Graph analytics systems today must process very large graphs that have billions of nodes and edges and require several TB of storage. Since the main memory of most computers is limited to a few 100 GB, graphs of this size must be processed either on clusters or by out-of-core processing. However, both these approaches have large overheads and they support only a limited set of graph processing algorithms.
Intel Optane DC Persistent Memory is a transformative memory technology which has higher density and lower cost than DRAM, but which can be accessed efficiently at the byte level like DRAM. This enables affordable machines with several TB of memory. In this talk, we describe our experience in using such a machine for in-memory analytics of massive graphs using the Galois system.
Keshav Pingali is a Professor in the Department of Computer Science at the University of Texas at Austin, and he holds the W.A."Tex" Moncrief Chair of Computing in the Oden Institute at UT Austin. Pingali is a Fellow of the IEEE, ACM and AAAS. He received the IIT Kanpur Distinguished Alumnus Award in 2013. Between 2008 and 2011, he was the co-Editor-in-chief of the ACM Transactions on Programming Languages and Systems. He has also served on the NSF CISE Advisory Committee.
My talk starts by turning back the clock to 1979-1983, introducing the ideas that culminated with the fundamental representation theorem of graphs (the Aldous–Hoover theorem). I will then show how these ideas connect to a probabilistic interpretation of matrix factorization methods, explaining why matrix factorization is fundamentally not as expressive as it could be to describe finite graphs. I will then turn to early machine learning attempts to represent graphs and how these attempts connect to graph mining algorithms. I will introduce the concept of representation learning with graph neural networks (GNNs) and explain its connections to statistical graph models and the Weisfeiler-Lehman isomorphism test. Finally, I will introduce a newly proposed general framework for graph representation learning using deep neural networks, which is directly rooted in the ideas that gave us the Aldous–Hoover representation theorem. This new representation framework points to novel graph models, new approaches to make existing methods scalable, and provides a unifying approach connecting matrix factorization, graph mining algorithms, and graph neural networks. I will end my talk with a few open problems.
Bruno Ribeiro is an Assistant Professor in the Department of Computer Science at Purdue University. He obtained his Ph.D. at the University of Massachusetts Amherst and did his postdoctoral studies at Carnegie Mellon University from 2013-2015. His research interests are in deep learning and data mining, with a focus on sampling and modeling relational and temporal data.
Deep learning is widely use in several cases with a good match and accuracy, as for example images classifications. But when to come to complex networks there is a lot of problems involved, for example how do we represent a network in a neural network without lost node correspondence? which is the best encode for graphs or is it task dependent? Here I will review the state of art and present the success and fails in the area and which are the perspective. I will also present a novel work done using chimera, an graph mining encoding that allow us to combine link and content overtime to predict and discovery communities in complex networks.
Ana Paula Appel joined IBM Research Brazil in February 2012 and since January 2018, She is in Visual Analytics and Insight group at IBM Research Brazil. The group's mission is to conduct research projects in data-driven aiming to develop novel technologies that can help in several industries as Finance, Agriculture and Natural Resources. She is a master inventor since 2016, member of AOT since 2017 and a Certified Master Data Science since 2019. In 2011, She worked as a professor at the Computer Science Department at Federal University of Espirito Santo (CEUNES - UFES), Brazil, in Database and Data Mining area. During this period she also did a post-doc at UFSCar under the guidance of Prof. Dr. Estevam Hruschka. Before that, she received a Ph.D. (2010) and a Master Degree (2004) in Computer Science from the University of São Paulo under the guidance of Prof. Dr. Caetano Traina Jr. She also, did one year internship at Carnegie Mellon University (CMU) under the supervision of Prof. Dr. Christos Faoutsos.
Her research interests lies in machine learning and data mining, more specifically: temporal data, graph mining, graph and deep learning, financial models and risk analyses. She is active in the machine learning and data mining research communities, having being organizing WinDS - Women in Data Science in the past three edition (2017, 2018 and 2019). She also served in the program committees of the ACM International Conference on Knowledge Discovery and Data Mining (ACM-KDD), Brazilian Conference on Intelligent Systems (BRACIS), Symposium on Knowledge Discovery, Mining and Learning (KDMILE) and Simposio Brasileiro de Banco de Dados (SBBD), among others.