Event Date: July 21, 2016 16:15
Deep Learning for Big Graph Data
Big data can often be represented as graphs. Examples include chemical compounds, communication and traffic networks, and knowledge graphs. Most existing machine learning methods such as graph kernels do not scale and require ad-hoc feature engineering. Inspired by the success of deep learning in the image and speech domains, we have developed neural representation learning approaches for graph data. We will present two approaches to graph representation learning. First, we present Patchy-SAN, a framework for learning convolutional neural networks (CNNs) for graphs. Similar to CNNs for images, the method efficiently constructs locally connected neighborhoods from the input graphs. These neighborhoods serve as the receptive fields of a convolutional architecture, allowing the framework to learn effective graph representations. Second, we will discuss a novel approach to learning knowledge base representations. Both frameworks learn representations of small and locally connected regions of the input graphs, generalize these to representations of more and more global regions, and finally embed the input graphs in a low-dimensional vector space. The resulting embeddings are successfully used in several classification and prediction tasks.
Bio
Mathias Niepert is a senior researcher at NEC Labs Europe in Heidelberg. From 2012-2015 he was a research associate at the University of Washington, Seattle, and from 2009-2012 also a member of the Data and Web Science Research Group at the University of Mannheim. Mathias was fortunate enough to win awards at international conferences such as UAI, IJCNLP, and ESWC. He was the principle investigator of a Google faculty and a bilateral DFG-NEH research award. His research interests include tractable machine learning, probabilistic graphical models, statistical relational learning, digital libraries and, more broadly, the large-scale extraction, integration, and analysis of structured data.
Sildes from Topical Seminar
The slides from Mathias Niepert's talk can be found here.
