Cooperative Communities (CoCo): Exploiting Social Networks for Large-scale Modeling of Human Behavior

Abstract

Human behavior modeling at a large-scale and under real-world conditions is still an open problem. Existing classification models do not always perform well on a diverse population. Training personalized models that incorporate different contextual conditions and individual user characteristics are effective in addressing this challenge. However, this approach burdens the users with collecting and manually labeling their own training data which is not scalable. In this article, we propose CoCo (Cooperative Communities), a learning framework that leverages different types of everyday social connections between people to personalize classification models. CoCo exploits social networks to selectively combine small contributions of labeled data from people with shared context or user characteristics. Under CoCo a personalized classifier is trained for each individual user, but by exploiting social networks, the burden of providing training data can be spread over the entire community.