Research On Collaborative Filtering Recommendation Algorithm Based On Improved Variational Auto-encoder

The traditional recommendation method uses the user item scoring matrix to calculate the similarity and carry out the recommendation task,but the recommendation effect will be poor because it can not make full use of the user item information.Based on this,a variational auto-encoder recommendation algorithm integrating complex a priori and attention mechanism is proposed.The combination of variational auto-encoder model and recommendation method can not only learn the potential characteristics of users and projects,but also compress the high-dimensional sparse user item matrix into a low-dimensional dense matrix.Firstly,the implicit vector prior distribution generated by multilayer neural network is used to replace the standard normal distribution as the hypothetical prior distribution,so that the model can obtain more potential representations;Then,based on the single-layer implicit vector,an auxiliary implicit vector is added to regenerate the auxiliary implicit vector and the data feature vector into an implicit vector,so as to enhance the low dimensional performance and decoupling of the implicit vector;Finally,the attention mechanism is introduced into the model to give greater weight to important nodes,so that the hidden vector can convey more important information.Experiments were conducted on Movielens-1M,Movielens-latest-small,Netflix and Movielens-20 M.The results show that adding complex a priori and attention mechanism modules can improve the accuracy of the model.Adding them to the model together can optimize the accuracy of the model.At the same time,compared with Multi-DAE,Multi-VAE,Macrid-VAE,CDAE and NCF,the proposed model is better in the evaluation index Recall@20,Recall@50,NDCG@100 On average,it increased by 12.95%,10.80%and 10.48%.Experiments show that the proposed method can significantly improve the recommendation effect.This paper has 35 figures,9 tables and 56 references.