FFDNN: Feature Fusion Depth Neural Network Model of Recommendation System

Faced with the various colorful items on the ecommerce platforms, it becomes a problem that how to effectively match the items to those who need. An indispensable tool for solving the problem above is recommendation system. Focusing on the massive data faced by the current recommendation algorithms, the sparse user-project interaction matrix, the cold start of new items and users, and this paper proposes an effective neural network-based recommendation model. Firstly, the user-item scoring matrix is constructed and filled to alleviate the sparseness problem. Secondly, the user and item rating eigenvector are extracted by matrix decomposition, whiles text eigenvectors are obtained by using text description information. Furthermore, user and item feature vectors are constructed by rating features and text features together. Finally, the constructed eigenvector of users and items are used as input to the long short-term memory (LSTM). The eigenvectors of users and items are trained, and the output of the LSTM network is processed by multi-layer perception (MLP) to predict the correlation between users and projects. The experimental and comparative results show that the proposed model can effectively alleviate the cold start problem of projects and users, and improve the recommendation performance.