AM Seminar: Are Graph Learning Methods Actually Learning?

Presenter: Seshadhri Comandur, Professor of Computer Science, UCSC

Description: There has been a lot of literature on graph machine learning over the past few years, and a bewildering array of new methods. This talk is based on a series of results making a provocative argument. Maybe many graph machine learning methods are not really that effective, and the progress we are seeing is an artifact of experimental design and measurement. I will talk about some results showing that low-dimensional embeddings with dot product similarity (arguably the most common graph ML technique) cannot capture salient aspects of real-world graphs. Follow-up work demonstrates that simple benchmarks seem to outperform fancier methods, and that there are significant shortcomings in existing accuracy measurement.

Bio: C. Seshadhri (Sesh) is a professor of Computer Science at the University of California, Santa Cruz and an Amazon scholar. Prior to joining UCSC, he was a researcher at Sandia National Labs, Livermore in the Information Security Sciences department, during 2010-2014. His primary interest is the theoretical study of algorithms, especially those with a mix of graphs and randomization. By and large, Sesh works at the boundary of theoretical computer science (TCS) and data mining. His work spans many areas: sublinear algorithms, graph algorithms, graph modeling, scalable computation, and data mining. In the theory world, his work has resolved numerous open problems in monotonicity testing and graph property testing. A number of his papers in the interface of TCS and applied algorithms have received paper awards at KDD, WWW, ICDM, SDM, and WSDM. He received the 2019 SDM/IBM Early Career Award for Excellence in Data Analytics. Sesh got his Ph.D from Princeton University and spent two years as a postdoc in IBM Almaden Labs.

Hosted by: Ashesh Chattopadhyay, Applied Mathematics Department