Open Collections

Behrouz, Ali

University of British Columbia

The problem of identifying anomalies in dynamic networks is a fundamental task with a wide range of applications from understanding brain diseases/disorders to fraud detection in financial networks. However, it raises critical challenges due to the complex nature of anomalies, lack of ground truth knowledge, and complex and dynamic interactions in the network. Most existing approaches usually study simple networks with a single type of connection. However many complex systems exhibit natural relationships with different types of connections, yielding multiplex networks. We first propose ANOMULY, a graph neural network-based unsupervised edge anomaly detection framework for multiplex dynamic networks. ANOMULY learns node encodings in different relation types, separately, and then uses an attention mechanism that incorporates information across different types of relations. To improve generalizability and scalability, we further propose ADMIRE, an inductive and unsupervised anomaly detection method that extracts the causality of the existence of connections by temporal network motifs. To extract the temporal network motifs, ADMIRE uses two different casual multiplex walks, inter-view and intra-view that automatically extract and learn temporal multiplex network motifs. Despite the outstanding performance of ADMIRE, using it in sensitive decision-making tasks requires explanations for the model's predictions. Accordingly, we introduce an interpretable, weighted optimal sparse decision tree model, ADMIRE++, that mimics ADMIRE, to provide explanations for ADMIRE's predictions. With extensive experiments, we show the efficiency and effectiveness of our approaches in detecting anomalous connections in various domains, including social and financial networks. We further focus on understanding abnormal human brain activity of people living with Parkinson’s Disease, Attention Deficit Hyperactivity Disorder, and Autism Spectrum Disorder to show how these methods can assist in understanding the biomarkers for these diseases.

Text

2023-08-17

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/85514

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/