A Multidimensional Link Prediction Method Based on the Node itself and Common Neighbors Contributions

Temporal complex networks can effectively describe the elements and their interrelationships in real-world complex systems, where link prediction is used to identify possible future links between two nodes, such as recommending new friends in social networks. Most previous work can be classified into two categories, learning-based and similarity-based, where learning-based methods lack interpretability and have high complexity, and similarity-based methods tend to focus only on local or global single-dimension network features and are limited to static networks. In fact, an effective linking method requires more topological information to be extracted from the network, and real-world complex systems are often temporal in nature. In this paper, we propose a multidimensional link prediction method based on node's own contribution and public neighbor's contribution, where we firstly consider node's own activity as the initiation probability, secondly consider node pair's public neighbor similarity as the positive contribution, and lastly, consider public neighbor's influence as the negative contribution, thus combining the node, network local and network global multidimensional structural features to perform node pair similarity metric. We conducted extensive experimental evaluations in real-world networks using AUC as an evaluation metric and compared it with eight benchmark algorithms, and the results show that our proposed method achieves high accuracy with small time complexity.