Empirical Analysis And Modeling Of The User Behavior In The Online Social Systems

Beginning with the oracle,‘know yourself',thousands years ago,we still have far way to go on the road of understanding human behavior patterns.Fortunately,the coming of Internet era,which recorded every detail of our behavior,provides us a great opportunity to study and uncovering the human behavior's patterns and mechanisms.The present thesisfocus on the user behavior in the online social systems,which has theoretical significance for better understanding human behaviors and practical significance for e-commerce and online serviceetc.Firstly,the thesis uncovers and models the memory effect of the online users' selecting and rating behavior that could reflect the periodical transfer of users' interest and tastes.Since users are allowed to rate on objects in many online systems,ratings can well reflect the users' preference.According to the method of Correlation Coefficient,we find the strong memory effect in users' selecting behavior,which is the sequence of qualities of selected objects,and the rating behavior,which is the sequence of ratings delivered by each user.In addition,the memory duration,which is presented to describe the length of a memory,exhibits the power-law distribution,i.e.,the probability of the occurring of long-duration memory is much higher than that of the random case that follows the exponential distribution.We further present a preference model in which a Markovian process is utilized to describe the users' selecting behavior,and the rating behavior depends on the selecting behavior.With only one parameter for each of the user's selecting and rating behavior,the preference model could regenerate any duration distribution ranging from the power-law form(i.e.strong memory)to the exponential form(i.e.weak memory).Secondly,based on the bipartite network theory,the thesis studies online users' interest from the prospective of local clustering properties and the object popularity respectively.The clustering coefficient of bipartite network C4,considering the properties of user-object systems,could describe the diversity of users' interests.We find that,one should not classify users only according to their activity levels,because users with the same activity level may have two totally different interest patterns,one of which is very concentrate but another is very diverse.Besides in the interest based on the structure clustering property,similar phenomenon is also found in users' preference on object popularity.Thus,according to the preference on object popularity,the thesis proposes Non-equilibrium Mass Diffusion and Non-equilibrium Heat Conduction recommendation algorithm.Compared with the classical mass diffusion and heat conduction method,the presented algorithm could largely improve the accuracy and diversity of the recommendation.The improvementproves from another perspective that,the users' preference is an important force in the evolutionof such kind of online socialsystems.At last,the thesis presents and studies the stability problem of object similarity.Similarity measuring two objects' potential relation is widely used constructing the gene co-expression networks,protein-to-protein networks,recommendation systems etc.But it would be unreliable and contains false information if the similarity is unstable,i.e.the similarity of a definite pair of objects is measured as different level before and after.In two online bipartite systems,we evaluate the stabilities of fifteensimilarity indexes when measuring object similarity.Results show that,more data could lead to more stable evaluations but most indexes except Preferencial Attachment,Common Neighbor index,Adamic-Adar index and Resource Allocation index,may have quite different evaluations using different data samples.While there are dozens of similarity indexes,most of them can be classified into three Clusters from the prospective of stability and indexes in the same cluster are generally based on the same considerations and have similar mathematical definitions.When a new index being proposed,one just need to identify which cluster it belongs to,and then could get deeper insight to this index by comparing with other indexesin the same cluster.In addition,we develop a top-n-stability method to study the object similarity stability's further effect on the recommendation.We find that,by taking only the stable similarities into account,the stability,accuracy and diversity of the recommendation could be improved.Overall,the present thesis is a significant step on the road of understanding human behavior pattern.Focusing on the online users' selecting behavior,interest preference and objects similarity stability,investigations and results in this thesis may shed some light on both theoreticalinvestigation and practical application,and attract more attention to get deeper insight in those important problems.