Anomaly detection and machine learning using graph patterns
Abstract
Relational data is often represented by directed graphs, where nodes represent entities, and directed edges represent relations between these entities. The rapid accumulation of such graphs and the need to automatically extract patterns from such graphs have led to the emergence of two related domains in graph theory: Machine Learning (ML) methods to classify nodes/edges and Anomaly Detection (AD) methods to detect anomalous behavior of graphs or specific nodes/edges. Such methods are now applied using deep networks formalisms. Current methods suffer from two main limitations that we plan to overcome: A) The implicit assumption that similarity = proximity. While this assumption is very often true, it misses an important group of distinct vertices with similar classifications. B) Ignoring the edge direction and only using the proximity in the graph to perform classification. In directed networks this misses a lot of important structural information contained in the edge direction and node topology. Our main goal is to overcome these two limitations, and propose formalisms to perform ML and AD explicitly using the information available from the direction of edges and topology of nodes in directed graphs. We plan to adapt existing classical and deep learning methods and develop new deep learning methods to incorporate the network detailed topology for ML and AD in both static and dynamic framework, using a combination of graph convolution networks and recurrent networks.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 15, 2019
- Source ID
- N629091912097
Entities
People
- Yoram Louzoun
Organizations
- Bar-Ilan University
- Office of Naval Research
- United States Navy