CMSA New Technologies in Mathematics: Word and Graph Embeddings for Machine Learning
Steven Skiena - Stony Brook University
DeepWalk is an approach we have developed to construct vertex embeddings: vector representations of vertices which be applied to a very general class of problems in data mining and information retrieval. DeepWalk exploits an appealing analogy between sentences as sequences of words and random walks as sequences of vertices to transfer deep learning (unsupervised feature learning) techniques from natural language processing to network analysis. It has become extremely popular, having been cited by over 4600 research papers since its publication at KDD 2014. In this talk, I will introduce the notion of graph embeddings, and demonstrate why they make such powerful features for machine learning applications. I will focus on more recent efforts concerning (1) fast embedding methods for very large networks, (2) techniques for embedding dynamic graphs, and (3) embedding spaces as models for knowledge generation.