Abstract:

In view of the problems of weak characterization ability and low computational efficiency of high-dimensional sparse undirected networks for mining potential correlation information among entities, this study proposes a non-negative nonlinear characterization latent factor model based on trust relationship under the framework of social recommendation model. The model shapes the feature space of the latent matrix through nonlinear mapping, which guarantees not only the non-negativity of the target matrix, but also improves the characterization ability of the model. With the introduction of the graph Laplace regularization term in the objective function of the model training, the structural consistency between before and after the mapping of the trust relationship is ensured. The experimental results based on six public data sets show that the proposed model has obvious superiority over other models.

Key words: high-dimensional sparse undirected networks, social recommendation model, trust relationship, non-negative nonlinear, feature space, graph Laplacian regularization, latent factor model, mini-batch gradient descent