In this talk I will present several works on the interplay between machine learning and social dynamics. These revolve around three themes.
The first theme explores how machine learning can be applied to predicting the outcomes of social processes. For the task of predicting the spread of information in networks, we propose a hybrid approach for discriminatively training generative models of information propagation. Learning and inference in this approach can be done efficiently via a certain kernel embedding technique, which can, in principle, be applied to any generative model.
The second theme focuses on using social-dynamic processes as inspiration for machine learning algorithms. For the task of graph-based semi-supervised learning, we suggest a method for propagating label information based on competitive infection dynamics. We analyze our method from multiple perspectives, uncover the underlying learning objective, and show a surprising relation to graph Laplacians. Our method applies to multiple learning settings and is highly competitive across diverse datasets.
The third theme considers the task supporting decision makers in online social systems. We take the perspective of content creators in social sharing platforms, and design a tool for choosing a set of tags that optimizes incoming traffic to an item. We show that the social dynamics in these systems lead to an objective that is monotone and submodular, and can thus be optimized with a greedy algorithm. We provide an efficient implementation of the algorithm, and evaluate its performance on data from real tagging systems.
Nir Rosenfeld is a Postdoctoral Fellow at Harvard's Center for Research on Computation and Society (CRCS), where he is advised by Yaron Singer and David Parkes. Prior to that, he was a Ph.D. student at the Hebrew University in Jeruslaem, where he was advised by Amir Globerson. He was also a long-term intern at Microsoft Research in Israel. Rosenfeld's main research interest is to develop machine learning methods for tasks involving dynamic social and behavioural data, and to apply them in order to gain insight into social and behavioral processes. He also draws on social dynamics as inspiration for
designing learning and inference algorithms. Lately, he has become interested in various aspects of incoporating humans into the learning process.