S2DNMF: A Self-supervised Deep Nonnegative Matrix Factorization Recommendation Model Incorporating Deep Latent Features of Network Structure

Most of the existing recommendation methods are based on shallow models that do not consider rating interaction information noise. To solve this problem, this paper proposes a recommendation model based on deep nonnegative matrix factorization (Self-supervised Deep Nonnegative Matrix Factorization, S2DNMF), which inherits the advantages of the self-supervised model, combines deep attribute fusion features of network structure, integrates network topology and sparse constraints of community membership, and achieves an intelligent recommendation task. Specifically, the model fully utilizes the observed rating information of each implicit layer by deep nonnegative matrix factorization. Then, the similarity rating is computed by using the common nearest neighbor method, and mapped to a multi-layer low-dimensional hidden space to obtain the network structure topological information of each hidden layer. Meanwhile, the \({\mathcal{L}}_{2,1}\) parametric constraint factor matrix is used at each hidden layer to remove random noise. Experimental results demonstrate the effectiveness and efficiency of the proposed method over alternatives.