Personalized recommendation model based on users multi-features fusion

Abstract

Personalized recommendation is one of the most effective means to extract specific information.Aiming at the problem that lacking of users feature in traditional recommendation methods, a hybrid recommendation model of generalized matrix factorization and deep long short-term memory network based on multi-features fusion (HMF) was proposed.The model used potential eigenvector factors to characterize the selective impact of age and occupation on the project, highlighting user personalization.Long short-term memory (LSTM) was used to obtain the temporal characteristics between users and projects, and then the deep temporal nonlinear higher-order interaction relationship was mined by multi-layer perceptron.The simple nonlinear low-order interaction obtained by generalized matrix factorization (GMF) was fused with the complex time-series nonlinear high-order interaction, and the final prediction score was obtained through the full connection layer.HMF effectively utilized the user’s multi-feature and user-project interaction information to realize personalized dynamic and accurate recommendation.In order to verify the validity and feasibility of the model, the test was conducted on the public dataset MovieLens-1M.The simulation experiment shows that when the potential eigenvector factors of HMF is 50 and the MLP layer is 7, HR@10 and NDCG@10 are 0.983 1 and 0.974 9 respectively, which are 27.61% and 54.29% higher than the optimal solution of the traditional single feature model NCF.

Keywords

• personalized recommendation;multi-features fusion;deep learning