Dynamic characterization of explanations during human–machine collaboration in cyber-physical systems

Abstract Cyber-Physical Systems often rely on human–machine collaboration to achieve superior outcomes and mutual learning. For effective collaboration, systems must be understandable and trustworthy. Explanations are instrumental in promoting system understandability and building human trust. However, crafting such explanations poses key challenges: How many explanations do users need? How much information can they process? What level of detail is appropriate? This work introduces a conceptual model to characterize explanations and a predictive model that determines when and what type of explanation to provide. The predictive model considers system context, user profile, and interaction behavior, using machine learning techniques to infer explanatory needs and tailor responses accordingly. Although we apply our approach to a smart home environment, the methodology is generalizable to other CPS domains. At design time, developers define possible explanation points and relevant parameters. At runtime, the system autonomously adapts explanations to user-specific conditions. To validate our proposal, we conducted an experimental study evaluating the resulting explanations in terms of understandability, obtrusiveness, and adaptability. Results showed modest but meaningful improvements in all three dimensions, suggesting that our framework facilitates explanation designs more aligned with user preferences and situational demands.