PhD Application Talk: Matrix factorization for the Link Prediction Problem

The problem of link prediction has received significant attention from the research community and a number of interesting approaches have been proposed. After a review of the state of the art we present a novel method based on extrapolation in the Lie-algebra of matrices, which, to the best of our knowledge, is the first attempt to apply it in this context. Interpolation in Lie-algebra is widely used in many areas (e.g. Computer Graphics, Robotics and Computer Vision) and can be considered a mature technology, however due to the nature of adjacency matrices cannot be applied to them directly.
We compare our method's performance to two state-of-the-art approaches – Supervised Learning and Spectral Growth Model – on both synthetic (generated by common graph models) and real-world networks (built from author collaboration information of arXiv.org papers) and examine directions for future work.