TY - JOUR
T1 - Multi-Topic Tracking Model for dynamic social network
AU - Li, Yuhua
AU - Liu, Changzheng
AU - Zhao, Ming
AU - Li, Ruixuan
AU - Xiao, Hailing
AU - Wang, Kai
AU - Zhang, Jun
PY - 2016/7/15
Y1 - 2016/7/15
N2 - The topic tracking problem has attracted much attention in the last decades. However, existing approaches rarely consider network structures and textual topics together. In this paper, we propose a novel statistical model based on dynamic bayesian network, namely Multi-Topic Tracking Model for Dynamic Social Network (MTTD). It takes influence phenomenon, selection phenomenon, document generative process and the evolution of textual topics into account. Specifically, in our MTTD model, Gibbs Random Field is defined to model the influence of historical status of users in the network and the interdependency between them in order to consider the influence phenomenon. To address the selection phenomenon, a stochastic block model is used to model the link generation process based on the users' interests to topics. Probabilistic Latent Semantic Analysis (PLSA) is used to describe the document generative process according to the users' interests. Finally, the dependence on the historical topic status is also considered to ensure the continuity of the topic itself in topic evolution model. Expectation Maximization (EM) algorithm is utilized to estimate parameters in the proposed MTTD model. Empirical experiments on real datasets show that the MTTD model performs better than Popular Event Tracking (PET) and Dynamic Topic Model (DTM) in generalization performance, topic interpretability performance, topic content evolution and topic popularity evolution performance.
AB - The topic tracking problem has attracted much attention in the last decades. However, existing approaches rarely consider network structures and textual topics together. In this paper, we propose a novel statistical model based on dynamic bayesian network, namely Multi-Topic Tracking Model for Dynamic Social Network (MTTD). It takes influence phenomenon, selection phenomenon, document generative process and the evolution of textual topics into account. Specifically, in our MTTD model, Gibbs Random Field is defined to model the influence of historical status of users in the network and the interdependency between them in order to consider the influence phenomenon. To address the selection phenomenon, a stochastic block model is used to model the link generation process based on the users' interests to topics. Probabilistic Latent Semantic Analysis (PLSA) is used to describe the document generative process according to the users' interests. Finally, the dependence on the historical topic status is also considered to ensure the continuity of the topic itself in topic evolution model. Expectation Maximization (EM) algorithm is utilized to estimate parameters in the proposed MTTD model. Empirical experiments on real datasets show that the MTTD model performs better than Popular Event Tracking (PET) and Dynamic Topic Model (DTM) in generalization performance, topic interpretability performance, topic content evolution and topic popularity evolution performance.
KW - Dynamic social network
KW - Influence phenomenon
KW - Multi-Topic Tracking Model
KW - Selection phenomenon
UR - http://www.scopus.com/inward/record.url?scp=84960912858&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84960912858&partnerID=8YFLogxK
U2 - 10.1016/j.physa.2016.02.038
DO - 10.1016/j.physa.2016.02.038
M3 - Article
AN - SCOPUS:84960912858
VL - 454
SP - 51
EP - 56
JO - Physica A: Statistical Mechanics and its Applications
JF - Physica A: Statistical Mechanics and its Applications
SN - 0378-4371
ER -