You Are What You Watch and When You Watch: Inferring Household Structures From IPTV Viewing Data

What you watch and when you watch say a lot about you, and such information at the aggregated level across a user population obviously provides significant insights for social and commercial applications. In this paper, we propose a model for inferring household structures based on analyzing users' viewing behaviors in Internet Protocol Television (IPTV) systems. We emphasize extracting features of viewing behaviors based on the dynamic of watching time and TV programs and training a classifier for inferring household structures according to the features. In the training phase, instead of merely using the limited labeled samples, we apply semisupervised learning strategy to obtain a graph-based model for classifying household structures from users' features. We test the proposed model on China Telecom IPTV data and demonstrate its utility in census research and system simulation. The demographic characteristics inferred by our approach match well with the population census data of Shanghai, and the inference of household structures of IPTV users gives encouraging results compared with the ground truth obtained by surveys, which opens the door for leveraging IPTV viewing data as a complementary way for time- and resource-consuming census tracking. On the other hand, the proposed model can also synthesize trace data for the simulations of IPTV systems, which provides us with a new strategy for system simulation.