ATD: Sparse and Localized Graph Convolutional Networks for Anomaly Detection and Active Learning

University Of Texas At Dallas, Richardson TX

Investigators

Abstract

Many applications produce massive quantities of data in which complex relationships and interdependency are naturally modeled as graphs, such as road networks, mobile networks, and public health surveillance networks. Despite considerable attention on graph convolution networks (GCNs), the vast majority of the existing works are limited to balanced-node classification, thus ineffective at detecting anomalous nodes accurately. This project aims at a sparse and localized GCN for detecting anomalous patterns. In addition, a novel active learning framework will be developed that learns the optimal query strategy to reduce the number of training labels. Theoretical investigations will be performed to interpret the anomalous patterns in an attempt to align with intuitions and clarifications from domain experts. The computational tools developed will be applicable, specifically in the fields of remote sensing and geospatial information. Furthermore, the investigators will incorporate the research results into developing new interdisciplinary courses with a focus on both theory and application. These courses will serve as a springboard for student recruitment, such as graduate students interested in using this research as the subject of a Ph.D. thesis and undergraduates interested in a summer research project. Preliminary studies on real datasets have strongly demonstrated a preference for sparse convolution filters on the graph. It is also desirable to have joint localization in both the vertex and graph frequency domains for computational efficiency and structural feature extraction. By combining sparsity and localization, the project aims at a sparse and localized GCN paradigm that efficiently detects anomalous nodes. In particular, this project will address the following questions: (1) how to design graph convolutions to effectively enforce sparse and localized patterns for anomaly detection; (2) how to optimize the network architecture and algorithm design; and (3) how to reduce the user’s burden in the collection of labeled data by active learning, which involves a sequence of dynamical query for labeling a small number of nodes that are most effective in anomaly detection. Overall, this project will advance the algorithmic and theoretical foundations of GNNs and anomaly detection. The research will contribute to the areas of time-frequency analysis, harmonic analysis, uncertainty quantification, and nonconvex optimization. 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.