Advancing Model Explainability in Pervasive Computing

Pervasive computing technologies are increasingly integral to numerous domains, necessitating enhanced explainability due to their human-centered design. However, task difficulty and data complexity pose substantial challenges for Explainable Artificial Intelligence (XAI) methodologies. Particularly, the predictive performance of interpretable models is often limited within these domains. Furthermore, the inherent complexities of pervasive computing, characterized by noisy, high-volume, and temporally dependent data, exacerbate the difficulty in developing effective explanatory methods. Prevalent XAI approaches, especially those dependent on saliency maps, typically fail to adequately elucidate the underlying decision-making processes of complex models.
This dissertation aims to advance the field of XAI within pervasive computing by addressing these critical challenges. It is structured around three principal objectives: First, it seeks to improve the predictive performance of interpretable models through the introduction of an innovative automatic feature engineering framework, coupled with an optimization algorithm specifically designed for pervasive environments. ... mehrSecond, the dissertation conducts a comprehensive literature review and model architecture analysis to systematically identify key data elements leveraged by high-quality models in the domain. Third, utilizing these key data elements, it proposes novel post hoc explanation methods, including a tree-search approach for exploring model decisions and a Markov chain-based technique to elucidate temporal dynamics. These contributions collectively represent a significant advancement in the field of XAI for pervasive computing, facilitating the development of more effective and interpretable artificial intelligence systems within this complex domain.