Research On Session Recommendation Model Based On Self-attention And Dilated Convolution

In the rapidly developing information age and digital economy,recommendation systems help users filter the content of interest from the vast amount of data and thus assist in decision making.Session recommendation is an important branch of recommendation systems,which aims to predict users’subsequent behaviors through interactions in a continuous period of time without user information.Through the study of traditional and deep learning session recommendation models,it is found that most models only consider the interaction behavior within or between sessions,but ignore the influence of contextual information on recommendations.At the same time,there are problems of inadequate feature extraction and under-representation for neighborhood information within sessions and between sessions.Accordingly,the related research is carried out,and a time-aware session recommendation model based on self-attention and dilated convolution is proposed,the main work is as follows:(1)Considering the impact of time interval perception on users’ preferences and the problem of inadequate feature extraction within sessions,incorporating dilated convolution in a time interval session recommendation model,and proposed time-aware session recommendation method incorporating dilated convolution.The time interval is used to obtain richer contextual information and user’s personalized behavior,while capturing users’ short-term interaction behaviors at long distances within sessions using dilated convolution to improve the recommendation performance of the model.Experiments on two datasets show that this improved approach can effectively improve the recommendation accuracy of the model.(2)In order to better capture the complex transition relationships between items in different sessions and more representative item features,the stacked self-attention block is further incorporated to propose a dilated convolutional time-aware session recommendation method incorporating self-attention to learn more representative long-term user behavioral preferences.At the same time,since deepening the number of network layers increases the complexity of the model,which in turn leads to model overfitting problems,for this reason,residual connections are made to each self-attention layer and feedforward neural network to improve the generalization ability of the model.The effectiveness and necessity of this improved method are verified by comparison experiments as well as ablation experiments.(3)The improved method proposed in this paper is applied to the news recommendation system,and the corresponding software design and development are carried out.Through actual operation and functional testing,it shows that the application system can run stably and also proves that the recommendation method proposed in this paper has certain practicality.