Graph neural networks (GNNs) have demonstrated promising results in enhancing machine learning and artificial intelligence techniques by addressing graph-related problems, including graph categorization, edge prediction, and node prediction. Despite the rapid growth of the field, a structured synthesis of recent application-oriented GNN research is necessary. This systematic literature review analyzes 363 peer-reviewed, application-oriented GNN studies published from 2019 through February 2026 from four scientific repositories. These studies are organized by GNN model variant, (including single and multimodel approaches), and four classification task categories: graph, link, node, and node and edge combined. In addition, the review assesses the use of learning settings, graph types, loss functions, aggregation mechanisms, embedding-representation models, implementation toolkits, datasets, and within-study base-model comparisons. In the reviewed corpus, the graph convolutional network (GCN) is the most common single-model family, which is most often associated with favorable within-study comparisons. In contrast, the model combining GCN with a graph attention network (GAT) is the most frequent multimodel configuration and is most often associated with favorable within-study comparisons. The reviewed corpus yields three broader insights: 1) multimodel architectures have become increasingly prominent after 2023, 2) heterogeneous graphs and a few broad-spectrum backbone families define a wide range of cross-domain applications, and 3) reporting transparency improves unevenly across implementation-oriented studies. This study aims to answer critical research questions regarding the analysis of dataset evolution, feature trends, and learning strategies in GNNs. The findings suggest that future GNN research should report learning settings, implementation toolkits, auxiliary representation models, and layer usage more systematically and evaluate multimodel proposals against strong, clearly identified single-model baselines. These findings provide valuable insight for researchers and practitioners in the field.

Large language models (LLMs) are increasingly integrated into recommender systems to support semantic reasoning, natural language understanding, and user-adaptive personalization. However, their reliance on static parametric knowledge and fixed representations limits robustness in dynamic environments, particularly under long-tail and cold-start conditions. Retrieval-augmented architectures have emerged to address these limitations by grounding LLMs in external, non-parametric knowledge sources. This systematic literature review synthesizes 138 peer-reviewed studies published between 2023 and 2025 in conferences and journals, focusing on retrieval-augmented and LLM-enhanced recommendation. We analyze these works through a three-dimensional framework covering: (i) domain application, (ii) semantic feature and representation design, and (iii) algorithmic strategies for retrieval and personalization. The review shows that current research is concentrated in general recommendation and information retrieval, that similarity/retrieval, user-item interaction, and textual content signals dominate semantic modeling, and that LLM and BERT-style encoders form the primary representation backbones, while graph-based, multimodal, and hybrid approaches remain comparatively underexplored. Algorithmically, most systems adopt generic LLM-centric modeling with limited use of retrieval optimization, reinforcement learning, or structure-aware strategies, and only sporadic attention to explicit cold-start, hallucination, and robustness treatment. By mapping co-occurrence patterns between domains, semantic features, representation choices, and strategy families, this review identifies concrete gaps and transfer opportunities for future work on retrieval-augmented recommendation and provides a structured reference for designing more context-aware, explainable, and data-efficient LLM-based recommender systems.

The escalating complexity and heterogeneity of modern energy systems—particularly in smart grid and distributed energy infrastructures—has intensified the need for intelligent and scalable security vulnerability classification. To address this challenge, we propose Vulnerability2Vec, a graph-embedding-based framework designed to enhance the automated classification of security vulnerabilities that threaten energy system resilience. Vulnerability2Vec converts Common Vulnerabilities and Exposures (CVE) text explanations to semantic graphs, where nodes represent CVE IDs and key terms (nouns, verbs, and adjectives), and edges capture co-occurrence relationships. Then, it embeds the semantic graphs to a low-dimensional vector space with random-walk sampling and skip-gram with negative sampling. It is possible to identify the latent relationships and structural patterns that traditional sparse vector methods fail to capture. Experimental results demonstrate a classification accuracy of up to 80%, significantly outperforming baseline methods. This approach offers a theoretical basis for classifying vulnerability types as structured semantic patterns in complex software systems. The proposed method models the semantic structure of vulnerabilities, providing a theoretical foundation for their classification.