Optimizing Link Prediction by Community Detection on Dynamic Networks

Link prediction is a machine learning technique employed to comprehend the structure of networks and identify missing and future connections among peer nodes on interconnection networks. One prominent category of methods used in link prediction is similarity-based approaches, which leverage both local and global topological information from the network to predict links for possible teammate workers. These methods aim to utilize local and global structural topology information on static networks. However, they are not well-suited for dynamic networks where interconnection nodes and links change over time. To this end, in this work we propose the use of community detection for link prediction on dynamic networks, pursuing a balance of accuracy and efficiency. We present a community detection based link prediction framework, utilizing parameterized influential regions of nodes and their contributions in community partitions to predict potential combinations of counterpart nodes. The community information based features can be also applied to optimize the machine learning models for link prediction based on other graphical and similarity-based features. Experimental evaluation results demonstrate that our community detection based approach enhances the link prediction model on the DBLP database to achieve a significant accuracy, with an AUCPR score exceeding 0.98.