Abstract:
The goal of link prediction is to predict missing links and possible future links based on known network structure information. However, most existing link prediction algorithms only focus on undirected andunweighted networks but ignores the weight contribution and node neighborhood structure information, which leads to the decrease of prediction accuracy. To address this problem, a link prediction of weighted network model combining the neighborhood structure and symmetric non-negative matrix factorizationis proposed, which performs tthe tasks of missing weight prediction and robustness of the weighted network. Firstly, the adjacency matrix and its transpose sum to calculate the local similarity, and then the similarity is mapped to the low-dimensional potential space to preserve the local structure information of the network. Secondly, the minimum spanning tree algorithm is used to search node neighborhood structure information, and the neighborhood similarity matrix based on minimum spanning treeis constructed. Thirdly, to preserve node neighborhood information, the similarity matrix based on minimum spanning tree is mapped to common low-dimensional potential space to preserve the weight structure information of the entire network. Finally, a unified weighted link prediction model is constructed by integrating the above two kinds of information.The local optimal solution was obtained by learning the parameters of the model using the multiplication updating rule, and the original weighted network was reconstructed with the minimum error to obtain the prediction score matrix.Compared with the existing representative methods, experimental results on the 8 real world network show that the AUC of the proposed algorithm is improved by 3.1%