ABSTRACT

Recent advances in sensing and computing techniques have led to a need for massive quantities of data to be aggregated from heterogeneous information sources in fields such as science, engineering, and business that are naturally modeled in the form of big attributed networks. A big attributed network (BAN) is characterized by a combination of (a) high-dimensional and heterogeneous network topologies and (b) high-dimensional and heterogeneous attribute data. Effective analysis of BAN data relies on simultaneous subgraph mining and feature selection for discovering complex patterns that are interesting or significant. However, as yet little has been done to bridge these two important research areas. The focus of this project is therefore to unify a wide range of complex pattern discovery tasks including, for example, the detection and forecasting of societal events (disasters, civil unrest), anomalous patterns (disease outbreaks, cyberattacks), discriminative subnetworks (cancer diagnosis), knowledge patterns (new knowledge building) and storylines (intelligence analysis), and to resolve the fundamental modeling, algorithmic, and interactive challenges associated with ubiquitous BAN data in today's big data era. This project incorporates the resulting research outcomes into the curricula of interdisciplinary courses on topics such as complex pattern detection in BAN data presented at seminars, tutorials, and workshops, and will include outreach activities such as a big data analytics summer camp for local K-12 education in New York's Capital Region.

The research objectives of this project are: (1) the development of a unified, theoretical framework for discovering complex patterns in BAN data in various kinds of tasks; (2) making the inference computationally tractable in extremely large combined spaces composed of vertices, edges, and attributes; and (3) rendering the detected heterogeneous patterns transparent and interpretable and incorporating heterogeneous user feedback into the detection process. The research approach includes the development of: (1) novel principled methods capable of learning complex patterns of interest directly from BAN data; (2) near-linear-time common inference algorithms capable of optimizing a variety of BAN-specific model objectives that are subject to different constraints on subgraph topologies and structured sparsity models; and (3) a computer-interpretable language-based system capable of modeling and interpreting rich user feedback schemes in BAN data. More details can be found at: http://www.cs.albany.edu/~fchen/projects/BAN/.

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.

Please report errors in award information by writing to: awardsearch@nsf.gov.