Network quantification despite biased labels

The increasing availability of participatory web and social media presents enormous opportunities to study human relations and collective behaviors. Many applications involving decision making want to obtain certain generalized properties about the population in a network, such as the proportion of actors given a category, instead of the category of individuals. While data mining and machine learning researchers have developed many methods for link-based classification or relational learning, most are optimized to classify individual nodes in a network. In order to accurately estimate the prevalence of one class in a network, some quantification method has to be used. In this work, two kinds of approaches are presented: quantification based on classification or quantification based on link analysis. Extensive experiments are conducted on several representative network data, with interesting findings reported concerning efficacy and robustness of different quantification methods, providing insights to further quantify the ebb and flow of online collective behaviors at macro-level.