A Cluster-Based Prediction Method of Specific Social Networks User Relationship: Take Academic Blog as an Example

Social networks user relationship prediction, also known as social network link prediction, predicts the probability of edge-linking between users by analyzing the network topology between users and its related algorithms, such as friend recommendation, association interaction. There are many methods of link prediction, which are mainly for the research of large-scale blog network, such as Twitter, Micro-Blog,and so on. The typical feature of such social networks is the more friends a user has, the higher active-level and the more blog posts. User relationship prediction methods usually focus on the study of network topology without considering other characteristics of users. However, through the research on some specialized blog networks, it is found that their network structure is different from that of general social networks, such as blogs in the academic fields as well as in the medical ones. The prominent feature of the latter is that the number of their active users’ friends is relatively small instead. Therefore the accuracy rate will be affected for such academic blog net-user to predict user relationship in a general link prediction method. It is also found that, due to a need for certain registration conditions, users of these academic blog networks have some highly recognizable characteristic information, such as research fields, geographical distribution, etc. This paper proposes Cluster-CAR, a two-step prediction model that integrates K-prototype clustering for mixed attribute data (numerical and categorical) and refines link prediction using a modified CAR similarity index within homogeneous clusters. By first clustering users based on domain-specific attributes, our method mitigates network sparsity and isolates structurally coherent subgroups. Subsequent intra-cluster link prediction leverages localized community structures, enhancing accuracy while reducing computational overhead. Experiments on academic blog data demonstrate that Cluster-CAR achieves a prediction accuracy of 0.973— outperforming baseline algorithms (CN, AA, CAR)— and exhibits lower computational complexity. This hybrid approach uniquely bridges attribute-driven clustering with topology-aware link prediction, addressing the nuanced demands of specialized social networks.