Intelligence analysts require automated tools to mine multi-source data, including answering queries, learning patterns of life, and discovering malicious or anomalous activities. Graph mining algorithms have recently attracted significant attention in intelligence community, because the text-derived knowledge can be efficiently represented as graphs of entities and relationships. However, graph mining models are limited to use-cases involving collocated data, and often make restrictive assumptions about the types of patterns that need to be discovered, the relationships between individual sources, and availability of accurate data segmentation. In this paper we present a model to learn the graph patterns from multiple relational data sources, when each source might have only a fragment (or subgraph) of the knowledge that needs to be discovered, and segmentation of data into training or testing instances is not available. Our model is based on distributed collaborative graph learning, and is effective in situations when the data is kept locally and cannot be moved to a centralized location. Our experiments show that proposed collaborative learning achieves learning quality better than aggregated centralized graph learning, and has learning time comparable to traditional distributed learning in which a knowledge of data segmentation is needed.

Date Published: 17 May 2016
PDF: 14 pages
Proc. SPIE 9842, Signal Processing, Sensor/Information Fusion, and Target Recognition XXV, 98420O (17 May 2016); doi: 10.1117/12.2225965

Published in SPIE Proceedings Vol. 9842:
Signal Processing, Sensor/Information Fusion, and Target Recognition XXV
Ivan Kadar, Editor(s)