Research on Recommendation Method Based on Multi-Convolution Neural Network

Abstract

Traditional recommendation methods mainly use the relational data of users and projects to extract characteristics of users, which is highly dependent on the relationship data and ignores the feature learning of the data of users and projects resulting in low recommendation accuracy. Moreover, the feature extraction of traditional recommendation algorithms principally rely to manual design causing the high time complexity of constructing feature model . In view of the above problems, this paper proposes a recommendation method based on multi-convolution neural network, which takes relational data as the basis of regression training. In addition to extracting the deep features of projects and users data, convolution neural network extracts the features of text data of projects or users, and constructs a recommendation model that can mine the deep features of users and projects. With Movielens movie data set as data source, this experiment is based on collaborative filtering, full-connected neural network, title text convolution neural network and the proposed recommendation method. The results show that MAE and RMSE have been optimized with multi-convolution neural network.