Group Behavior Modeling Based On Event Sequences Analysis

Group behavior refers to the collection of many individual behaviors in certain spatial-temporal spaces.With the scientific and technological progress and society development,group behavior analysis has attracted more and more attention from researchers in various fields and has gradually formed a cross discipline based on anthropology,sociology,statistics and computer science.Now,with the help of ubiquitous network technology represented by sensor network and mobile internet,large amounts of data related to group behaviors have been collected.These data can help us understand group behavior purposes,describe group behavior dynamics,infer influential factors of group behaviors and even understand group's classification and development.It has great research value and wide application prospect in many fields,such as social security,urban planning,business investment and conservation.For the purpose of better analysis and application of group behavior data,a group behavior model based on event sequences analysis is set up and used to model behaviors of users in Internet Protocol Television(IPTV)system.According to structures and properties of data,we analyze practical problems in data collection and analysis,and study the following three problems in depth: 1)How to analyze intrinsic structure of group behaviors based on aggregate data of event sequences? 2)How to classify event sequences from partial labeled data? 3)How to model relationships among sequences and those among events jointly? Focusing on the problems above,we propose different models and methods,respectively,and achieve encouraging results for behavior analysis,prediction and system simulation.Specifically,this thesis' s contributions include:For analyzing group behavior from aggregate data of event sequences,a mixture of Markov chains(MMCs)model and corresponding learning algorithm are proposed and applied to group flow estimation,trajectory inference and IPTV user analysis.We analyze intrinsic structure of group behaviors based on aggregate data from the aspect of transition process and propose a learning algorithm of multiple transition processes based on a mixture of Markov chains model.Given a pairwise sparse regularization on transition matrices and specified active state sets corresponding to each chain of MMCs,we learn each transition matrix and its weight in MMCs by alternating optimization.We also analyze and verify the feasibility,complexity and robustness of the algorithm.Experimental results demonstrate that the proposed MMCs model and learning algorithm have potential to applications of group flow estimation,trajectory inference and IPTV user analysis.For classifying event sequences from partial labeled data,a feature extraction method based on low-rank estimation algorithm and a classifier based on semi-supervised learning are proposed and applied to IPTV user analysis.Considering IPTV user viewing sequences as event sequences,we aggregate them based on low-rank model and obtain histogram-based viewing features.Combining quick label propagation algorithm with nonparametric estimation,a graph-based semi-supervised learning method is proposed to classify IPTV users accordingly and reflect the household structures of users.This method not only provides useful information for fast census and population structure tracking but also provides a new strategy for IPTV system simulation.For modeling relationships among sequences and those among events,a point process-based model is proposed and applied to analyze event sequences.Specifically,we propose a multi-task multi-dimensional Hawkes Process(MMHP)model and its learning algorithm,which are applied to model and analyze IPTV user behaviors.MMHP models the dynamics of multiple sequences jointly by imposing structural constraints and thus systematically uncovers clustering structure among sequences — using an intrinsic intensity matrix,a structural infectivity tensor and a triggering kernel to describe dynamics of multiple sequences.We propose an iterative optimization algorithm to learn MMHP,which takes advantages of alternating direction method of multipliers(ADMM),majorization minimization and Euler-Lagrange equations.During learning process,by imposing sparse and low-rank constraints on the infectivity tensor,the learning algorithm avoids over-fitting successfully and obtains triggering patterns within events and clustering structures across sequences simultaneously.The robustness and the superiority of the proposed model are demonstrated on both synthetic and real-word data set.Applied to IPTV user behavior analysis,the proposed MMHP model describes users' s viewing preferences,dynamics of behaviors and clustering result of users jointly.