Collaborative Research: CIF: Medium: Robust Learning over Graphs

Connected sensors, autonomous systems and the Internet all produce vast amounts of structured data that must be analyzed. Graphs can model complex networked systems, and as such graph-based machine learning systems aiming at inferences from structured data, have gained significant traction. Data-driven systems, on the other hand, must deal with noisy, uncertain, and outlying measurements, as well as data provided by untrustworthy or even malicious sources. In addition, graph-aware systems must cope with errors in the estimated graph structure. Aspiring to address these challenges, this project introduces a comprehensive framework for robust learning over graphs that can identify and mitigate outliers, adversarial attacks, and errors in the graph structure. The development of scalable, robust, and trustworthy learning can enable active and timely inference and decision-making in domains such as social analytics, crowdsourcing, health informatics, and Internet-of-Things security. The overarching goal of this project is to fortify learning over graphs against noise, anomalies, errors in the data, and adversaries. This project consists of three intertwined thrusts: (T1) identifying anomalies in nodal processes over network graphs; (T2) Graph-aware deep learning under perturbed graph structure, and; (T3) information and decision fusion to quantify the reliability of information sources. T1 leverages random sampling and consensus, as well as graph signal processing tools to pinpoint outlying nodes. Dealing with noisy or perturbed graph topologies, T2 endows graph convolutional networks with dithering-inspired modules that are robust to changes in graph links. Finally, T3 will design and analyze graph-cognizant unsupervised ensemble learning and crowdsourcing algorithms to assess the reliability and usefulness of various information sources. Results from this project will be disseminated to the broader research community through publications, workshops, and code sharing. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.