Group Profiling For Understanding Evolution In Networks

The popularization of online social software is reshaping the way people communicate.People are no longer satisfied to join offline groups or interest communities which can create a sense of belongling,but turn their attention to joining online groups.People who share common attributes tend to form groups in online social network.Group profiling provides prior knowledge for users to decide whether to join a group.Additionally,group profile makes it possible to understand how group form and evolve.Previous research on group profiling is limited to feature selection or probabilistic models,which lacks considering diversification in the process of generating profiles.Meanwhile previous research on group evolution is limited on concluding the pattern of evolution in the view of network topology,which lacks considering group semantic features to account for the process of group evolution.Therefore,to overcome the shortcomings of previous research,this work designs a fine-granularity group profiling model based on topic sub-group and propose an algorithm to diverfy keywords in group profile.Finally,by combining group profile and network topology features,this work does an analysis on group evolution.The main work includes:Firstly,this work takes explicit group in online social network as research object.A program is developed to crawl topics information and the comments below during a certain time period.These text data will be used for constructing group profile.Besides the follow relationships between group members have also been crawled.Furthermore,a topic sub-group discovery model has been proposed to enhance content homogeneity of sub-group to quantify individual's contribution to group.The model applies dynamic word embedding and topic model to solve topic missing in short text and divide people who have similar topic distribution into the same subgroup.Based on sub-group,this work proposed an improved mixed probabilistic model to alleviate polysemy brought by irregular use of words in online social network and extract group features.The profile's diversification can be guaranteed through a feature selection method which fuses sub-group features and diversify them.Lastly,this work sends questionaires to a number of actives group members selected to acquire keywords of the group based on different time granularity.Manual ground-truth will be achieved based on these keywords.The conclusion can be made by comparing the model in this work and models in related works: the model in this work can construct more precise and more diversified group profile compared to existing models.In addition,this work quantifies the process of group evolution by showing group profiles and group topology features in multi-granularity time scale.Then,this work tries to explain how the group evolution in the view of group profile.Based on the results,this work design and implement a system for showing group profile process.