In today's hyperconnected world, demonstrated clearly by the rapid global spread of Covid-19, understanding complex networks has never been more crucial. Network science—the study of interaction patterns across physical, biological, social, and information systems—provides powerful frameworks for solving our most challenging problems. By analyzing these intricate connections, researchers can uncover hidden patterns and predict behaviors across diverse domains.
The 2021 Graph Exploitation Symposium (GraphEx), virtually hosted by MIT Lincoln Laboratory, gathered pioneering network science experts to showcase breakthrough innovations and real-world applications in this rapidly evolving field.
"Our mission is to discover how leveraging graph data can provide essential technological solutions to address the most critical challenges facing our nation today," explains Edward Kao, symposium organizer and technical staff in Lincoln Laboratory's AI Software Architectures and Algorithms Group.
The virtual conference centered on pressing contemporary issues, including analyzing social media disinformation campaigns, developing sophisticated pandemic propagation models, and utilizing graph-based machine learning models to accelerate pharmaceutical development.
"The focused sessions on influence operations and Covid-19 at GraphEx highlight how network and graph-based analysis helps us understand these complex, impactful aspects of modern society, while also suggesting potential pathways forward as we advance our knowledge of graph manipulation techniques," notes William Streilein, who co-chaired the event with Rajmonda Caceres, both from Lincoln Laboratory.
Social network analysis
Multiple symposium presentations examined how network science enables the detection and analysis of influence operations (IO)—coordinated efforts by state or non-state actors to disseminate deceptive narratives across digital platforms.
Lincoln Laboratory researchers have developed innovative tools to identify and measure the impact of social media accounts likely involved in IO, such as those deliberately promoting false Covid-19 treatments to susceptible populations.
"IO accounts function as coordinated echo chambers that amplify deceptive narratives. Vulnerable populations then engage with and spread this misinformation," states Erika Mackin, a researcher behind RIO (Reconnaissance of Influence Operations), a sophisticated detection framework.
To identify IO accounts, Mackin's team trained an algorithm to recognize probable IO accounts within Twitter networks based on specific hashtags or narratives. One case study involved #MacronLeaks, a disinformation campaign targeting Emmanuel Macron during the 2017 French presidential election. The algorithm evaluates accounts based on multiple indicators, including interactions with foreign news sources, frequency of shared links, and language diversity. Their model then applies statistical methods to calculate each account's influence in propagating narratives within the network.
The research team discovered that their classifier surpasses existing IO detection tools by identifying both automated bot accounts and human-operated ones. They also found significant overlap between accounts pushing the 2017 French election disinformation and those currently spreading Covid-19 misinformation. "This indicates these malicious actors adapt their strategies to promote whatever disinformation narrative serves their current objectives," Mackin observes.
AI-powered pandemic modeling techniques
Throughout the Covid-19 crisis, policymakers have relied on epidemiological models to forecast disease transmission and guide public health decisions. Alessandro Vespignani, director of the Network Science Institute at Northeastern University, has been at the forefront of Covid-19 modeling efforts in the United States and delivered a keynote presentation on this critical work.
Beyond incorporating biological disease parameters like incubation periods, Vespignani's model excels by integrating community behavior patterns. To create realistic disease transmission simulations, his team develops "synthetic populations" using comprehensive, publicly available datasets about U.S. households. "We construct populations that, while not real individuals, are statistically representative of actual communities, and then map the interaction patterns between these synthetic individuals," he explains. This information feeds back into the model to generate accurate predictions about disease spread.
Currently, Vespignani is exploring ways to incorporate viral genomic analysis into population modeling to better understand variant transmission. "This is an extremely promising work in progress," he notes, adding that this approach has proven valuable in modeling the dispersal of the SARS-CoV-2 Delta variant.
As researchers continue modeling virus transmission, Lucas Laird at Lincoln Laboratory is investigating how network science can inform effective control strategies. His team is developing a model for customizing intervention approaches for different geographic regions. This work was motivated by observing significant variations in Covid-19 spread patterns across U.S. communities and recognizing the need for more nuanced intervention modeling.
To demonstrate their approach, they applied their planning algorithm to three counties in Florida, Massachusetts, and California. By accounting for region-specific characteristics such as susceptible population size and infection rates, their system implements tailored strategies throughout the outbreak duration.
"Our approach eliminates disease transmission within 100 days while using significantly more targeted interventions than broad, one-size-fits-all approaches. Essentially, you don't need to implement nationwide shutdowns," Laird explains. He adds that their planning tool provides a "sandbox environment" for testing and refining intervention strategies for future outbreaks.
Graph neural networks for drug discovery
Graph-based machine learning is gaining significant attention for its ability to "learn" complex relationships within graphical data, enabling new insights and predictions about these connections. This interest has spurred the development of a new algorithm class called graph neural networks. Today, these networks are being successfully applied in domains including drug discovery and materials design, with remarkable results.
"We can now apply deep learning techniques far beyond traditional applications like medical imaging and biological sequence analysis. This opens exciting new frontiers in data-rich biology and medicine," says Marinka Zitnik, an assistant professor at Harvard University who presented her research at GraphEx.
Zitnik's research focuses on the complex interaction networks between proteins, drugs, diseases, and patients, encompassing billions of relationships. One application involves identifying treatments for diseases with limited or no approved therapies, such as Covid-19. In April, Zitnik's team published research employing graph neural networks to evaluate 6,340 drugs for their potential effectiveness against SARS-CoV-2, identifying four promising candidates that could be repurposed to treat Covid-19.
Similarly, Lincoln Laboratory researchers are applying graph neural networks to the challenge of designing advanced materials, such as those capable of withstanding extreme radiation or capturing carbon dioxide. Like drug development, traditional materials design relies on time-consuming and costly trial-and-error approaches. The Laboratory's team is developing graph neural networks that can learn the relationships between a material's crystalline structure and its properties. This network can then predict various properties from any new crystal structure, dramatically accelerating the screening process for materials with desired characteristics for specific applications.
"Graph representation learning has emerged as a dynamic and flourishing research area for incorporating inductive bias and structured priors into machine learning processes, with wide-ranging applications including drug design, accelerated scientific discovery, and personalized recommendation systems," Caceres notes.
A thriving research ecosystem
Lincoln Laboratory has hosted the GraphEx Symposium annually since 2010, with only last year's cancellation due to Covid-19 interrupting this tradition. "A key insight from this year's event is that despite last year's postponement and the virtual format, the GraphEx community remains as vibrant and engaged as ever," Streilein reflects. "Network-based analysis continues to expand its reach and is being applied to increasingly important areas of science, society, and defense with growing impact."
Beyond Lincoln Laboratory participants, technical committee members and co-chairs of the GraphEx Symposium included researchers from Harvard University, Arizona State University, Stanford University, Smith College, Duke University, the U.S. Department of Defense, and Sandia National Laboratories.
