Research On Collaborative Recommendation Based On Network Community Analysis

With the development of networks, the amount of information increases beyond our ability to process it. The clients may feel that they are submerging in the information sea and difficult to find what they need. The information overload is getting worse and worse. Recommendation technology attracts more attentions of scholars and internet users in that it is expected to help clients efficiently find what they need online.The development of recent social network services (SNS) and network communities hinders recommendation technology to work well under simple relationships between various resource objects and users. The user's requirement for resources cannot be simply fulfilled by personal recommendation. Better recommendation can be regarded as a dialectic integration of users'personalities and their overall characters. This paper builds the entire recommendation system with network collaboration, which will affect the information exchange between people and others. As for many problems in collaborative recommendation: sparseness, cold start, scalability, interest change, cognitive feedback and other issues, this article presents some methods to solve those problems based on the theory of network community. The research works are listed as follows:1. The network architecture of recommendation system on expanding user-object network is presented. The recommendation network system is divided into three layers:user-resource object bipartite network, resource object network and user social network. (1) User-resource object network:users and resource objects constitute the network of relationships. (2) Resource object network:resource object network can be established with the relationships of resource objects. (3) User social network:the online relationships among different users.The three-tier network better describes the recommendation system because some recommendation methods can be looked as special cases of the system. Significantly, the network community constitutes special unit of recommendation system, being regarded as theoretical and practical base for solving many recommendation problems.2. The network communities of recommendation system are analyzed and theoretical and practical significance of network community-based collaborative recommendation are discussed in detail. A new network community detecting algorithm based on representative energy is presented. The integration of reality and the theory of network community i.e. socialization of recommendation system makes the study of issues be more accurate and efficient. In this paper, the limited resolution of network community discovery based on the use of "modularity" is described in detail. A new network community discovery method based on competing representative energy is proposed. This method does not require modularity optimization; rather it depends on the behavior of competing for community representatives. The method can divides network into those communities, it can also draw the skeleton of each community directly. Since the network connections within community are denser than those between communities, the members can obtain higher representative energy from those nodes in same community than from nodes in other communities. Community representatives emerge through competitive process naturally, and the number of representatives is the number of communities. When community affinities of entire network community are not obvious, the number of representatives who win out in the competition will decrease. Therefore, compared with other methods of network division, this method based on energy competition can show the natural law of evolving groups.3.In this work, a network community-based collaborative filtering recommendation method is proposed. The method is the combination of two tasks i.e. division of the network and collaboration recommendation within community, which consistently predicts the preference of user depending on his nearest neighbors. First of all, representative energy analysis is employed to discover the community in co-rating network. Then, the community ratings are normalized so that collaborative recommendation can implement with the similarity of preferences between users. The nearest user neighbors and the nearest object neighbors constitute the nearest community which is the nearest space to target prediction, and the community can be projected into the nearest neighbor matrix. Moreover, some empty data in the matrix can be filled with values which are predicted with object-based collaborative filtering algorithm. User's nearest neighbors are selected from candidates, which contributes to predict different objects. The nearest community-based collaborative filtering can approximate the target forecast in a super-linear way.4. A time-weighted network recommendation method is presented. This method weighs the edge of resource allocation network with time and models the dynamic network of recommendation system, and analyses the influence of interest varied with time in user-object network, which describes the decay of interest through time attenuation. The latter selecting can make bigger contribution to the energy of recommendation. As the interest diversion is concerned, the paper presents diversion delay factor to model the recommendation flow among different resource objects. Redundant energy transition is discovered by analyzing recommendation energy; the paper tries to eliminate the redundant energy with overlapping degree of transfer node. The recommendation method based time-weighted network enable the recommendation system with structural dynamic prediction ability.5. This paper proposes a method of human-machine collaborative aggregating recommendation based on network cloud community, and constructs a Bayesian feedback cloud model. Under the analysis of cloud computing, an aggregating recommendation method based on network cloud community is presented. A Bayesian feedback cloud model is constructed with the combination of human being's apriority and feature of cloud model. Based on analysis of transformation between quantitative and qualitative measurement of uncertain concepts, the paper designs cloud drop tester in detail, and presents statistical description of Bayesian feedback cloud. Preferences cloud can be obtained by using the model within the clouds community, and the prediction for recommendation can be aggregated on cloud community with preference. The method based on object cloud community has been applied in alleviating new user cold-start problem in recommendation.