“Scalable and Robust Submodular Maximization: From Dynamic Ground-Sets to Unknown Utility Functions”

Speaker: Ehsan Kazemi

Abstract: The emergence of datasets of an unprecedented scale across different scientific disciplines presents new computational challenges. To remedy the situation, there has been a lot of interest in using combinatorial structures and information measures that allow for efficient and fast computations. In many information gathering and data mining applications, the underlying utility function we aim to optimize has additional structures, which can add to the complexity of the problem and/or provide opportunities to find solutions efficiently. We will consider optimizing a general class of discrete functions, called submodular functions, that exhibit an intuitive diminishing returns property.

It has long been known that solutions obtained from optimization methods can demonstrate striking sensitivity to the parameters of the problem. Robust optimization, in contrast, is a paradigm in the mathematical programming community with the aim of safeguarding the solutions from the changes in the underlying parameters. In this talk, we study the problem of scalable submodular maximization under the unifying umbrella of robust computation where the dataset may change over time (i.e., some portion of data or sensitive features are deleted) or the utility function is not fully known.

Speaker Bio: Ehsan Kazemi is a postdoctoral research fellow at Yale Institute for Network Science (YINS). His main research interests are machine learning, data mining and applications of these fields in computational biology and social networks. He is working with Prof. Amin Karbasi. He completed his PhD. at LCA4, EPF Lausanne. He received both his B.S. and M.S. degrees in Communication Systems from Sharif University of Technology. He won NOKIA Mobile Data Challenge (MDC) on the Next Place Prediction. He is the recipient of Swiss National Science Foundation (SNSF) Post-Doctoral Fellowship. 

.