Researches Of The Potential User Mining Based On Complex Network

Audience targeting is to identify the potential users who has the similar behaviors with the seed users by the users' behaviors on the internet.In order to help the enterprises or advertisers pick out the potential consumers,the research of potential users mining based on the different topologies of the complex social network is conducted.According to the different topologies of complex network,we have studied user mining based on user behavior,user mining based on location and user behavior,as well as user mining based on interest tags and user behavior.The content includes:Firstly,on user mining based on user behavior network,we study the problem about user mining based on multiple behavior relations,and propose a method based on the combination of community detection and link in the complex network.The community is mined by the method based on network integration or the method based on similarity and modularity optimization in the complex networks at first.The community detection based the method of network integration is that the soft community indicators are obtained by Laplace operator and are used to cluster for detecting communities.The community detection based on the method of similarity and modularity optimization is that the similarity between two users is calculated in terms of their neighbors and the connection between the users,the similarity matrix is established and is divided by the modularity optimization.In order to prediction the link between the user and the seed user,the method of user mining based on the multiple link relations is proposed in the paper,which firstly the link intensity between two users on the single behavior relation is calculated by their common neighbors,the link intensity on all the relations is calculated by the relation calculus and pick out the users who have a high link intensity with the seed users as the potential users.Secondly,on user mining based on the users' location and behavior network,we focus on the problem about user mining based on the users' location and behavior network and propose a method based on the combination of community detection and neighbor selection in the heterogeneous information network.The communities are detected by the method based on multimode network or R-L model at first.The community detection based on multimode network is that the Laplace operator of the nodes of each type is calculated and the soft community indicator for each operator is attained,the joint soft community indicators of all the operators are established and are used to cluster.The community detection based on the R-L model is that the connections between two users are mapped to a relation node,and a new relation network which is composed of multiple relation nodes is established by the Tanimoto coefficient between two nodes.Then the new network is divided into lots of relation communities by the existing algorithms,and the relation communities are mapped to the corresponding user communities.In order to predict the connection between the target user and geo-location,the neighbor selection strategy on the multiple relations is proposed in the paper.All the neighbors which have a strong relationship with the target user are picked out by a variety of relations,and then the proportion of the neighbors who connect to the target geo-location is regarded as the link intensity between the target user and the target geo-location,and the users who have the high link intensity with the target geo-location are picked out as the potential users.Thirdly,on user mining based on the users' tags and behavior network,we focus on the problem about user mining based on the users' interest tags and behavior network and propose a method based on the combination of meta-path and supervised learning.The meta-path features between user and tag are defined by a variety of connections in the heterogeneous network,such as the connection between two users,the connection between user and tag,the connection between two tags.The meta-path feature value between user and tag is calculated and the feature vector is formed.In order to the potential tags of the target users,the two methods,which are called as the tag prediction based on logistic regression model and the tag prediction of tag based on community detection,are proposed in the paper.The tag prediction based on logistic regression model is that the model is established in terms of the meta-path feature vectors in the training data,and the user's potential interest tags are predicted by the meta-path feature vector using the logistic regression model.The tag prediction based on community detection of the interest tags is that the overlap communities are detected about all the interest tags in the heterogeneous network.Then the classifier for each interest tag community is learning by the training data and the meta-path feature vector between user and tag is classified by exploiting all the classifiers,the frequency rate of each tag is calculated in terms of the classify results and the interest tags which have the high frequency rate are picked out as the potential tags of the user.Finally,the potential users who have the given tags are picked out by the link between user and tag.In summary,we have studied user mining in the different network topologies,which is useful and brilliant perspective on precision advertising.