Research On Session Sequence Recommendation Method Based On Deep Learning

IIIWith the increasing amount of online information,recommendation system has developed into a basic tool to help users quickly select high-value information,and plays an important role in real business.The session based recommendation system only provides accurate prediction based on the item data that users click during the short session.Most of the traditional recommendation methods can only predict the items that the user clicked last time in the session,but ignore the items that the user clicked during the session.Recently,the recurrent neural network model has been proposed for session based recommendation tasks,which can be predicted based on the items that users click during the session.Although the current session based recommendation algorithm has greatly improved compared with the traditional recommendation method in representing the item information that users click during the session,the existing research work still faces the following problems: 1)The session based recommendation mainly forecasts the item information that users click during the session,and cannot represent the association between items that users click during the session.2)User preferences may change dynamically at different times,so it is impossible to determine the impact of items clicked by users during the session on the recommendation of the next item.In order to solve the above problems,this paper proposes a recommendation model based on bi-directional GRU neural network and attention mechanism--Bi GRUAA-Rec(Recommendation Model Based on bi-directional GRU And Attention).In Bi GRUAA-Rec,we first obtain the features of items that users click during the session through embedding layer,then use bi-directional GRU neural network model to represent the relationship between items that users click during the session,and apply attention mechanism to further capture the impact of items that users click during the session on the items that users are interested in,and finally the bi-linear similarity calculation scheme is used in the output layer of the recommended model to effectively reduce the training of the model Parameters and significantly improve the performance of the recommended model.In this paper,the proposed recommendation model is compared with other recommendation algorithms by using the YOOCHOOSE data set.The experimental results show that the proposed model achieves 66.69% and 31.13% in Recall@20 and MRR@20,respectively,which is 3.84% and 3.64% higher than the "Improved GRU-Rec" recommendation model.Experimental analysis shows that the proposed model for long session data sets achieves 65.17% and 27.38% on Recall@20 and MRR@20,respectively,which is 4.05% and 3.09% higher than the "Improved GRU-Rec" recommendation model.The advantages of the proposed model are illustrated by the experimental results and analysis.