A Comprehensive Review of Deep Temporal, Graph-Based, and Bayesian Learning Frameworks for Risk Assessment and Community Detection

Abstract

Risk assessment and structural pattern discovery in complex systems—such as environmental monitoring networks, spatiotemporal infrastructures, and large-scale relational data—pose significant challenges due to nonlinear temporal dynamics, latent graph structures, and pervasive uncertainty. In recent years, the rapid development of deep learning has led to a diverse body of methods integrating temporal neural models, graph-based representation learning, and Bayesian inference to address these challenges. However, existing studies are often scattered across different application domains and methodological paradigms, lacking a unified and systematic perspective.

This survey presents a comprehensive review of recent advances in deep temporal modeling, graph neural networks, and uncertainty-aware Bayesian learning frameworks for risk assessment and community detection. We first examine sequence-based models, including Long Short-Term Memory (LSTM) networks and Transformer architectures, highlighting their strengths and limitations in capturing long-range temporal dependencies for hazard source identification and risk prediction. We then review graph neural network–based approaches for community detection, with particular emphasis on hybrid frameworks that combine graph convolution or attention mechanisms with classical clustering and modularity optimization to enhance structural awareness and interpretability. Furthermore, we analyze Bayesian deep learning models and operator-learning frameworks that incorporate probabilistic reasoning, Markov priors, Fourier spectral modeling, and gauge-equivariant constraints to achieve calibrated prediction, robustness, and trustworthy decision support.

To provide a structured understanding of the field, we introduce a methodological taxonomy and comparative analysis across key dimensions, including temporal modeling capability, graph structure integration, uncertainty quantification, and interpretability. Finally, we discuss open challenges and future research directions, such as scalable dynamic graph learning, unified temporal–graph–Bayesian modeling, and explainable uncertainty-aware systems. This survey aims to serve as a reference for researchers and practitioners seeking principled and reliable intelligent modeling approaches for complex, uncertain, and interconnected real-world systems.