Building Trustworthy GNNs: Enhancing Explainability and Robustness on Heterogeneous Urban Graphs
In the current era of comprehensive digitization, the rapid expansion of intelligent city infrastructure and advancements in positioning technologies have resulted in the generation of vast amounts of urban data that depict human activities. Examples of such data include traffic flow data, energy data, air quality and noise level data, among others. The extraction of valuable insights from these rich sources has provided solutions for addressing real-world urban issues, including human mobility understanding, smart transportation, urban planning, public safety, etc. Due to the inherent network structure of urban streets and city facilities, these data naturally possess a graph structure or can be modeled using graphs, such as public transportation networks and urban social graphs. The escalating number, volume, and heterogeneity of urban graph data have rendered traditional data mining methods, particularly statistics-based methods, inadequate in handling the vast amount of available data.In response to the daunting task of analyzing vast and diverse urban graph data, researchers in recent years have increasingly turned to the utilization of Graph Neural Networks (GNNs) and their augmented variants for the purpose of predicting and analyzing urban issues. Through the application of GNNs, remarkable advancements have been made, attaining state-of-the-art performance in various domains, including urban traffic flow prediction, urban anomaly detection, and urban event detection, among others.GNNs have obtained great performance in capturing the complex spatial correlations of urban data in forecasting tasks. Enhancing the explainability of GNN models in the context of predicting urban problems furnishes decision-makers with compelling evidence to effectively address specific urban challenges. However, deep neural networks have always suffered from the low interpretability of their black-box representations. While recent progress in the field of interpretable machine learning has given rise to the emergence of a multitude of GNN models that possess either intrinsic interpretability or can be enhanced with post-hoc explanation techniques, the explainability of GNNs in the field of urban study is still insufficiently studied due to unique urban properties, such as high spatial heterogeneity.Moreover, the heterogenous nature of urban data sources introduces considerable challenges when performing cross-domain analysis of distinct urban graphs. Effectively integrating and modeling diverse types of nodes and edges within these graphs becomes a significant undertaking. Additionally, constructing a robust cross-domain association model necessitates capturing and comprehending the semantic associations inherent in the data, which, in turn, involves intricate spatiotemporal data alignment procedures. While Extensive Heterogeneous GNNs (HGNNs) have been explored in the context of heterogeneous graphs, the development of flexible HGNNs that can mitigate the impact of adversarial attacks on prediction outcomes remains an unresolved challenge.Therefore, this research proposal aims to build trustworthy GNNs that enhance the explainability of GNNs in the context of urban studies and enable effective and robust cross-domain models on heterogenous urban graphs. By developing novel techniques tailored to the specific challenges posed by urban data, my research seeks to bridge the gap between GNN-based predictions and decision-making processes in urban environments.