DeepNNMF:deep nonlinear non-negative matrix factorization to address sparsity problem of collaborative recommender system

A recommender system (RS) is a data filtering technique that suggests the appropriate information to the end-user. Collaborative filtering is the most frequently deployed algorithm in this domain, which analyzes the users’ past behavior and recommends the products to an intended user based on correlation score. However, this technique often faces data sparsity problems. We propose a deep nonlinear non-negative matrix factorization (DNNMF) technique to address the above problem. First, we impose non-negative constraints in the embedding layer to generate non-negative vectors. Then pass them to the deep neural networks (DNN) to extract the nonlinear interactions between the users and products. The latent features and parameters are simultaneously updated using Adam optimizer. Experimental outcomes on MovieLens 100K and 1M datasets signify that the proposed model improves MAE by 3.2% and RMSE by 5.8% than the best benchmark techniques. Also, from the sparsity analysis, it is observed that the proposed model handles the sparsity issue of CF.