NSF Award
2323533
Graphs (networks) are a versatile scientific framework to represent and analyze biological, social, and human-made complex systems. Such complex systems are inherently dynamic—for example, social interactions and human activities are intermittent; links appear and disappear in functional brain networks. Despite “time” playing a central role in those systems, most of the classic studies on graphs are based on the topological properties of static graphs (graphs that do not change over time). The existing works on dynamic graphs show only limited scalability for large-scale practical datasets. This proposed research aims at designing fast, scalable methods for revealing dynamic behaviors of a socio-technical system by developing innovative algorithmic and computing techniques. The host site, Berkeley Lab, will provide unique expertise and mentoring and facilitate access to leading supercomputer facilities to achieve the proposed research goals. The project will generate new algorithmic techniques and scalable software tools to advance graph-based data science and high-performance scientific computing. The PI includes an underrepresented graduate student in this research. Educational and training modules will also be developed for PI’s institution from the techniques and results emerging from this project. Thus, the project will enhance the scientific research, training, and education capacity of the PI’s jurisdiction.<br/><br/>The goal of this EPSCoR proposal is to develop fast and scalable methods for mining and analyzing large dynamic graphs. Examples of such graphs include social networks, human contact networks, web graphs, and functional brain networks. The proposal addresses substructure-based problems such as finding evolving communities and enumerating interesting temporal subgraphs or motifs with applications in neuroscience, bioinformatics, infrastructure, and social domains. Even though there exists a rich literature for static graphs, the literature for dynamic graphs is very nascent. Existing parallel algorithms for dynamic graphs demonstrate limited scalability due to their low ratio of compute to memory operations and the irregular memory access patterns. Consequently, such algorithms show weak spatial and temporal locality, leading to poor cache utilization and high communication volume. The proposed research will utilize a unique collaboration with the Performance and Algorithms Group of Berkeley Lab to avail the most advanced user facilities and leading expertise to tackle the above technical challenges. The proposal aims at developing scalable parallel methods with efficient load-balancing and communication-avoidance techniques, data reduction approaches with sampling and sparsification, and efficient formalization of temporal metrics. Algorithmic methods generated from this proposal will be applicable in understanding dynamic properties of various real-world systems—for instance, locating key neurons in cortical (brain) networks, route-planning for time-varying traffic in infrastructure networks, modeling disease/virus or information propagation in social/contact networks. Therefore, the project will expand the PI’s research capacity to build impactful software/technology tools and also enhance his ability to serve a diverse student population at his host institution as both a research mentor and an educator.<br/><br/>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.
