Filters

Guidelines

Provide Contextual Information

Placing explanations together with additional contextual information enhances their relevance and interpretability to domain-expert users, e.g., physicians [10.1109/TVCG.2021.3114836,10.2196/28659], AI engineers [10.1145/3613904.3642106,10.1109/VAST50239.2020.00010], etc. Moreover, the depth of information in explanations should be adaptable to the user’s expertise, ensuring that both detailed and summary information are available without overwhelming the user [10.1007/978-3-031-60606-9_13, 10.1016/j.ijhcs.2022.102941, 10.1145/3290605.3300831]

Interactivity and Control

Effective XUIs empower users by supporting active engagement with the system. Supporting interactive exploration was found useful for users, for example allowing selection, filtering, justapoxition, and smooth transitions between views [ 10.1109/TVCG.2023.3327389, 10.1109/TVCG.2020.3030418]. Moreover, features such as what—if exploration, guided navigation, and real-time updates were found to enhance user engagement and comprehension [10.1109/PacificVis48177.2020.7090, 10.1145/3510003.3510129]. Interaction is also a core requirement for white-box models to iteratively refine the understanding of explanations [10.1145/3490099.3511111 , 10.1145/3514258]. Regarding reinforcement learning systems, interaction modalities such as questioning the model and navigating through explanation spaces help users clarify why actions were taken, increasing trust and understanding of the system [10.1109/PacificVis53943.2022.00020].

Navigating explanations

Users should have freedom and control in accessing and navigating the explanations by including UI controls such as “next”, “previous”, or “exit”.