<p>The goal of link prediction is to derive the probability of current or future nodes forming links according to the existing network structure information. At present, the algorithm based on non-negative matrix factorization has become popular because of its dimension reduction and high prediction accuracy. However, existing algorithms based on non-negative matrix factorization encounter the following challenges: (1) NMF does not have the ability to directly capture neighbors of nodes. (2) NMF cannot fully mine global structure information, especially in sparse networks. To deal with the above shortcomings, a link prediction model called graph regularized symmetrical non-negative matrix factoring via degree-related spectral clustering and topological structure (GSNMFDT) is proposed in this paper. Specifically, firstly, the similarity of node clustering coefficients coupled with resource allocation is constructed to mine implicit information about the common neighbors of nodes. Secondly, a degree-related spectral clustering similarity is developed to fully explore the clustering information of the whole network nodes. Finally, we collect all these terms with symmetric non-negative matrix factorization loss functions into a unified link prediction model to preserve different types of network structures and countering network sparsity. Then, we design a linear scaling factor multiplication update rule to optimize the proposed model and provide proof of convergence. Extensive experiments results performed on ten real-world networks demonstrate that the proposed method outperforms the state-of-the-art methods.</p>