arxiv: 2605.05238 · v1 · submitted 2026-05-02 · 💻 cs.IR · cs.LG· cs.SI

Dynamic Graph with Similarity-Aware Attention Graph Neural Network for Recommender Systems

Aadarsh Senapati , Neha Kujur , Vivek Yelleti This is my paper

Pith reviewed 2026-05-09 17:52 UTC · model grok-4.3

classification 💻 cs.IR cs.LGcs.SI

keywords dynamic graph neural networksrecommender systemsuser similarity graphsattention fusioncollaborative filteringgraph transformermulti-view embeddings

0 comments

The pith

Dynamic reconstruction of four user similarity graphs during training improves recommender recall over static models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents DG-SA-GNN to move beyond fixed user-item graphs and single similarity measures in collaborative filtering. It builds four separate user similarity graphs from different functions, refreshes them at set points in training so they track the current embeddings, and combines their signals through attention layers to produce final recommendations. A sympathetic reader would care because this setup could let systems respond to gradual shifts in what users like without restarting from scratch each time. The reported results on MovieLens100K show higher recall than the LightGCN baseline.

Core claim

The central claim is that constructing four parallel user similarity graphs using Cosine, Jaccard, Discounted Pearson Correlation Coefficient, and IPIJ functions, processing each with its own UserGNN module, fusing the views in a Graph Transformer, and refining the embeddings with a CrossAttention module to item representations produces stronger recommendation performance when the graphs are rebuilt at scheduled epochs during training.

What carries the argument

Dynamic reconstruction of four user similarity graphs at scheduled epochs during training, each routed through a dedicated UserGNN, then fused by a Graph Transformer and refined by CrossAttention.

If this is right

User representations can track preference drift by incorporating fresh similarity relations as embeddings improve.
Multiple similarity functions together capture complementary aspects of user-user relations that a single measure misses.
Attention-based fusion lets the model weigh the four graph views differently for each user.
Mini-batch training with hard negative sampling keeps the dynamic updates scalable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same scheduled-reconstruction pattern could be tested on larger or sparser datasets to see how often the graphs need updating.
Adding a symmetric dynamic mechanism on the item side might strengthen the bipartite modeling.
The four chosen similarity functions might be replaced by learned similarity predictors without changing the overall architecture.

Load-bearing premise

Reconstructing the four user similarity graphs at scheduled epochs will stably adapt to the evolving embedding space without causing training instability or overfitting to the current embeddings.

What would settle it

Run DG-SA-GNN on MovieLens100K with the scheduled graph reconstruction disabled and check whether Recall@20 drops to or below the LightGCN baseline value of 0.162.

Figures

Figures reproduced from arXiv: 2605.05238 by Aadarsh Senapati, Neha Kujur, Vivek Yelleti.

**Figure 1.** Figure 1: Schematic diagram of the DG SA-GNN next as the model moves away from random initialisation. Frequent reconstruction during this phase ensures that the neighbourhood structure keeps pace with these fast-moving embeddings. As training progresses into the later phase, embeddings stabilise and change more gradually, so less frequent reconstruction is sufficient to maintain graph quality without incurring unne… view at source ↗

**Figure 2.** Figure 2: Graph Fusion operation in the DG-SA-GNN E. CrossAttention User-Item Alignment Item embeddings are enriched through interaction with user embeddings. The raw item embedding matrix I ∈ R |I|×d is augmented by the user context: I ′ = L2Norm I + R⊤U , (13) where R⊤U propagates user embeddings to items they have interacted with, providing items with a user-informed context. For each user u, the top-50 items … view at source ↗

**Figure 3.** Figure 3: Cross Attention Mechanism in the proposed DG-SA-GNN view at source ↗

**Figure 4.** Figure 4: Mini-Batch Training with Hard Negative Sampling followed in the current work view at source ↗

read the original abstract

Recommender systems are essential components of modern online platforms which presents personalized content in various domain. The traditional collaborative filtering methods depends on static user-item interaction graphs and a limited subset of similarity measures which fail to capture the changing nature of preferences of an individual. Recent graph neural network (GNN) based approaches focus on user-item bipartite graphs which do not use explicit user-user relational modelling and dynamic graph evolution during training. To address these limitations, this paper proposes a Dynamic Graph SimilarityAware Attention Graph Neural Network (DG-SA-GNN) framework that integrates dynamic user similarity graph construction with multi-similarity propagation and attention-based aggregation. The proposed architecture constructs four parallel user similarity graphs using Cosine, Jaccard, Discounted Pearson Correlation Coefficient (Discount PCC), and IPIJ similarity functions, each processed by a dedicated UserGNN module. A Graph Transformer fuses the four graph views, and a CrossAttention module refines user embeddings through interaction with item embeddings. Crucially, the graphs are reconstructed at scheduled epochs during training, enabling the model to adapt to the learned embedding space constituting the dynamic graph component. Mini-batch training with hard negative sampling improves scalability and convergence. Experiments on the MovieLens100K benchmark demonstrate that DG-SA-GNN achieves a Recall@20 of 0.162 and NDCG@20 of 0.065 which is better than the LightGCN baseline in recall. The results validate that dynamic multi-similarity graph construction coupled with attention-based fusion which produce recommendation performance

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This adds a multi-similarity dynamic graph layer to GNN recommenders but the reported gains rest on incomplete experiments with no baselines or stability checks.

read the letter

The main thing to know is that DG-SA-GNN builds four parallel user similarity graphs (Cosine, Jaccard, Discount PCC, IPIJ) from current embeddings, runs them through separate UserGNNs, fuses the views with a Graph Transformer, adds CrossAttention to item embeddings, and rebuilds the graphs at scheduled epochs. That specific mix of dynamic reconstruction plus multi-view fusion is not a routine extension of LightGCN-style models. The paper does a clear job explaining why static bipartite graphs miss evolving user preferences and why multiple similarity measures might help capture different aspects of user relations. The hard-negative mini-batch training is a sensible scalability choice. On MovieLens100K they give Recall@20 of 0.162 and NDCG@20 of 0.065 and say it beats LightGCN on recall. Those numbers are concrete enough to be checked later. The soft spots are the missing pieces that matter for trusting the central claim. The abstract supplies no LightGCN baseline numbers, no error bars, no significance tests, and the final sentence is incomplete. Only one small dataset appears. The dynamic reconstruction creates a direct loop where embeddings shape the graphs that then shape the embeddings; without details on reconstruction frequency, any damping, or regularization, it is unclear whether the adaptation stays stable or simply overfits to transient similarities. No ablations isolate the dynamic component from the extra attention layers. This work is aimed at people already working on GNN recommenders who are looking for multi-graph fusion experiments. A reader wanting a reliable new baseline or reproducible advance will not find enough here. If the full paper supplies the missing comparisons, ablations, and stability evidence, it deserves peer review in a specialized venue; otherwise the claims are too hard to evaluate.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes the DG-SA-GNN framework for recommender systems. It constructs four user similarity graphs (Cosine, Jaccard, Discounted Pearson Correlation Coefficient, and IPIJ) from user embeddings, processes each with a dedicated UserGNN, fuses them using a Graph Transformer, and refines user embeddings with item embeddings via CrossAttention. The graphs are dynamically reconstructed at scheduled epochs during training to adapt to the evolving embedding space. Experiments on MovieLens100K report Recall@20 of 0.162 and NDCG@20 of 0.065, claimed to exceed the LightGCN baseline.

Significance. Should the empirical improvements be substantiated with proper baselines, statistical analysis, and evidence that the dynamic reconstruction contributes without instability, this could represent a meaningful step in dynamic graph neural networks for recommendations by leveraging multiple similarity measures and attention mechanisms. The idea of scheduled graph updates is interesting but requires rigorous validation to establish its value over static multi-view approaches.

major comments (2)

[Experiments] The reported results claim that DG-SA-GNN outperforms LightGCN in recall with Recall@20 = 0.162 and NDCG@20 = 0.065, but the manuscript provides no numerical results for the LightGCN baseline, no standard deviations or error bars, and no mention of statistical significance testing. This makes the central performance claim difficult to evaluate.
[Proposed Method] The dynamic aspect relies on reconstructing the four similarity graphs at scheduled epochs, but the paper does not specify the epoch schedule, the fusion weights for the similarity functions, or any techniques (such as regularization or gradual updates) to mitigate potential instability or overfitting from the circular dependency where graphs are built from embeddings that are learned using the graphs.

minor comments (2)

[Abstract] The abstract concludes with an incomplete sentence: 'The results validate that dynamic multi-similarity graph construction coupled with attention-based fusion which produce recommendation performance'.
[Abstract] The abstract mentions improvement over LightGCN but does not include the baseline metrics for comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below and commit to revisions that strengthen the experimental reporting and methodological transparency without altering the core contributions.

read point-by-point responses

Referee: [Experiments] The reported results claim that DG-SA-GNN outperforms LightGCN in recall with Recall@20 = 0.162 and NDCG@20 = 0.065, but the manuscript provides no numerical results for the LightGCN baseline, no standard deviations or error bars, and no mention of statistical significance testing. This makes the central performance claim difficult to evaluate.

Authors: We agree that the current manuscript does not report the explicit numerical results for the LightGCN baseline, standard deviations across runs, or statistical significance tests. This omission weakens the evaluation of the claimed improvements. In the revised version, we will add a full results table including LightGCN (and other baselines) with mean Recall@20 and NDCG@20 values, standard deviations from multiple random seeds, and p-values from paired statistical tests to substantiate the performance gains. revision: yes
Referee: [Proposed Method] The dynamic aspect relies on reconstructing the four similarity graphs at scheduled epochs, but the paper does not specify the epoch schedule, the fusion weights for the similarity functions, or any techniques (such as regularization or gradual updates) to mitigate potential instability or overfitting from the circular dependency where graphs are built from embeddings that are learned using the graphs.

Authors: We acknowledge the lack of explicit details on the reconstruction schedule, fusion weights in the Graph Transformer, and stabilization techniques for the embedding-graph dependency. We will revise the method section to specify the epoch schedule (e.g., reconstruction every 10 epochs after initial warm-up), clarify how the four similarity views are fused (including any learned or fixed weights), and add discussion of mitigation strategies such as L2 regularization on embeddings, gradual update schedules, or auxiliary losses to reduce instability and overfitting risks. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical model proposal with explicit dynamic training procedure.

full rationale

The paper proposes the DG-SA-GNN architecture that explicitly defines dynamic user similarity graph reconstruction from evolving embeddings at scheduled epochs, followed by parallel UserGNN processing, Graph Transformer fusion, and CrossAttention. All central claims consist of measured Recall@20 and NDCG@20 on MovieLens100K versus the LightGCN baseline. No first-principles derivation, uniqueness theorem, or prediction is offered that reduces to its own inputs by construction; the dynamic component is an openly stated training heuristic rather than a self-referential tautology. No self-citations or ansatzes are invoked in the provided text to justify load-bearing steps.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical performance of a composite neural architecture whose behavior depends on many unstated training choices and the assumption that scheduled graph updates improve embeddings.

free parameters (2)

scheduled reconstruction epochs
The epochs at which the four similarity graphs are rebuilt are chosen by the authors and directly affect the dynamic component.
similarity function weights in fusion
How the Graph Transformer combines the four views is learned or set via hyperparameters not specified.

axioms (1)

domain assumption User similarity graphs constructed from current embeddings remain useful for recommendation when rebuilt periodically.
Invoked in the description of the dynamic graph component.

pith-pipeline@v0.9.0 · 5582 in / 1459 out tokens · 35315 ms · 2026-05-09T17:52:26.107727+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references

[1]

Recommender systems survey,

J. Bobadilla, F. Ortega, A. Hernando, and A. Guti ´errez, “Recommender systems survey,”Knowledge-Based Systems, vol. 46, pp. 109–132, 2013

2013
[2]

Hybrid recommender systems: Survey and experiments,

R. Burke, “Hybrid recommender systems: Survey and experiments,” User Modeling and User-Adapted Interaction, vol. 12, no. 4, pp. 331– 370, 2002

2002
[3]

Graph neural networks in recommender systems: A survey,

S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui, “Graph neural networks in recommender systems: A survey,”ACM Computing Surveys, vol. 55, no. 5, pp. 1–37, 2022

2022
[4]

LightGCN: Simplifying and powering graph convolution network for recommenda- tion,

X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “LightGCN: Simplifying and powering graph convolution network for recommenda- tion,” inProc. 43rd Int. ACM SIGIR Conf. Research and Development in Information Retrieval (SIGIR), 2020, pp. 639–648

2020
[5]

Temporal graph networks for deep learning on dynamic graphs,

E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. Bronstein, “Temporal graph networks for deep learning on dynamic graphs,” inProc. ICML 2020 Workshop Graph Representation Learning and Beyond, 2020

2020
[6]

Graph neural networks with similarity-aware attention for recommender systems,

A. Senapati, N. Kujur, A. Pujahari, and V . Yelleti, “Graph neural networks with similarity-aware attention for recommender systems,” in Proc. Int. Conf. (SA-GNN), 2025

2025
[7]

Graph convolutional neural networks for web-scale recom- mender systems,

R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale recom- mender systems,” inProc. 24th ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2018, pp. 974–983

2018
[8]

Knowledge-aware graph neural networks with label smoothness regularization for recommender systems,

H. Wang, F. Zhang, M. Zhang, J. Leskovec, M. Zhao, W. Li, and Z. Wang, “Knowledge-aware graph neural networks with label smoothness regularization for recommender systems,” inProc. 25th ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2019, pp. 968–977

2019
[9]

MixGCF: An improved training method for graph neural network-based recommender systems,

T. Huang, Y . Dong, M. Ding, Z. Yang, W. Feng, X. Wang, and J. Tang, “MixGCF: An improved training method for graph neural network-based recommender systems,” inProc. 27th ACM SIGKDD Conf. Knowledge Discovery and Data Mining, 2021, pp. 665–674

2021
[10]

A deeper graph neural network for recommender systems,

R. Yin, K. Li, G. Zhang, and J. Lu, “A deeper graph neural network for recommender systems,”Knowledge-Based Systems, vol. 185, p. 105020, 2019

2019
[11]

Multi-behavior graph neural networks for recommender system,

L. Xia, C. Huang, Y . Xu, P. Dai, and L. Bo, “Multi-behavior graph neural networks for recommender system,”IEEE Trans. Neural Networks and Learning Systems, vol. 35, no. 4, pp. 5473–5487, 2022

2022
[12]

Inductive representation learning on temporal graphs,

D. Xu, C. Ruan, E. Korpeoglu, S. Kumar, and K. Achan, “Inductive representation learning on temporal graphs,” inProc. Int. Conf. Learning Representations (ICLR), 2020

2020
[13]

Graph neural networks for social recommendation,

W. Fan, Y . Ma, Q. Li, Y . He, E. Zhao, J. Tang, and D. Yin, “Graph neural networks for social recommendation,” inProc. The World Wide Web Conf. (WWW), 2019, pp. 417–426

2019
[14]

Survey of similarity functions on neighborhood-based collaborative filtering,

H. Khojamli and J. Razmara, “Survey of similarity functions on neighborhood-based collaborative filtering,”Expert Systems with Appli- cations, vol. 185, p. 115482, 2021

2021
[15]

DGRec: Graph neural network for recommendation with diversified embedding generation,

L. Yang, S. Wang, Y . Tao, J. Sun, X. Liu, P. S. Yu, and T. Wang, “DGRec: Graph neural network for recommendation with diversified embedding generation,” inProc. 16th ACM Int. Conf. Web Search and Data Mining, 2023, pp. 661–669

2023
[16]

TransGNN: Harnessing the collaborative power of transformers and graph neural networks for recommender systems,

P. Zhang, Y . Yan, X. Zhang, C. Li, S. Wang, F. Huang, and S. Kim, “TransGNN: Harnessing the collaborative power of transformers and graph neural networks for recommender systems,” inProc. 47th Int. ACM SIGIR Conf. Research and Development in Information Retrieval, 2024, pp. 1285–1295

2024
[17]

PONE-GNN: Integrating positive and negative feedback in graph neural networks for recommender systems,

Z. Liu, C. Wang, S. Zheng, C. Wu, K. Zheng, Y . Song, and N. Mou, “PONE-GNN: Integrating positive and negative feedback in graph neural networks for recommender systems,”ACM Trans. Recommender Systems, vol. 3, no. 2, pp. 1–23, 2025

2025
[18]

SoURA: A user-reliability- aware social recommendation system based on graph neural network,

S. Dawn, M. Das, and S. Bandyopadhyay, “SoURA: A user-reliability- aware social recommendation system based on graph neural network,” Neural Computing and Applications, vol. 35, no. 25, pp. 18533–18551, 2023

2023
[19]

CaGE: A causality-inspired graph neural network explainer for recommender systems,

S. Yu, Y . Li, S. Wang, T. Tang, Q. Zhang, J. Zhou, L. Ivan, and F. Xia, “CaGE: A causality-inspired graph neural network explainer for recommender systems,”ACM Trans. Information Systems, vol. 43, no. 6, pp. 1–29, 2025

2025
[20]

Graph neural networks with deep mutual learning for designing multi-modal recommendation systems,

J. Li, C. Yang, G. Ye, and Q. V . H. Nguyen, “Graph neural networks with deep mutual learning for designing multi-modal recommendation systems,”Information Sciences, vol. 654, p. 119815, 2024

2024
[21]

Vertical federated graph neural network for recommender system,

P. Mai and Y . Pang, “Vertical federated graph neural network for recommender system,” inProc. Int. Conf. Machine Learning (ICML), 2023, pp. 23516–23535

2023
[22]

Macro graph neural networks for online billion-scale recommender systems,

H. Chen, Y . Bei, Q. Shen, Y . Xu, S. Zhou, W. Huang, H. Fang, W. Senzhang, and X. Huang, “Macro graph neural networks for online billion-scale recommender systems,” inProc. ACM Web Conf. 2024, 2024, pp. 3598–3608

2024
[23]

BPR: Bayesian personalized ranking from implicit feedback,

S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “BPR: Bayesian personalized ranking from implicit feedback,” inProc. 25th Conf. Uncertainty in Artificial Intelligence (UAI), 2009, pp. 452–461

2009