Sanjukta Krishnagopal | Methods in networks and machine learning: from hate speech to personalized predictive medicine | January 11, 2024

NOTE: Their is an audio issue with this video - the questions from the audience cannot be heard for the most part. Apologies for the inconvenience.

Abstract: In this talk I will discuss some aspects at the intersection of machine learning and networks, and introduce interdisciplinary methods for data-driven analysis of social and biological complex systems. 

First, I will present some nuanced aspects of social media. In recent work, we study, using knowledge graphs, sentiment analysis, and natural language processing, how online hate speech and public sentiment evolves for different marginalized groups, and how this leads to hate crimes and reactions that is different between e.g. Black Vs LGBTQ+ groups.

Next, in order to investigate this more rigorously I introduce a new mathematical model of opinion dynamics, which (unlike previous models) captures phenomena such as temporary consensus (eventually falling out of consensus) and opinion crossing. Human behaviour (e.g. masking vs not masking) is also tied to health outcomes and disease spreading. 

I will use these ideas and heterogeneous behaviour, symptoms, and genes, to present a new method in multilayer networks for early and personalized prediction of disease subtype with remarkable success in prediction of Parkinson's subtype five years in advance.

Human opinions evolve, and lead to changing alliances, behavior, and outcomes. Higher order networks such as simplicial complexes (where more than two people interact simultaneously) are particularly useful in capturing properties of higher-order human interaction effectively. I will present theory of simplicial complexes and then some applications on (a) studying strategies and influence in passing of bills in Congress, and (b) the privilege of structural position in coauthorship network on hiring using graph neural networks. 

Lastly, I will discuss topics in explainable machine learning, especially for prediction on graphs. Machine learning is tremendously successful but often a black box. Here, I will present some recent rigorous results on understanding what precisely are the weight dynamics in the graph neural network during 'learning', where we show how properties of the graph (the spectrum of the Laplacian) plays a role in training. I will also briefly discuss work on more 'biological' ways of training machine learning models of cognition. 

Bio: 
Sanjukta's research interests are multidisciplinary, with the goal of developing tools to answer questions about real world systems. She received her PhD from the University of Maryland where she was also a fellow of the Combine (Computational and Mathematics in Biological Networks) program. She was then a postdoc at University College London, where she worked with Google DeepMind on computational models of cognition. She is currently a UC Presidential postdoc with a joint appointment at UC Berkeley CS and UCLA Math. She enjoys traveling to remote corners of the world and has lived on four continents. In her free time she enjoys dancing, diving and hiking.