Learning with Graphs using Kernels from Propagated Information

Abstract

Traditional machine learning approaches are designed to learn from independent vector-valued data points. The assumption that instances are independent, however, is not always true. On the contrary, there are numerous domains where data points are cross-linked, for example social networks, where persons are linked by friendship relations. These relations among data points make traditional machine learning diffcult and often insuffcient. Furthermore, data points themselves can have complex structure, for example molecules or proteins constructed from various bindings of different atoms. Networked and structured data are naturally represented by graphs, and for learning we aimto exploit their structure to improve upon non-graph-based methods.<br /> However, graphs encountered in real-world applications often come with rich additional information. This naturally implies many challenges for representation and learning: node information is likely to be incomplete leading to partially labeled graphs, information can be aggregated from multiple sources and can therefore be uncertain, or additional information on nodes and edges can be derived from complex sensor measurements, thus being naturally continuous. Although learning with graphs is an active research area, learning with structured data, substantially modeling structural similarities of graphs, mostly assumes fully labeled graphs of reasonable sizes with discrete and certain node and edge information, and learning with networked data, naturally dealing with missing information and huge graphs, mostly assumes homophily and forgets about structural similarity. To close these gaps, we present a novel paradigm for learning with graphs, that exploits the intermediate results of iterative information propagation schemes on graphs. Originally developed for within-network relational and semi-supervised learning, these propagation schemes have two desirable properties: they capture structural information and they can naturally adapt to the aforementioned issues<br /> of real-world graph data. <br /> Additionally, information propagation can be efficiently realized by random walks leading to fast, flexible, and scalable feature and kernel computations. Further, by considering intermediate random walk distributions, we can model structural similarity for learning with structured and networked data. We develop several approaches based on this paradigm. In particular, we introduce propagation kernels for learning on the graph level and coinciding walk kernels and Markov logic sets for learning on the node level. Finally, we present two application domains where kernels from propagated information successfully tackle real-world problems.