Enhancing trust through a human-center evaluation framework from an accessibility perspective: The case of graph anomaly detection

Artificial intelligence (AI) significantly enhances operational efficiency and decision-making across sectors like transportation, finance, and healthcare. However, the increasing complexity of AI systems challenges their comprehensibility and reliability for users (especially non-experts). While Explainable AI (XAI) aims to make models more transparent, it tends to cater more to experts rather than laypeople. This study emphasizes enhancing user trust in AI results by improving accessibility, defined through three core dimensions: comprehensibility, verifiability, and timely feedback. Comprehensibility reduces knowledge barriers for non-experts; verifiability allows reasonable validation of AI conclusions with interactive tools; and timely feedback supports swift responses, fostering a positive interaction cycle. To illustrate these principles, we focus on graph anomaly detection (GAD) as the case. We introduce GADTrust, a user-centered visual evaluation framework designed to enhance the accessibility of GAD results. GADTrust features perceptible trust-oriented visual evaluations and an interactive trust-analysis system to provide clear explanations, sufficient verification means, and prompt feedback. Through extensive experiments on real-world datasets, we demonstrate that GADTrust effectively aids users in building trust in GAD results, thus bridging the gap between AI technologies and users.