My interests lie at the intersection of graph mining and machine learning. Much of my research is about creating efficient representations of various graph-structured objects that can be used in downstream machine learning tasks.
KDD 2018, Audience appreciation award runner up
Comparing graphs is one of the cornerstones in graph analytics. Expressive and efficient measures for graph comparison are very rare - and there were none that scaled to both large graphs and large graph collections. We propose the Network Laplacian Spectral Descriptor (NetLSD): the first theoretically grounded and efficiently computable graph representation that allows for straightforward comparisons of large graphs.
Node representations (aka graph embeddings) are a key tool for modern machine learning on graphs. Past research has addressed the problem of extracting such representations by adopting methods from words to graphs, without defining a clearly comprehensible graph-related objective. We show how the objectives used in past works research implicitly utilize similarity measures among graph nodes. We propose VERtex Similarity Embeddings (VERSE), a simple and efficient method that derives graph embeddings that explicitly preserve the distributions of a selected vertex-to-vertex similarity measure such as personalized PageRank.