Network failures, whether due to random disruptions or malicious attacks, pose significant challenges for uncrewed aerial vehicle (UAV) swarm networks. One critical concern is determining which failed UAVs to recover or replace under limited resource conditions to enhance the robustness of their communication networks. Current research primarily considers static structural characteristics of the network and struggles to uncover deep features that influence network robustness, and the efficiency cannot meet the real-time needs in UAV swarm scenarios. To address these issues, we introduce a Prioritized Recovery strategy for failed nodes based on graph reinforcement learning (PRGRL). This approach integrates a random SAmpling neighbor method with a multihead attention mechanism to create a novel graph convolutional kernel (SAGCK). This kernel is designed to extract global structural information and relative positional information of nodes within the graph. Additionally, we develop a deep policy network (DPN) that explores the intricate relationships between graph-level and node embedding features, enabling the assessment of nodes’ impact on overall robustness. PRGRL’s network parameters are automatically updated and optimized using scalable deep reinforcement learning. Importantly, PRGRL prioritizes the recovery of boundary nodes within connected components to enhance network robustness further. Our experiments, conducted on both simulated and real-world networks, demonstrate that PRGRL outperforms existing methods of robustness enhancement across various recovery ratios, attack strategies, and network sizes while delivering superior real-time performance.

The robustness of Internet of Things (IoT) communication topologies against internal failures and external perturbations is a fundamental prerequisite for maintaining system stability. This paper studies IoT topology robustness from two perspectives: robustness metric and optimization. Existing robustness metrics, primarily based on the maximum connected subgraph, neglect contributions from other connected subgraphs, thereby inadequately capturing dynamic topological changes. To address this issue, we propose a robustness metric based on topology data reachability, which sensitively reflects the data transmission capability of an IoT topology under arbitrary perturbations. Regarding robustness optimization, most existing methods adopt centralized strategies that rely on global information, resulting in inefficiencies and limited adaptability in decentralized IoT environments. We propose DecTRO, a Decentralized Topology Robustness Optimization method for IoT via multi-agent graph reinforcement learning. To mitigate partial observability, DecTRO employs a scalable graph attention network enhanced with multi-modal sampling, which aggregates cross-agent information and captures spatiotemporal correlations. Furthermore, a topology robustness-oriented node sampling method is introduced to reduce action-space complexity and accelerate convergence, while a decentralized heuristic reward function enables efficient online decentralized learning. Experimental results show that DecTRO achieves up to two orders of magnitude (5–125×) greater improvement in robustness per unit time compared with state-of-the-art baselines, striking a favorable balance between robustness enhancement and computational efficiency.

Community detection in multiplex networks seeks groups of nodes from several types of relations. Attention mechanisms can integrate these relations, but weights based on embedding compatibility do not explicitly measure the community support provided by a node’s neighborhood. We propose Community-Guided Attentive Fusion (CGAF), an end-to-end graph neural network framework that uses neighborhood community evidence to guide relation fusion. Graph contrastive learning encodes refined relation graphs and a learned fused graph to capture relation-specific and shared information. Shared prototypes produce provisional soft memberships, which are used to calculate degree-corrected neighborhood evidence. This evidence adjusts node-specific attention, and a learnable residual coefficient combines the relation embeddings with the fused-graph representation. The prototypes then assign communities to the resulting representations. Joint optimization with contrastive, self-training, and modularity objectives connects representation learning, relation weighting, and community assignment. On ACM, DBLP, and Yelp, CGAF achieves the highest observed means across four community detection metrics among eight comparison methods under the stated protocols. Ablation and parameter analyses show how community evidence, structural supervision, and residual fusion contribute to the learned partitions.

Textual-edge graphs represent interactions through both connectivity and relation-specific text. Learning community structure from these graphs requires an encoder that preserves the relevance of individual relations when combining semantic and structural information. We propose Dual-Stage Attention Embedding for Community Detection on Textual-Edge Graphs (DAETG). The first attention stage summarizes incident-edge text, and the second transfers its coefficients into neighboring-node aggregation as a learnable prior. Following topology-reconstruction pretraining, Kullback–Leibler self-training and differentiable modularity jointly refine the encoder and a centroid-based community assignment. Training uses fixed node and edge text embeddings and a supplied community count, without category-membership supervision. Experiments on four benchmark graphs with base and large Bidirectional Encoder Representations from Transformers (BERT) embeddings show competitive community-detection performance, with modest mean improvements that vary across datasets and evaluation metrics. Ablation studies on two graphs support the contributions of correctly associated edge text, the dual-stage encoder, and joint community refinement. Together, these results support using relation-specific text alongside topology to learn community representations.

Dynamic community detection must preserve community continuity while responding to abrupt structural changes. Temporal smoothing can improve stability but can also suppress the signals of community formation, splitting, or dissolution. To address this stability–plasticity tension, we propose Role-Decomposed Memory for Dynamic Graphs (RDMG), a self-supervised framework that separates persistent, appearing, and disappearing edges before temporal modelling. Role-specific graph convolutional branches supply an identity representation from persistent edges and separate recurrent memories for edge births and deaths. An event-aware assignment head adds a gated birth-minus-death residual to the identity representation, with its contribution controlled by local edge turnover. An optional community-state gated recurrent unit updates soft community representations across snapshots and combines their temporal-state affinities with node-level assignment logits. Training combines differentiable modularity, role-specific reconstruction, and gate-entropy regularisation without community-label supervision. Experiments on four real-world dynamic networks and three controlled synthetic benchmarks show competitive clustering and event responsiveness. RDMG obtains the best or second-best result in eleven of twelve real-world metric–dataset combinations and the highest event-restricted adjusted Rand index on all three synthetic variants. These findings support separating persistent structure from event-related edge changes when learning dynamic community assignments.