Abstract:

Personalized recommender systems are playing an increasingly important role in the online consumption platform.Low-rank and deep matrix factorization have been widely used in recommendation systems to optimize the recommendation performance.In order to overcome the bilinear property of traditional matrix factorizations,deep matrix factorizations establish the deep neural network models based on the feature vectors of users and items.The existing methods show a poor performance and a long running time when the data scale is large and the sparsity is high.For this purpose,a new deep matrix factorization model is proposed whose input is the latent feature vectors of users and items.The network structure is composed of two parallel multi-layer perceptrons and a weighted inner product operator for prediction.For the proposed model,a two-stage solution method is designed.First,the low-rank matrix fitting algorithm is used to complete the missing data so that two latent feature matrices are determined simultaneously.Then,the generated feature engineering is fed into the deep neural network and a nonlinear mapping is established with the output as the prediction score.The effectiveness of the proposed method is verified in public recommendation data sets.Experimental results show that the proposed method greatly improves the recommendation performance compared with the traditional matrix factorization methods and that compared with the existing deep matrix factorization methods,the running time is significantly reduced.

Key words: recommendation systems, low-rank matrix fitting, deep matrix factorization, deep neural network, deep learning