关键词: 复杂网络, 深度非负矩阵分解, 链路预测, 隐特征

Abstract:

Link prediction is predicting latent edges or unknown edge according to the existing network structure information and it has become one of the hot topics in complex networks. Traditional non-negative matrix factorization has been used in link prediction directly projects the original network into the latent space. However, non-negative matrix factorization can not fully characterize the deep latent structure information of complex networks. As a result, the prediction ability in sparse networks is limited. Aiming at the above problems, this paper proposes link prediction method based on Deep Non-negative Matrix Factorization（DNMF）. DNMF forms a multi-level network structure learning model through the multiple factorization of the coefficient matrix. By multiple factorization of the coefficient matrix, the objective function of deep latent feature model can be obtained. Pre-decomposition results are obtained by layer decomposition and then the training parameters can be adjusted as a whole. The similarity matrix can be obtained by the different basis matrices and coefficients matrix. DNMF method can guarantee the expression of deep structure of the real network information, at the same time, it can get much richer and more comprehensive network structure information. Experiments on ten typical real networks show that the proposed method has competitive performances than state-of-the-art link prediction methods.

Key words: complex networks, deep non-negative matrix factorization, link prediction, latent feature