Open research questions in Advanced Graph Neural Networks

What the literature leaves open

• While multi-relational GNNs (MR-GNNs) have been adopted for representation learning and recommendation, their application for user modeling remains unexplored.

  Sequential Learner Modeling Using Multi-Relational Graph Convolutional Networks · 2026

• Although prior studies have explored specific mitigation strategies, existing evidence remains fragmented across degradation types, datasets, tasks, and model families, leaving TAG robustness insufficiently understood.

  OpenRTAG: A Comprehensive Benchmark for Robust Text-Attributed Graph Learning under Data Quality Degradation · 2026

• We discover that the representa- tion information in the latent space remains under-explored. Additionally, we intend to investigate self-supervised learning techniques to effec- tively mitigate the challenges posed by covariate distribution shifts, particularly in scenarios where labeled data is scarce or unavailable.

  Graph data augmentation with contrastive learning on covariate distribution shift · 2026 · DOI

• We introduced TalentFlow, which reformulates internal talent mobility as next-role link prediction over a heterogeneous talent graph and pairs a relational GNN encoder–decoder with dual attribute and structural explanations. On a benchmark anchored to real HR data with a transparent, socially grounded mobility mechanism, relational message passing outperformed strong non-graph and homogeneous-graph baselines, and ablations identified the managerial/peer relation—the carrier of social capital—as the most important structural channel, empirically supporting the view that mobility is a relational phenomenon. The explanations expose interpretable drivers spanning both employee attributes and network structure. Future work includes validation on proprietary internal-mobility data, richer temporal modeling of heterogeneous encoders such as attention over meta-paths,, and integrating fairness constraints and debiasing, so that relational mobility recommendations are not only accurate and explainable but also equitable.

  TalentFlow: Explainable Internal Talent Mobility Prediction Using Graph Representation Learning · 2026 · DOI

• neural networks[C]//Proceedings of the AAAI conference on artificial intelligence: 33. 2019: 346-353. graph with WANG M, LIU X, WU X. Visual Classification by $\ell _1$ -Hypergraph Modeling[J/OL]. IEEE Transactions on Knowledge and Data Engineering, 2015, 27(9): 2564- 2574. ZHANG Z, LIN H, GAO Y, et al.. Dynamic hypergraph structure learning.[C]//IJCAI. 2018: 3162-3169. JI R, CHEN F, CAO L, et al.. Cross-modality microblog sentiment prediction via bi-layer multimodal hypergraph learning[J]. IEEE Transactions on Multimedia, 2018, 21(4): 1062-1075. 10 EAI Endorsed Transactions on e-Learning | Volume 10 | 2024 | RETRACTED A Review of Hypergraph Neural Networks TSUYUZAKI K, ISHII M, NIKAIDO I. Uncovering hypergraphs of cell-cell interaction from single cell RNAsequencing data[J]. BioRxiv, 2019: 566182. YU L, SHEN X, JIANG X, et al..

  Retraction Notice: A Review of Hypergraph Neural Networks · 2026 · DOI

• attention with networks[C]//Proceedings of the 2021 SIAM international conference on data mining (SDM). SIAM, 2021: 82-90. hypergraph DING M, LIN X, ZENG B, et al.. Hypergraph neural networks with attention mechanism for session-based recommendation[C]//Journal of Physics: Conference Series: 2082. IOP Publishing, 2021: 012007. LIU Q, ZENG Y, MOKHOSI R, et al.. STAMP: shortterm attention/memory priority model for session-based the 24th ACM recommendation[C]//Proceedings of SIGKDD conference on knowledge international discovery & data mining. 2018: 1831-1839. WU S, TANG Y, ZHU Y, et al.. Session-based recommendation neural networks[C]//Proceedings of the AAAI conference on artificial intelligence: 33. 2019: 346-353. graph with WANG M, LIU X, WU X. Visual Classification by $\ell _1$ -Hypergraph Modeling[J/OL]. IEEE Transactions on Knowledge and Data Engineering, 2015, 27(9): 2564- 2574. ZHANG Z, LIN H, GAO Y, et al.. Dynamic hypergraph structure learning.[C]//IJCAI. 2018: 3162-3169. JI R, CHEN F, CAO L, et al.. Cross-modality microblog sentiment prediction via bi-layer multimodal hypergraph learning[J]. IEEE Transactions on Multimedia, 2018, 21(4): 1062-1075. 10 EAI Endorsed Transactions on e-Learning | Volume 10 | 2024 | RETRACTED A Review of Hypergraph Neural Networks TSUYUZAKI K, ISHII M, NIKAIDO I. Uncovering hypergraphs of cell-cell interaction from single cell RNAsequencing data[J]. BioRxiv, 2019: 566182. YU L, SHEN X, JIANG X, et al..

  Retraction Notice: A Review of Hypergraph Neural Networks · 2026 · DOI

• for that enhance session-based Short-Term Attention/Memory Priority (STAMP) [40] and Session-based Recommendation with Graph Neural Networks (SR-GNN) [41] are two advanced baseline models session-based recommendations by introducing attention mechanisms and GNNs, respectively. Table 2 records the commonly used datasets recommendation, YooChoose, and Diginetica. As shown in Table 3, hypergraph neural network models SHARE, DHCN, and S2-DHCN outperform SR-GNN in terms of the mean reciprocal ranking (MRR@K) on YooChoose1/4, with improvements of 0.22%, 4.83%, and 8.24%, respectively. On Diginetica, the MRR@K values improved by 0.46%, 12.99%, and 13.35%, respectively. HGNNA also outperformed STAMP, further proving the significant effectiveness of hypergraph neural network models in this task. 4.2. Clustering Network clustering involves partitioning network vertices into several clusters, with the goal of making vertices within the same cluster closely connected while having weaker connections between different clusters. This clustering structure is widespread in fields such as bioinformatics, computer science, physics, and sociology, and it holds significant importance. In fact, hypergraph partitioning can also achieve clustering objectives, so in many studies, hypergraph partitioning algorithms are classified as clustering algorithms. Research on custom edge weight functions is also an important method in hypergraph clustering. Additionally, projecting hypergraph structure information and node attribute information into vector space through hypergraph representation methods, and then clustering based on spatial distance, is also an effective approach for hypergraph clustering. This method preserves graph structure characteristics while also considering node attribute information.

  Retraction Notice: A Review of Hypergraph Neural Networks · 2026 · DOI

• information, text DHCN, SHARE, and HGNN are hypergraph neural network models based on session-based recommendations. DHCN is a dual-channel hypergraph convolutional network that combines hypergraphs and line graphs, where hypergraph convolution captures higherorder relationships at the item level, and line graph convolution learns relationships at the session level. S2- DHCN is a variant of DHCN that incorporates selfsupervised learning. SHARE models each session as a hypergraph, with items as vertices and hyperedges linking all items within a contextual window. It employs a hypergraph attention mechanism to assess the significance of items to sessions and the influence of sessions on items. HGNNA, a recommendation model based on hypergraph neural networks and attention mechanisms, first utilizes a hypergraph neural network to learn item associations. It then applies a self-attention mechanism to aggregate session information and a graph attention mechanism to reveal the relevance between sessions. Table 2.

  Retraction Notice: A Review of Hypergraph Neural Networks · 2026 · DOI

• In this paper, we systematically reviewed the concepts, definitions, methods, and applications of hypergraph learning in different fields. First, we introduced the definition of hypergraphs and their related symbols, clarifying the unique advantages of hypergraphs in handling higher-order relationships, making them an important tool for modeling complex data structures. We then explored the main current methods of hypergraph A Review of Hypergraph Neural Networks learning, particularly spectral analysis methods and neural network methods. By categorizing and comparing expansion-based and non-expansion-based methods, we revealed their applicability and limitations in different application scenarios. On the application level, we discussed in detail the widespread use of hypergraphs in recommendation systems, clustering, node classification, visual tasks, and biological networks, demonstrating the strong capability and flexibility of hypergraphs in handling complex and high-dimensional data.

  Retraction Notice: A Review of Hypergraph Neural Networks · 2026 · DOI

• This paper proposed CAGNN-Sum, a novel extractive summarization model based on a customized heterogeneous graph neural network (HGNN) designed to leverage the rich semantic structure of long-form texts. By modeling sentences, named entities, and BERTopic-derived latent topics as distinct node types and linking them through four typed semantic edges—resolving the LDA/SBERT incompatibility that affects all prior heterogeneous summarizers—CAGNN-Sum performs type-aware message passing that captures higher-order semantic representations. The model is further strengthened by ML-GAP with input-conditioned scalar gating and relation-aware aggregation. Experiments on arXiv and BillSum demonstrate consistent improvements of more than 1.2-−1.5 ROUGE-1 points over the best baseline G-Seek-2, with BERTScore confirming that gains reflect semantic quality rather than superficial n-gram overlap. Cross-domain transfer (arXiv→BillSum, ROUGE-1 = 42.07) confirms structural generalization. Retrained ablations with Wilcoxon significance testing establish that ML-GAP ( p<0.001) and encoder depth ( p<0.001) are architecturally critical, while the input-conditioned gates provide graceful degradation when individual auxiliary node types are removed. The monotonic increase in graph-based advantage with document length (+5.33 points for >100 sentences) confirms that heterogeneous topology succeeds where flat retrieval degrades. Regarding future work: (i) confidence-aware graph construction and noise-aware attention weighting will reduce error propagation from inaccurate preprocessing modules; (ii) NLIbased faithfulness metrics such as SummaC and dedicated human evaluation will be applied for validated hallucination assessment; (iii) the framework will be extended to multilingual and multi-document settings; (iv) targeted ablations with increased statistical power will isolate the marginal contribution of each auxiliary node type under varying data regimes; and (v) external knowledge graphs and domain-specific ontologies will be incorporated to enhance semantic relations and robustness in specialized domains. Author Contributions S.W.A.B. contributed to the study conception and design, performed material preparation, data collection, and analysis, and wrote the first draft of the manuscript. Y.Y. provided supervision, resources, and reviewed the manuscript. All authors read and approved the final manuscript.

  Context-aware heterogeneous graph neural networks with attention-based pooling for long-document summarization · 2026 · DOI

• Kairouz P, McMahan HB, Avent B, Bellet A, Bennis M, Bhagoji AN, Bonawitz K, Charles Z, Cormode G, Cummings R et al (2021) Advances and open problems in federated learning.

  Parameter broadcasting on participant graph for federated heterogeneous graph learning · 2026 · DOI

• Carlota Delgado Vera * and Andrea Sinche-Guzmán Facultad de Ciencias Agrarias, Universidad Agraria del Ecuador, Guayaquil, Ecuador Artificial intelligence (AI) has achieved extraordinary progress in recent years, yet this progress reveals a deep educational and epistemic imbalance. Neural architectures have mastered prediction but often obscure the grounds of their outputs. This Perspective argues that knowledge graphs (KGs) are more than a technical advance: they are an intellectual bridge between symbolic and neural paradigms, and a pedagogical opportunity to reform university-level AI curricula. The true frontier of explainable AI is educational, not only technological. By reintroducing symbolic reasoning into advanced AI curricula and professional training, we can prepare students who design, build, deploy, and evaluate AI systems to understand and justify system outputs. The focus is higher education for future developers, deployers, and auditors of AI systems, not general AI literacy for everyday users of AI tools. Through historical analysis, theoretical synthesis, and pedagogical reflection, we show that knowledge graphs are not only computational infrastructures but also catalysts for cognitive transformation in how we teach, learn, and conceptualize intelligence. KEYWORDS AI education, explainable AI (XAI), knowledge graphs, neuro-symbolic artificial intelligence, symbolic reasoning Introduction: the educational blind spot of artificial intelligence In the last decade, deep learning has become the dominant language of artificial intelligence (AI). The proliferation of high-level frameworks such as TensorFlow and PyTorch has democratized access to powerful models (Abadi et al., 2016; Paszke et al., 2019), and pretrained model repositories and low-code interfaces have further reduced the need to implement models from first principles. However, this success has created a cognitive asymmetry. In many university computer science, data science, and engineering programs, students learn how to train models that predict with high precision but cannot readily explain their logic. The symbolic tradition of AI, once central to the identity of the field, has been marginalized to the point of invisibility. Systems such as MYCIN, an expert system for infectious-disease diagnosis, and DENDRAL, a system for inferring molecular structures from mass spectrometry data, once embodied a philosophy of intelligence that prized explanation and accountability (Lindsay, 1980; Shortliffe, 2012). Today, many curricula privilege efficiency and performance metrics over reasoning and transparency. This neglect has consequences that reach far beyond pedagogy.

  Knowledge graphs as pedagogical bridges for symbolic reasoning in hybrid AI systems: a perspective · 2026 · DOI

• 5.1 Conclusion In this paper, we carry out systematic research on bitcoin transaction anomaly detection task and propose several inno- vative methods: firstly, to meet the practical requirements, we define the dynamic heterogeneous graph semi-supervised bitcoin anomaly detection task and design the Bi-directional Fusion Heterogeneous Graph Network (BF-HGN) to construct the basic framework. Second, in feature extraction, we improve upon RGCN to construct EvolveRGCN and combines EvolveGCN to design a gradual scheme. It also introduces LSTM to capture temporal features and deeply mines dynamic features through a fusion strategy. Further, we propose the Multi-type Feature Fusion Extractor. This improves the dynamic relationship modeling capability by capturing the upper and lower time-point subgraph associations. Lastly, we address the class imbalance problem caused by unlabeled anomalous samples by designing Class-balanced Classifiers. These classifiers balance the training data class distribution by generating pseudo-abnormal nodes constrained by AA and AFSR loss function. 5.2 Outlook Future research can be extended to a broader range of financial transaction scenarios, thereby strengthening risk pre- vention and control capabilities. Further exploration of the optimization space of feature extraction and fusion strate- gies reveals potential associations in complex data and injects richer semantic information into the model. Meanwhile, continuous efforts should be made to refine the optimization path of loss functions to improve the generation quality of pseudo-anomalous nodes, so as to promote the security and stability of anomaly detection technologies in Bitcoin transactions and related fields. In addition to technical advancements, future studies should incorporate regulatory, ethical, and societal considerations into the design of anomaly detection systems. Inspired by the sociotechnical framework proposed by Rahman et al. [62], responsible and trustworthy FinTech development can be better supported in blockchain transaction surveillance, particularly with respect to regulatory compliance, transparency, and social accountability.

  Bidirectional fusion heterogeneous graph networks for semi-supervised Bitcoin transaction anomaly detection in dynamic transaction graphs · 2026 · DOI

•  Aligning cross-modal heterogeneous feature spaces.  Maintaining semantics of modality in aggregation and adaptive graph structure learning [54].  Reducing balance and noise across the data sources. ISSN: 0363-8057 218 www.gradiva.it Volume 65 | Issue 05 | May 2026 DOI: 10.5281/zenodo.20303701 Finding methods to reach principled multimodal fusion without compromising on interpretability and stability is an important research direction. 6.5 Toward Unified and Trustworthy Multilayer Learning The overlap in explainability, self-supervision, temporal modeling, and multimodal reasoning leads to next-generation graph intelligence architectures marked by:  Clear-cut and responsible decision-making.  Learning when there is a data bottleneck and when labels are scarce.  Time -dependent flexibility and resilience.  Cross-domain and multimodal ability to reason. The described integrated systems should become the foundation of high-tech uses in smart cities, healthcare analytics, financial intelligence, social network analysis, and cyber- physical infrastructures. 7. CONCLUSION This article has given a broad, theoretically based overview of how machine learning algorithms can be applied to multilayer graphs, synthesizing insights into approaches to multilayer networks science, geometric deep learning and recent graph neural applications. We have built a unified modeling paradigm that generalizes classical single-layer GNN models to cohesive, heterogeneous, and interdependent systems of relations by combining multilayer representations based on tensors with message passing models. In addition to architectural synthesis, this review offered formal theoretical expression of multilayer propagation dynamics such as expressivity amplification by inter-layer coupling, stability conditions of spectral radius and formal expression of convergence behavior and over-smoothing as well as over-squashing. These findings help to explain the trade-offs inherent to representational power and stability in multilayer systems in principle and provide a principled basis on which any future model should be developed. We have shown the development of multilayer graph learning using systematic benchmarking discussion, as well as, bibliometric trend analysis, as evolving the paradigm of network learning models. This interdisciplinary growth of the field is also emphasized by the convergence of robustness, scalability, federated learning, and multimodal reasoning. transformer-based and self-supervised theory to Future research directions involve:  Scalable multilayer training in a large-scale and streaming setup.  Spectral regularization methods to reduce cross-layer smoothing and squashing.  Explainable multilayer attention models based on information theory.  Bases of heterogeneous and multimodal graph systems.  Hands-on implementation in transportation, biological, financial and cyber-physical systems. With the growth of multilayer interactions in large scientific and industrial systems, multilayer graph learning will become a prime component in intelligent information processing. The theoretical rigor and innovation of architecture by integrating both ISSN: 0363-8057 219 www.gradiva.it Volume 65 | Issue 05 | May 2026 DOI: 10.5281/zenodo.20303701 theoretically and educationally, through this review, will be used as a reference to the next generation of multilayer machine learning research.

  Comprehensive Review of Machine Learning on Multilayer Graphs: Theory, Architectures, Benchmarks, and Emerging Intelligent Paradigms · 2026 · DOI

•  Modeling inter-layer interactions (asynchronous).  Dealing with concept drift and time errors.  Computational efficiency with large-scale dynamics. Theoretical and engineering There is a significant theoretical and engineering challenge in developing stable temporal propagation operators on multilayer GNNs. 6.4 Multimodal and Cross-Domain Graph Learning [34], [45], [46] Heterogeneous data modalities, such as text, images, relation, and sensor are becoming more integrated with modern applications.

  Comprehensive Review of Machine Learning on Multilayer Graphs: Theory, Architectures, Benchmarks, and Emerging Intelligent Paradigms · 2026 · DOI

•  Scalable explanatory strategies of deep and heavily interconnected multilayer GNNs.  Stable cross-layer attribution and explanation.  Finding a balance between interpretability and predictive accuracy and complexity of the model. It is an open research question to develop principled explanation frameworks theory- grounded and computationally efficient. 6.2 Self-Supervised and Contrastive Learning [21], [43], [44] Multilayer environments are sparsely labeled, which leads to adoption of self-supervised learning (SSL) paradigms, which make use of inherent graph structure.

  Comprehensive Review of Machine Learning on Multilayer Graphs: Theory, Architectures, Benchmarks, and Emerging Intelligent Paradigms · 2026 · DOI

• in enterprise RAG has become a dominant pattern for grounding LLMs corpora, but baseline implementations face well-documented limitations in long, structured, multilingual, and governance- constrained environments. Knowledge graphs and GraphRAG-style retrieval provide a complementary path by introducing explicit structure that can support multi-hop reasoning, corpus-level summarization, and enhanced traceability. At the same time, enterprise feasibility requires careful attention to graph construction cost, noise, and operational constraints. This paper has presented the context and state of the art motivating the KNOWL project, with emphasis on multilingual English/Spanish proposal workflows over PDF corpora, governance under the EU AI Act and GDPR, and the practical landscape of open-weight LLM families. The next paper will describe the project’s specific methods and will report empirical evaluation results under a rigorously governed, non- fabricated experimental protocol, with accuracy as the primary objective.

  KNOWL: Retrieval-augmented generation and hybrid knowledge-graph retrieval for more reliable enterprise LLM applications — Context and state of the art · 2026 · DOI

• This paper has several limitations: • Rapidly evolving literature: RAG/GraphRAG and open-weight LLMs evolve quickly; some cited sources are preprints and may be updated. • Domain specificity: proposal/bid workflows are used as the motivating domain; generalization to other enterprise domains (legal, HR, code) may require different retrieval/graph strategies. • Implement observability: retrieval sets, prompts, outputs, latency, and user feedback for continuous evaluation. log 10.2. When to consider GraphRAG GraphRAG-style indexing becomes attractive when: • Questions are corpus-level (“global”) or multi-hop. • Document structure relationships matter clauses). and (templates, cross-document repeated • You need more explainable navigation through entities and evidence chains.

  KNOWL: Retrieval-augmented generation and hybrid knowledge-graph retrieval for more reliable enterprise LLM applications — Context and state of the art · 2026 · DOI

• Future work will explore the adaptive path weight learning through meta-learning, the extension to temporal graph dynamics where KG edges are introduced sequentially and combining tissue-specific expression data to contextualise disease modules are not limited by structural KG topology alone.

  MPRDR: A Multi-Path Relational Drug Repurposing Framework Grounded in Graph-Theoretic Principles · 2026 · DOI

• In this paper, we introduced the Laplacian Associative- Projective Neural Network (LAPNN) and evaluated its performance in human–machine collaborative industrial tasks using spatio-temporal data from a modified version of the collaborative action (CoAx) dataset. Experimental results show that the LAPNN achieves strong performance Page 13 of 15 35 in graph-level classification and robust cross-subject gener- alization under reduced-data conditions. While graph-level performance is comparable to that of the THGNN baseline, with a single outlier subject disproportionately affecting aggregated metrics, the LAPNN consistently demonstrates data-efficient generalization across unseen human subjects. In addition to efficiency, the LAPNN contributes a cognitive alternative to weight-based deep learning approaches. By disentangling structural graph knowledge from situational context, the LAPNN enables analogical- style reasoning and the potential for generalization across heterogeneous industrial domains. Learned spatiotempo- ral interaction motifs can be transferred and recombined with novel human subjects or physical components with- out retraining over old datasets. Given this disentangle- ment capability, VSAs could enable new decentralized approaches to multi-modal, multi-agent AI proposed by other researchers for HCPS. In contrast to other VSAs employed in graph learning, the LAPNN offers a holistic approach, capable of learning, transferring, and reasoning over diverse human–machine interactions. Future work can be focused on overcoming the current limitations of the LAPNN, using methods that align with the model architecture, such as learning causal behavior from the temporal interactions without relying on correlation matrices, and enabling real-time functional node feature predictions. Other directions include integration of neuro- morphic hardware such as field-programmable gate arrays to employ parallel associative recall, as well as human health and EEG signals in order to further harness human context. Overall, these directions align with the goals of Industry 5.0, emphasizing human-centrism, resilience, and sustain- ability, positioning the LAPNN as a promising approach to cognitive HCPS design. Supplementary Information The online version contains supplementary material available at h t t p s : / / d o i . o r g / 1 0 . 1 0 0 7 / s 1 2 5 5 9 - 0 2 6 - 1 0 5 7 4 - x. Author Contributions CG created the hardware and software architec- ture, performed the experiments, curated the figures/tables, and wrote the manuscript. JVD and JOM supervised the experimental design, en- sured the statistical soundness of the evaluation, advised and contribut- ed to the interpretation, positioning, and editing of the manuscript. All authors approved the final version and are accountable for the integrity and accuracy of the work. Funding JOM wants to recognize that this research has been par- tially supported by the Ministerio de Ciencia e Innovación of Spain (Grant Ref. PID2022-137748OB-C31 funded by MCIN/ AEI/10.13039/501100011033) and ”ERDF A way of making Europe”. JVD further acknowledges funding from the Hochschule Heilbronn and the Dieter Schwarz Stiftung through the program ”Innovative Son- derprojekte im Bereich Forschung und Bildung 2025” to the HyQCA Project, which supports strategic, high-excellence research initiatives at the interface of digital technologies, AI, and quantum computing science. Cognitive Computation (2026) 18:35 1 3 35 Page 14 of 15 Data Availability The original CoAx dataset from [23] is available at h t t p s : / / d l g m t z s . g i t h u b . i o / d a t a s e t - c o a x /. Code Availability The complete LAPNN model and experiment source code are available to reviewers as supplementary files and will be made public upon manuscript acceptance. In such cases, a link to the GitHub repository will be provided in this section.

  A Brain-Inspired Model for Efficient Graph Learning in Human Cyber Physical Networks · 2026 · DOI

• The paper does not provide empirical validation or real-world applications of the proposed MultiSoft MultiExpert Graph framework beyond the illustrative supplier-hub subnetwork example.

  Note for Soft MultiExpert Graph and MultiSoft MultiExpert Graph · 2026 · DOI

• Future work includes examining further extensions based on Fuzzy Graphs, Neutrosophic Graphs, Quadri-Partitioned Neutrosophic Graphs, HyperGraphs, Neutrosophic HyperGraphs, SuperHyperGraphs, Rough Graphs, and Plithogenic Graphs.

  Note for Soft MultiExpert Graph and MultiSoft MultiExpert Graph · 2026 · DOI

• While transfer principles are well-studied in the context of dense neural networks in language and vision tasks, they remain comparatively under-explored for graph neural networks (GNNs).

  Hyperparameter Transfer in Graph Neural Networks · 2026

• However, most existing works concentrate on node-level and edge-level tasks, while graph-level tasks, which require capturing more complex structural and feature information, remain relatively underexplored.

  GLIP: Graph and LLM Joint Pretraining for Graph-Level Tasks · 2026

• Despite ongoing efforts by the authors of these models to create specialized datasets, there remains a lack of standardized, curated databases tailored to specific tasks.

  A survey of models composed of Graph Neural Networks and Large Language Models for Molecular Science · 2026 · DOI

Most-cited papers in Advanced Graph Neural Networks

• Graph neural networks · Nature Reviews Methods Primers · 2024 · 367 citations
• A Comprehensive Survey on Deep Graph Representation Learning · Neural Networks · 2024 · 283 citations
• Gated Graph Recurrent Neural Networks · IEEE Transactions on Signal Processing · 2020 · 175 citations
• Exploring the Potential of Large Language Models (LLMs)in Learning on Graphs · ACM SIGKDD Explorations Newsletter · 2024 · 170 citations
• Stability Properties of Graph Neural Networks · IEEE Transactions on Signal Processing · 2020 · 154 citations
• Permutation Equivariant Graph Framelets for Heterophilous Graph Learning · IEEE Transactions on Neural Networks and Learning Systems · 2024 · 137 citations
• Graph Filters for Signal Processing and Machine Learning on Graphs · IEEE Transactions on Signal Processing · 2024 · 119 citations
• Generative AI-Driven Semantic Communication Networks: Architecture, Technologies, and Applications · IEEE Transactions on Cognitive Communications and Networking · 2024 · 119 citations
• Reconstructed Graph Neural Network With Knowledge Distillation for Lightweight Anomaly Detection · IEEE Transactions on Neural Networks and Learning Systems · 2024 · 118 citations
• Multi-view graph convolutional networks with attention mechanism · Artificial Intelligence · 2022 · 114 citations

Most recent work

• One model connects all graphs: Towards training one unified model for multi-domain graph pre-training using adaptive vector quantization · Information Fusion · 2026
• Graph intelligence for IoT and CPS security: From connectivity patterns to cyber-physical reasoning · Computer Science Review · 2026
• GNS2CCL: A graph network semi-supervised concept-cognitive learning model for node classification · Pattern Recognition · 2026
• FedLDP: Model-heterogeneous Federated Graph Learning via Logits Decoupling and Personalization · Knowledge-Based Systems · 2026
• A spectral heterogeneous graph neural network with multi-sensor fusion for machine fault diagnosis · Engineering Applications of Artificial Intelligence · 2026
• Unsupervised dual-masked graph autoencoder feature learning based on hyperspectral image classification · Geo-spatial Information Science · 2026
• KNOWL: Retrieval-augmented generation and hybrid knowledge-graph retrieval for more reliable enterprise LLM applications — Context and state of the art · Zenodo (CERN European Organization for Nuclear Research) · 2026
• Note for Soft MultiExpert Graph and MultiSoft MultiExpert Graph · Integra: Journal of Integrated Mathematics and Computer Science · 2026
• Leveraging a graph attention network-based academic recommendation framework for Indian higher education institutions · Journal of Artificial Intelligence and Technology · 2026
• Dismantling complex networks based on higher-order graph neural network · Communications Physics · 2026

Related topics in Computer Science