Research On Graph-based And Rezero-based Recommendation Models And Training Acceleration

With the continuous development of Internet technology,big data technology and artificial intelligence technology,personalized recommendation systems(RS)have been widely used in various e-commerce platforms and multimedia websites,which not only filter information and optimize experience for users,but also attract users and increase revenue for enterprises.In recent years,deep learning(DL)has brought new progress to RS domain.Thanks to the powerful representation ability and flexible module structure of neural networks,DL-based collaborative filtering(CF)recommendation algorithms and sequential recommendation(SR)algorithms have overcome many obstacles in traditional algorithms and performed better.Currently,DL-based recommendation algorithms still have some limitations in accuracy,scalability,and training efficiency.To alleviate these problems and challenges,we propose some solutions:(1)We propose a deep CF recommendation algorithm Graph CF with a graph neural network,which explicitly models interaction information between users and items,so as to improve the recommendation accuracy of the model;(2)We propose a deep SR algorithm RZRec that integrates Rezero operation into all residual blocks and extends more intermediate hidden layers for the model,so as to improve the expression ability and recommendation accuracy of the model;(3)We find that the hidden layers of deep SR models are similar.Inspired by this,we propose a training acceleration algorithm Stack Rec based on knowledge transfer for deep SR models.Through layer stacking and knowledge transfer,Stack Rec achieves more than2 times training acceleration without any loss of recommendation accuracy.We conduct extensive experiments and ablation analysis to verify that the proposed methods can effectively alleviate the above limitations and improve the recommendation performance.In summary,we propose some solutions to improve the accuracy,scalability and training efficiency of DL-based recommendation models,and verify their effectiveness through experiments on real datasets,which shows great theoretical significance and application value.