Interpretations and Actions towards Trustworthy Graph Neural Networks

点击翻译按钮获取中文摘要

Graph neural networks (GNN) have been widely adopted to achieve state-of-the-art performance in many real-world applications. Big US companies such as Google, Amazon, Facebook, and Microsoft have been developing large business applications based on GNNs. Most of these models are built and used as black boxes that are not easily interpreted to human. The lack of interpretability makes it difficult to verify and control the behavior of GNNs. This lays severe security problems, reduces the trust from users, and prohibits massive deployments of GNNs. In response to these issues, governments are drafting laws to entitle citizens the right to understand the inner workings of automated models. As a result, trustworthy GNNs have been a critical demand of governments, industry companies, and billions of end users involved in businesses built on GNNs. Developing trustworthy GNNs provide us unprecedented power to leverage big data in a reliable way, and, at the same time, post grand technical challenges. This proposed research program is to embrace the opportunities and address the technical challenges in a timely and practical manner. The long-term objective of this research program is to develop trustworthy business intelligence based on reliable and consistent interpretations on GNNs. In general, the decisions of trustworthy business intelligence should be interpretable to human, fair to users, and robust to adversarial attacks. We divide our long-term objective into four short-term objectives. First, we will develop principled approaches to produce provably reliable and consistent interpretations on GNNs that are originally built and used as black boxes. Second, we will develop transparent GNNs that are natively designed to be interpretable while achieving comparable performance as conventional GNNs. Third, we will take actions based on interpretations to enhance fairness and robustness of GNNs. Last, we will conduct case studies in real applications to verify and evaluate our research outcome and produce practical impact. Our research will have great impacts in both academia and industry. In academia, the proposed research will gain principled understanding on black-box GNNs, establish effective ways to build trustworthy GNNs, and explore a large family of principled approaches to use interpretations for better model performance. The research will advance the frontier in the fast-growing area of trustworthy big data analytics and train competitive HQP to produce substantial impact in academia. In industry, some of our strategic industry partners have already adopted our previous techniques on trustworthy machine learning; and we expect more of our research outcome to be used to boost trustworthy business intelligence. Moreover, this research will systematically train a diverse group of HQP to equip them with highly competitive skill sets. Exporting advanced techniques and talented HQP to industry will certainly produce significant impacts.