A Framework for Spatial-Temporal based Socio-Cyber-Physical System

The increasing complexity and interconnected nature of Smart Environments (SE) necessitates sophisticated frameworks for understanding, representing, and analyzing their dynamics. This thesis introduces a novel, end-to-end framework that integrates SE, Complex Network Analysis (NA), and Factual Explanation (FE) to enhance SE profiling. Unlike existing approaches that often treat these components in isolation, our framework leverages their synergistic potential to provide deeper insights into environmental dynamics. At its core, the framework comprises three innovative components: (1) a semi-autonomous symbolic rule extraction pipeline that transforms raw data into semantic representations, incorporating human expertise for validation; (2) a dynamic network analysis mechanism that captures temporal evolution through Graph Neural Network (GNN) and Community Detection (CD); and (3) a FE system that employs homophily and heterophily analysis to evaluate information flow and validate community structures. To validate our findings, we conducted extensive experiments using three diverse datasets: a smart workplace dataset, a social interactions dataset, and a movie recommendation dataset. Details of each dataset and their features are provided in the respective chapters. Our results indicate significant improvements in understanding community dynamics and relationship patterns, achieving notably accurate factual explanations through network analysis metrics. Specifically, when applied to dynamic Knowledge Graph (KG), our approach achieved an accuracy of 58.23% and an F1-score of 53% in predicting dynamic interactions, outperforming several baseline methods by 10.28%. These outcomes demonstrate the framework’s robustness and adaptability across various contexts. Beyond its immediate applications, this thesis paves the way for future research, particularly through two experimental studies exploring directional entropy and Large Language Models for graph query generation. This work not only bridges the critical gap between knowledge representation, NA and Explainable AI (XAI) but also provides a comprehensive solution for profiling SE in an interpretable and scalable manner. Collectively, these contributions pave the way for more sophisticated approaches to understand and manage SE.