A Combinational LDA-Based Topic Model for User Interest Inference of Energy Efficient IPTV Service in Smart Building

Internet multimedia computing and service is becoming one of the main contributors to global energy consumption in smart building for the coming decades. In particular, Internet protocol television (IPTV) has been regarded as an important source for the Internet multimedia computing and service in our daily lives. With the development of IPTV user interest inference for network resource allocation and energy efficient design, the quality of experience (QoE) has been considered as one of the most prominent evaluation indicators for IPTV service. However, due to the explosive increase of TV programs and user's mixed interests, how to quickly infer user interest and thus effectively allocate IPTV network resources are still a challenging and urgent problem. Moreover, with the increasing amount of user data, the requirement of real-time processing is more important than before. To handle these issues, we first put forward a new indicator called watching ratio to calculate the proportion of program's time watched by user, which could measure user's subjective watching QoE from objective perspective. Then, considering both the users' current interests and their prospective interests, we propose a combinational latent Dirichlet allocation (cLDA) model and design the associated batch-based Gibbs sampling algorithm for user interest inference. Furthermore, considering the real-time processing demand in applications, we also design an online-based inference algorithm for the proposed cLDA (online-cLDA) model. Experimental results show that the proposed models and algorithms can improve the accuracy of user interest inference which can be beneficial to IPTV resource allocation in an energy efficient design.