Collective Behavior Classification In Social Networks

With the rapid development of social networks,people's activities on social networks have become increasingly frequent.The interactive activities of the online community generate a large number of collective behaviors.These collective behaviors data are full of valuable knowledge,and present opportunities and challenges to the study of collective behaviors in a social networking environment.How to find out the interaction and influence mechanism of collective behaviors from the social network data and achieve effective classification of collective behaviors has become one of the hot fields in social network these days.However,compared with the traditional classification data,individuals in a social network interconnect through different types of links.Therefore,the training samples and the testing samples are in the same network and difficult to be divided.In addition,the classification of collective behaviors is different from the traditional classification of objective things,which is the classification of people's behaviors in social networks.The underlying mechanisms that impact the behaviors of individuals are the influences between them,which are challenging for traditional classification algorithms.Thus,classifying collective behaviors from social network data has important research and application value.In this dissertation,we focus on multi-label classification on networked data,for which behaviors are represented as labels and an individual can be assigned multiple labels.Our main contribution are as follows.(1)We propose a novel algorithm,called MORN(Multi-Order Relational Neighbor classifier)to mine second-order neighbors for non-isolated nodes and high-order neighbors for isolated nodes.Existing relational classification methods classify an individual based only on the individual's first-order neighbors,this may lead to the connectivity have not been made full use of.In addition,networked data often contain sub-networks which are known as isolated components,and all these isolated nodes'labels cannot be predicted by using traditional relational classification methods.We test MORN on real datasets and it demonstrates that MORN can classify the labels of isolated nodes,and it has better classification performance than traditional relational classification methods.(2)We propose a novel Multi-label relational Classifier by Clustering Analysis(MCCA)method to classify collective behavior from networked data with a large number of multi-labeled nodes.We first identify similar nodes for each unlabeled node based on local network structure.Then we perform clustering on nodes with known labels.We introduce an aggregated class probability to capture the correlations between nodes and class labels based on the clustering results.Experiments with real-world datasets demonstrate that our proposed method improves classification performance comparing to the existing approaches.(3)We propose a novel Multi-label Classification algorithm which distinguishes Peer influence and Personal preference(MCPP).Existing relational learning methods exploit the connectivity of individuals and they have shown better performance than traditional multi-label classification methods.However,the behaviors of an individual can be influenced by factors other than connectivity,e.g.personal preference,which is not captured by connectivity-based methods.We innovatively apply propensity score matching to identify and quantify the causal effect of peer influence on a node's labeling and thus to obtain the weights of peer influence and personal preference regarding their respective contributions to the labeling of a node.The weights are then used in the design of a multi-label relational classifier.Experiments on real-world datasets demonstrate that our proposed methods improve classification performance over existing methods.