An Integrated Recommendation Model Based on Two-stage Deep Learning

In recent years, deep learning technology has been widely used in the field of recommendation systems and has achieved great success. However, the input quality of the deep learning models has a great influence on the learning results. A sparse input feature vector will not only increase the difficulty of subsequent model training, but also will lead to the learning results falling into local optimum. In this article, an integrated recommendation model based on two-stage deep learning is proposed. Firstly, two individual marginal stacked denoising auto-encoders (mSDA) models with closed-form parameter calculation are used to extract the high-level abstract features of the users and the items. Then the resulted user abstract feature and the item abstract feature are connected as the input vector of the deep neural network (DNN) model, and the parameter learning and model optimization are performed through joint training. In addition, in order to model low-order feature interactions, a logistic regression model based on original feature vector is also integrated into the recommendation model. Extensive experiments with two real-world datasets indicate that the proposed recommendation model shows excellent recommendation performance compared with the state-of-the-art methods, especially in the data sparse and the cold start environments.