Font Size: a A A

Research On Mining Individual User Behavior On Online Social Network

Posted on:2017-02-20Degree:DoctorType:Dissertation
Country:ChinaCandidate:X TangFull Text:PDF
GTID:1368330542492957Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Online Social network(OSN)has been a main tool for people to maintain social relations.Usually,users share information,show attitudes and comment on hot topics in OSN.With information propagating through OSN,there are various kinds of security,fraud and privacy issues,which bring new challenges for the management of the network information.In order to deal with these issues,research on how user behaves in propagating information is worthy.User propagate information in OSN,and conduct information propagation.Moreover,relationships between users are the main means for propagating.Therefore,investigating individual user behavior is helpful for providing a deep understanding of information propagation,and also gain insights to the motivation and rule of information propagation.In this dissertation,we study four individual user behaviors related to information propagation.They are respectively spam user behavior,individual user retweet behavior,user link behavior and influence behavior.The main contributions are outlined as follows.1.We propose semi-supervised machine learning algorithms for dealing with hand labeling in spammer detection.Usually,methods for spammer detection fall into two categories.One is supervised learning,which in great need of labeled data.Another one is unsupervised learning,which requires additional information to ensure performance.In this dissertation,we introduce active learning framework into spammer detection.The proposed algorithm is able to achieve comparable accuracy with less hand labelings.Moreover,the self-paced learning is set as a complement component of active learning,leads to Active Self-paced Learning(ASPL)paradigm.The ASPL is further formulated as optimization problem,which well explains the ASPL.According to the experiments on Twitter and Weibo dataset,proposed algorithms are able to achieve comparable classification accuracy with less hand labeling work.2.We propose an algorithm for predicting individual retweet behavior based on multitask learning.Predicting individual retweet behavior is mainly based on individual retweet dataset.However,individual retweet dataset is sparse,which is not suitable for predicting behavior.We introduce multi-task learning to solve this problem.Based on the data analysis of individual retweet dataset,we connect each individual task with retweet behavior similarity according to the behavior homogeneity theory in sociology.Taking individual factor and similarity factor into consideration,our model solves the sparse problem in individual retweet behavior prediction.Besides,this dissertation simulates the information propagation with individual retweet probability on real social network,which indicates that the propagation model considering individual retweet probability is more close to the reality.3.We propose social circle detection algorithms based on structure and topic information respectively.User link behavior leads to ego network,while ego network is the main means for ego to communicate with others.Social circle detection is the main method to solve information overload in ego network.This dissertation proposes a novel social circle detection algorithm based on ego network structure.After analyzing the insufficiency of structure algorithm,we further propose a social circle detection algorithm incorporating structure and topic distribution in the information propagation.Experiments on weibo dataset show the effectiveness of the algorithm.Moreover,the simulation on real social network is conducted to show the effect of local social circle on information propagation.4.We propose a data-driven multi topic user influence ranking algorithm.User influences are originated from information propagation,while also act on information propagation.Usually,influence ranking is based on theoretic model,which fails to reflect influence in real data.We propose a data-driven method for finding user influence ranking.We analyze the user influence behavior in real social network,further divide user influence into direct influence and indirect influence,propose a multi-topic user influence model.Based on this model,we introduce a ranking algorithm related to topic,the ground nodes represented as topic pools are used in algorithm.Experiments indicate that topic-related ranking algorithm is able to rank user influene according to specific topic,and data-driven methods ensure the prediction of real influence.
Keywords/Search Tags:Online Social Network, Individual user behavior, Active learning, Self-paced learning, Multi-task learning, Ego network, Influence ranking
PDF Full Text Request
Related items