Mining Interesting Meta-Paths from Complex Heterogeneous Information Networks

Meta-paths in heterogeneous information networks are almost always hand created and have, so far, only been attempted on data sets with very small type systems like DBLP, IMDB, etc. Most real-world heterogeneous information networks have large and complex type systems. As the size and complexity of the type-system grows it becomes more and more difficult for humans to form reasonable meta-path queries. This work introduces a new technique to discover a new market for data called interesting meta-paths from complex heterogeneous information networks. Our interestingness measure is based on classical knowledge discovery principles, but have been applied in such a way that only interesting meta-paths are mined from the hundreds-of-thousands of possible choices. As in classical pattern mining literature, precision and recall statistics are difficult to obtain, instead we evaluate the effectiveness of our results using a quantitative node-similarity analysis as well as a large user study. Finally, we apply the newly discovered interesting meta-paths to find similar nodes on the Wikipedia heterogeneous information networks.