochastic structural perturbations in GCL on user-item bipartite graphs, this paper proposes a coupled reconstruction principle and the corresponding PRGCL framework. Specifically, PRGCL constructs replacement and addition contrastive views by combining retained core interaction edges with top-qq weighted auxiliary shortcuts derived from learnable user/item affinity matrices, and jointly optimizes BPR, contrastive alignment, and Shannon-entropy affinity regularization losses. The method is validated on three sparse public datasets through baseline comparisons and ablation studies.

Critical strengths

(1) The paper provides a well-motivated summary of the three failure scenarios (structural reliability degradation, different signal interference, and propagation topology inconsistency) caused by traditional graph augmentations on bipartite graphs, and makes a structured effort to formalize these topological constraints via truncated path complexes (Propositions 1–3).

(2) The conceptual figures (Figures 1 and 2) clearly and concisely illustrate the authors' core claims and the dual-view path reconstruction workflow, making the overall intuition easy to follow.

Critical weaknesses

(1) In Eq. (8), the contrastive objective I(hu(0),hu(1),hu(2))\mathcal{I}(h_u^{(0)}, h_u^{(1)}, h_u^{(2)}) places the addition view hu(2)h_u^{(2)} into the denominator expectation (the negative contrastive pool), mathematically repelling it rather than aligning it. What exact loss is implemented in code?

(2) In Table 2, replacing learned reconstructed paths with random paths of the same form (+w/o AL) leads to only a marginal drop in Recall@10 on Douban-book (from 0.1247 to 0.1222) and Alibaba-iFashion (from 0.0837 to 0.0831). The contribution of the proposed path learning module appears minimal.

(3) Removing all contrastive terms (+w/o CL) causes an unexplained collapse in Recall@10 on Douban-book to 0.0580, far below the LightGCN baseline (0.1025 in Table 1). What causes this degradation?

(4) The third contrastive term in Eq. (8), I(hu(0),hi(0),hi(0))\mathcal{I}(h_u^{(0)}, h_i^{(0)}, h_i^{(0)}), is never ablated separately, leaving the true source of improvement unclear.

(5) While the paper compares against recent feature-mixing/dimension-augmentation baselines (e.g., MixRec, DimCL), it noticeably omits the most directly analogous baseline NCL [1] and the recognized structural denoising benchmark LightGCL [2], despite explicitly highlighting both in Section 2.

References:

[1] Lin Z, Tian C, Hou Y, et al. Improving graph collaborative filtering with neighborhood-enriched contrastive learning[C]//Proceedings of the ACM web conference 2022. 2022: 2320-2329.

[2] Cai X, Huang C, Xia L, et al. LightGCL: Simple yet effective graph contrastive learning for recommendation[J]. arXiv preprint arXiv:2302.08191, 2023.

Questions for authors

See weaknesses.

3 comments

We thank reviewer LoH3 and provide our responses below.

W1

We agree that the term "alignment" in Section 3.2.3 is imprecise and may suggest direct alignment between the two reconstructed views. Our intention is to leverage the complementary information provided by the replacement and addition views to learn more discriminative node representations. Specifically, the contrastive objective uses the original view representations as anchors, the corresponding replacement view representations as positive pairs, and the addition view representations as the contrastive candidate set. The implementation follows the contrastive formulation in Eq. (8), and the reported experimental results are obtained using this objective. We have revised the wording in manuscript to clarify its purpose as contrastive representation learning from reconstructed views. The mathematical formulation and experimental results remain unchanged.

Revised wording in Section 3.2.3:

  1. Section 3.2.3 title: Learning Discriminative Representations from Reconstructed Views
  2. "Given the two shortcut-augmented views, we next discuss how PRGCL leverages their complementary propagation information to learn discriminative node representations."
  3. "Beyond contrastive representation learning, we regularize the learnable frequency domain affinity matrices induced by P and Q so that each learnable row exhibits anisotropy after softmax normalization."

W2

In the +w/o AL variant, we retain the affinity shortcuts and replace the learned Top-q selection with random selection. Thus, this ablation evaluates the contribution of learned shortcut selection. To further examine the role of affinity shortcuts, we conducted an additional experiment that removes the affinity shortcuts and randomly reconstructs propagation paths. The results are shown below.

DatasetRecall@10NDCG@10Recall@20NDCG@20
Douban-book0.11720.15190.16730.1339
Alibaba-iFashion0.08070.04690.11600.0557
Tmall0.05020.04290.07980.0585

PRGCL consistently outperforms Random Paths across all datasets and metrics. These results demonstrate the importance of affinity shortcuts in path reconstruction, while the +w/o AL results show the additional contribution of learned Top-q selection. We will clarify these two ablation settings and include the additional results in the revised paper.

W3

Following the SimGCL implementation[1], our GCN encoder excludes the initial embeddings from the final layer aggregation, which differs from LightGCN. This setting may contribute to the larger performance drop of +w/o CL on Douban-book, while its performance remains comparable to LightGCN on the other two datasets. Moreover, removing contrastive learning prevents reconstructed path signals from contributing to representation optimization, leaving the model mainly guided by LBPR\mathcal{L}_{BPR}, while Lcar\mathcal{L}_{car} regularizes the affinity matrices. To further examine the effect of initial embeddings, we conduct an additional experiment by including initial embeddings in the +w/o CL variant.

DatasetRecall@10NDCG@10Recall@20NDCG@20
Douban-book0.10510.09730.15540.1103
Alibaba-iFashion0.06580.03900.10130.0526
Tmall0.04230.04020.06960.0503

W4

The user–item contrastive term in Eq.8 can be decomposed into alignment and uniformity terms. For normalized embeddings, we have sui=h^u⊤h^i=1−12∥h^u−h^i∥2s_{ui}=\hat h_u^\top\hat h_i=1-\frac{1}{2}\|\hat h_u-\hat h_i\|^2. Expanding the InfoNCE objective gives:

\mathcal{L}_{UI} &=-\mathbb{E}_{(u,i)} \log\frac{\exp(s_{ui}/\tau)} {\sum_{j\in\mathcal B}\exp(s_{uj}/\tau)}\\\\ &=\mathbb{E}_{(u,i)} \left[-\frac{s_{ui}}{\tau} +\log\sum_{j\in\mathcal B} \exp\left(\frac{s_{uj}}{\tau}\right)\right]\\\\ &=\underbrace{\frac{1}{2\tau}\mathbb{E}_{(u,i)} \\|\hat h_u-\hat h_i\\|^2}_{\text{Alignment}} \quad+\quad \underbrace{\mathbb{E}_{u} \log\sum_{j\in\mathcal B} \exp\left(-\frac{\\|\hat h_u-\hat h_j\\|^2}{2\tau}\right)}_{\text{Uniformity}}. \end{aligned}$$ The alignment term brings interacted users and items closer, while the uniformity term encourages the separation of candidate representations. These two effects provide additional regularization and help maintain a stable representation space. We further evaluated PRGCL using only UI-CL, with the results shown below. |Dataset|Variant|Recall@10|NDCG@10|Recall@20|NDCG@20| |-|-|-|-|-|-| |Alibaba-iFashion|PRGCL|0.0837|0.0483|0.1223|0.0586| ||UI-CL|0.0822|0.0465|0.1110|0.0564| |Tmall|PRGCL|0.0560|0.0537|0.0890|0.0677| ||UI-CL|0.0538|0.0519|0.0862|0.0661| As shown in the table, UI-CL achieves performance close to the full PRGCL on most metrics, indicating the importance of this regularization. The full PRGCL further improves recommendation performance, showing the additional benefits of learning from reconstructed views. We will include these results in the revised Section 4.3.

W5

We have further included NCL and LightGCL in our experiments under the same evaluation settings. The additional results are presented below.

DatasetModelRecall@10NDCG@10Recall@20NDCG@20
Douban-bookNCL0.11210.14850.16470.1539
LightGCL0.10690.13930.15700.1455
PRGCL0.12470.17210.18110.1757
Alibaba-iFashionNCL0.07020.03810.10090.0467
LightGCL0.07460.04110.10260.0485
PRGCL0.08370.04840.12230.0586
TmallNCL0.04590.04290.07500.0553
LightGCL0.05310.05080.08330.0637
PRGCL0.05600.05370.08900.0677

As shown in the table, PRGCL consistently achieves the best performance across all four metrics on the three datasets. These results further demonstrate the effectiveness of PRGCL compared with neighborhood-based contrastive learning and structural denoising methods. We will include these additional comparisons in the revised Section 4.2.

Reference

[1] Junliang Yu, et al. Are Graph Augmentations Necessary? Simple Graph Contrastive Learning for Recommendation. [C]// SIGIR 2022.

Thanks for the additional experiments. They address part of my initial concern, but two issues remain:

(1) For W1, if Eq. (8) uses hu(0)h_u^{(0)} as the anchor, hu(1)h_u^{(1)} as the positive pair in the numerator, and the addition view (hu(2)h_u^{(2)}) as the denominator candidate set, then for the target node uu, its own addition-view representation hu(2)h_u^{(2)} is absent from the numerator and only penalized in the denominator, even though G~(2)\tilde{\mathcal{G}}^{(2)} is designed to preserve all original edges. Also, Appendix A.6 (Eq. 44) defines the positive pair as −sim(zv(1),zv(2))-\text{sim}(z_v^{(1)}, z_v^{(2)}), not sim(hv(0),hv(1))\text{sim}(h_v^{(0)}, h_v^{(1)}). How should these two formulations be reconciled?

(2) For W4, the new results suggest that UI-CL (I(hu(0),hi(0),hi(0))\mathcal{I}(h_u^{(0)}, h_i^{(0)}, h_i^{(0)})) on the original graph alone already accounts for most of the Recall@10 gain over LightGCN (0.0822 vs. 0.0837 for full PRGCL on iFashion, compared to 0.0669 for LightGCN; and 0.0538 vs. 0.0560 on Tmall, compared to 0.0435 for LightGCN; note that Douban-book is omitted in W4), and even outperforms +w/o PR in Table 2 (0.0805 and 0.0522). Together with the small gap between PRGCL and +w/o AL in W2, the actual gain from path reconstruction over UI-CL seems quite limited.

Comments on the paper

1 comment
Area_Chair_jStNOpenReview ↗

Author- Reviewer Discussion

Dear Reviewers, the authors have submitted their response. Kindly review it and update your review where appropriate.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.