Fengxiao Tang

Large language models (LLMs) have recently been adapted to multivariate time series forecasting (MTSF), but their direct use for temporal data remains limited by frequency entanglement: raw or lightly transformed time series often mix stable spectral patterns with fluctuation-sensitive local variations, making it difficult for LLMs to form reliable macro-context representations. To address this problem, we propose Dominant Spectral Retrieval (DSR), a forecasting framework that combines adaptive frequency decoupling with macro-context retrieval.DSR uses a dual-branch encoder to construct dominant spectral summaries through magnitude-based spectral selection while preserving full-signal representations from the original series. The spectral summaries provide compact macro-context cues to the LLM without imposing a fixed frequency-band prior. The full-signal representation then queries the LLM-generated context through Macro-Dynamic Prototype-Querying (MDP-Q), allowing local dynamics to be refined by summary-conditioned macro guidance.Experiments on eight real-world datasets show that DSR achieves competitive performance against representative LLM-based, frequency-domain, and Transformer-based forecasting methods. Additional ablation and perturbation analyses further support the roles of adaptive spectral selection and macro-context retrieval in improving forecasting robustness.

Multi-agent trajectory prediction relies heavily on information exchange, where incomplete state sharing, limited interaction frequency, and communication uncertainty readily induce collaborative instability. Current approaches primarily rely on heuristic interaction rules, failing to explicitly quantify the trade-off between communication overhead and prediction gains. This implicit assumption often leads to suboptimal information exchange and lacks evolutionary stability for communication strategies in dynamic traffic environments. Specifically, we propose the Stability-aware Communication via Weighted Potential Games (SC-WPG) framework. First, we leverage the continuous-time dynamics of Liquid Neural Networks (LNNs) to design a dynamic phase-coupling mechanism that aligns the temporal evolution of agent states, alleviating strategy asynchrony. Second, we establish a liquid state-sharing pool to exchange compact latent representations encoded from local observations, while promoting efficient implicit state fusion through spatio-temporal attention mechanisms. Finally, communication scheduling is formulated as a weighted potential game to explicitly quantify the trade-off between gains and costs. A stability-aware weight adaptation mechanism based on latent-state variation is introduced to regularize communication behavior under dynamic constraints. Experimental results on the NGSIM and HighD datasets show that SC-WPG achieves the lowest average root mean square error (RMSE) of 0.89 and 0.55, respectively. These results correspond to improvements of 10.1% and 6.8% over the strongest competing baselines under the same communication-constrained setting.

Accurate performance prediction is essential for wireless network optimization and management. Graph neural networks (GNNs) as the state-of-the-art method for network modeling, leverage the pairwise relationships in graph structure of wireless networks to propagate information. However, network performance variations are influenced by diverse environmental factors, such as the operational states of nodes within the same access point (AP) network and the conditions for nodes and links traversed by flows. The lack of many-to-many high-order information will lead to the decline of prediction accuracy. Therefore, we propose HGNet, a wireless network performance prediction model based on hypergraph learning, which constructs multi-type hyperedge groups with hyperedge characteristics and typed-weighted hypergraph message passing mechanism with temporal dynamics. Furthermore, to support privacy-preserving multi-domain deployment, we propose FedHGT, a personalized federated hypergraph structure-aware optimization strategy that addresses higher-order heterogeneity in federated hypergraph training, including the multi-category of nodes and multi-connectivity of hyperedges. The server performs adaptive weighted aggregation based on the contribution and connectivity of hyperedges. We evaluate the performance of HGNet and FedHGT in ns-3 and the results demonstrate that they outperform state-of-the-art baselines by average margins of 15.9% and 5.6%, respectively.

As global demand for efficient network services increases, the limitations of traditional terrestrial wireless networks are becoming more apparent. Space-air-ground integrated networks (SAGIN) have emerged as a promising solution to advance next-generation network infrastructure. However, SAGIN faces significant security challenges—including the lack of a robust security architecture, trust and data reliability issues in multi-hop transmissions, and the need to enhance network performance without compromising security—rendering traditional boundary-based defenses inadequate. Therefore, we propose SECURELINK, a decentralized zero-trust architecture tailored for SAGIN, which replaces traditional perimeter defenses with a ”never trust, always verify” approach. This approach strengthens network security and flexibility through continuous verification and adherence to the principle of least privilege. SECURELINK integrates blockchain technology to establish a multi-layered security verification and data processing scheme, addressing the dynamic and decentralized features of SAGIN. Additionally, we introduce DFRIO, a zero-trust traffic offloading method based on decentralized federated reinforcement learning and blockchain, designed to enhance network performance within maintaining security. Simulation results demonstrate that our solution significantly enhances SAGIN’s defense capability without compromising network stability, outperforming the two baseline schemes by 18.3% and 42.6%, respectively.

Fault diagnosis in microservice systems requires high availability, driving research towards multimodal learning that leverages heterogeneous monitoring data, including logs, metrics, and traces. These data inherently contain time series and data streams across various modalities. However, existing frameworks fail to exploit temporal dependencies in these dynamic streams effectively. First, existing methods rely too much on original time-domain metrics, which makes it difficult for them to capture periodic patterns and sudden events. Second, most language-bound integrations treat timestamps only as sequential labels or numerical inputs, which ignores the rich contextual information within timestamps. As a result, existing models struggle to comprehend the temporal sequences and data characteristics associated with faults in multimodal data. In this work, we introduce FD-TE Diagnosis, a framework for microservice fault diagnosis that addresses these challenges by applying frequency domain feature analysis and time encoding. To enhance the detection of abnormal events, we integrate a frequency-domain approach that utilizes the Fast Fourier Transform (FFT) to extract robust features from metric data. Next, we encode the timestamps of all events using a dedicated time encoding layer. These temporal representations are then incorporated to strengthen event embedding for fault inference. Experimental results demonstrate that our method enhances the sensitivity of multimodal models in interpreting temporal features, resulting in improved diagnostic outcomes.