Research On User Trust Analysis In Online Social Networks

Trust is one of the bases for people to collectively live and work.Various online social services connect people’s living to the mobile world.By taking advantage of trust researches,online social services are greatly improved and secured,and online social networks(OSNs)related analyses and tasks are refined.Therefore,to study user trust will benefit both the OSN world and researches depend on user trust.However,challenges exist for researches on user trust in OSNs,as user trust in OSNs is not always clearly defined across studies,characteristics of user trust are various,the mechanism for forming a user trust relationship is complex,sources for user trust acquisition are implicit,etc.To study user trust in OSNs,the research of this dissertation was carried out in two folds:firstly,as trust acquisition sources for user trust evaluation,user profile components such as user identity and user influence,and user participation with respect to their belonging groups were explored;and secondly,the model for quantifying and inferring user trust in OSNs was researched and implemented.The main contributions of this dissertation are as follows:1.Based on the analyses conducted on a Weibo dataset of spammers and their posts and posting activities,a measurement combining multiple types of user characteristics and information was proposed to evaluate the level of a user’s being a spammer.Comparing to other methods that used only single type of features or made use of multiple yet complex types of features,the proposed measurement has a simple and plain form,and can be well interpreted and in accordance with authors’ real purposes of writing or posting.Experiments have been conducted to validate that the measurement can help identify spammers from the dataset.2.According to a statistical analysis on numerous information spreading events in Weibo,the findings which have statistical significance are that super-spreading in information propagation exists in OSNs and the influence of super-spreaders is quite bigger than average users.Based on compartmental models,an information propagation model,SAIR,was proposed to characterize super-spreading events and super-spreaders in information propagation.To validate such capability of the model,an experiment using a random use case of information propagation from Weibo was conducted.The proposed SAIR model can characterize information propagation events with super-spreading phenomenon more accurately than the conventional SIR model.Using the reproduction numbers calculated for different user groups in a super-spreading event,the influence of super-spreader users in such an event was quantified.3.To study how overlapping user groups affect a member’s preference may shed light on the way how user trust is evaluated.A novel method based on symmetric non-negative matrix factorization was proposed to detect communities in a network.By using this technique,the method can identify interpretable user groups in an OSN according to the relevant task’s requirement of identified groups.Based on the user-community partition matrix obtained during overlapping community detection,four strategies to aggregate user preferences were proposed.Experimental results on public datasets showed that the proposed community detection method could control how overlapping the resulting community partition was,and using the group identification method and the aggregation strategies for group recommendation tasks led to more accurate predicted results,and how overlapping user groups affect a member’s preference was further discussed.4.Based on the above studies and characteristics of user trust in OSNs,five assumptions were made to further aid trust researches.Taking the assumptions into account,various types of features that embed user trust information were proposed;a trust inference model based on conditional random field was proposed;by using loopy belief propagation as its probabilistic inference algorithm,the model has good interpretability for inferring user trust as it can be understood in the way of message passing.Experiments were conducted on a real-world dataset and the results showed that the proposed model was effective in user trust inference in OSNs.Furthermore,taking advantage of parallelized and GPU accelerated computing to achieve high performance,an implementation of undirected probabilistic graphical models based on the proposed trust inference model was presented.A performance experiment on a big real-world OSN dataset was conducted,and the result showed that the implementation running on a hybrid hardware configuration achieved 5x speedup over simple CPU configurations.