Discriminative Learning of Social Dynamics

Опубликовано 30 марта 2017, 22:04

As a field concerned with prediction, machine learning is a natural candidate for approaching social-related prediction tasks. However, in contrast to many other domains were machine learning has been successfully applied, social phenomena remain notoriously hard to predict. In this talk I will argue that in addition to a large and rich data and powerful computational resources, the success of a learning approach critically depends on its ability to incorporate prior knowledge. It is therefore natural to ask – how can social and behavioral theory be incorporated into predictive learning frameworks?
I will discuss ways in which this can be achieved, and present a general method for incorporating models of social behavior. Our method learns a class of predictors inspired by generative models of social dynamics, but under a discriminative objective. Thus, it utilizes the explanatory power of social models, while optimizing predictive accuracy. This is achieved by efficiently learning an optimal (linear) combination of (non-linear) social-dynamic models, which implicitly serve as an infinite, continuous set of features in a dual kernel space. I will present results on the tasks of influence estimation and network value prediction, and elaborate on possible extensions to tasks in domains like behavioral economics, crowdsourcing, and others.
Finally, I will discuss how ad-hoc predictive methods can be used to expose and illuminate social phenomena and support behavioral hypotheses, and how such goals can motivate challenging and novel problems for machine learning.

See more on this video at microsoft.com/en-us/research/v...