Graph Neural Networks GNNs) have emerged as powerful models for learning over complex graph-structured data, with applications spanning social networks, biology, and recommendation systems. Despite their success, the black-box nature of GNNs poses significant challenges for interpretability and trustworthiness—particularly in high-stakes domains such as healthcare and finance. This talk presents a data-driven perspective on explainable and reliable GNNs, introducing four novel paradigms. First, view-based explanations capture both fine-grained subgraphs and higher-level patterns, enabling queryable and configurable interpretability. Second, we define robust counterfactual witnesses, explanation structures that remain stable under adversarial perturbations, and design efficient verification and generation algorithms, including parallel strategies. Third, we propose skyline explanatory queries, which provide multi-objective, diversified explanations with provable approximation guarantees. Finally, we introduce counterfactual evidence, which is a new counterfactual paradigm that grounds explanations in similar but differently labeled real examples, avoiding infeasible perturbations. We demonstrate the effectiveness of these approaches in real-world applications such as revealing biases in GNN models and ensuring robustness in security- critical settings. Together, these contributions advance the goal of building transparent, scalable, and trustworthy graph learning systems.
Bio Dazhuo Qiu is a Ph.D. candidate in Computer Science at Aalborg University, Denmark, supervised by Prof. Arijit Khan and co-supervised by Prof. Yan Zhao. His research focuses on graph data management, graph neural networks GNNs), and explainable AI, with an emphasis on developing trustworthy and user-centric explanation frameworks. He has published at top database and AI venues, including SIGMOD, KDD, and ICDE. His broader research vision lies at the intersection of data management and large language models LLMs), focusing on retrieval-augmented generation DB4LLM and LLMs as intelligent interfaces for databases LLM4DB. He has received multiple recognitions, including the KDD 2025 Student Travel Award, the SIGMOD/PODS 2024 Student Scholarship, and the Best Industry, Systems, and Apps Award at IEEE MDM 2022. Beyond research, he has served as a PC member for AAAI 2024 and external reviewer for top venues such as SIGMOD, PVLDB, ICDE, and NeurIPS. He has given invited talks on explainability and fairness in GNNs CIKM 2024 Workshop on Responsible AI and served as a panelist at KDD 2025 Undergraduate and Masterʼs Consortium.
