Research On Session-based Recommendation Integrated With Attention And GRU

With the development of the times and the progress of society,more and more people are experiencing the benefits of technology,but with it comes information overload.Internet users have difficulty finding the information they need in a complex data environment.At the same time,it is important for merchants to filter the data and recommend it to users appropriately when they are faced with huge amount of data.Recommendation systems still need to be studied due to multiple external factors.Recommender systems can help users select the items they need from a large number of products by predicting their preferences through user logs,product reviews and ratings,or using their personal information,and help merchants accurately determine user preferences and recommend different products for different users.Traditional recommendation algorithms make recommendations based on user information.However,due to the cold start,the accuracy is often not high under realistic conditions.Cold start refers to the problem that traditional recommendation algorithms have difficulty in making effective recommendations to users under conditions where user information is unknown or poorly known,such as when users are not logged in or browsing items anonymously.To solve the cold start problem,scholars have proposed Session-based Recommendation algorithms.In recent years,due to the wide application of deep learning,Recurrent Neural Network(RNN)has been applied to Session-based Recommendation,and good results have been achieved.RNN uses the hidden unit of each layer to represent user information,and the current state of the hidden unit is determined by the hidden state of the previous moment and the initial input of the user,which enables RNN to accurately This enables RNN to accurately predict users’ current preferences and build session models.However,few previous research works have focused on the influence of user neighborhood sessions on users’ current behavior patterns,and some research works have not fully captured users’ short-term preferences.To address the above issues,the contributions of this thesis are as follows.In the research work of the thesis,a Session-based Recommendation algorithm that fuses Bi GRU and memory networks is proposed.The Graph Convolutional Network is fused on the basis of RNN to combine the advantages of both,and the neighborhood session and historical session features of users are fused and modeled.The user’s historical behavioral features are extracted by fusing scaled dot product self-attention and Bidirectional Gated Recurrent Unit(Bi GRU).User neighborhood information is obtained using a Graph Convolutional Network and attention across sessions,which is used to improve the Memory Network and determine the contribution of auxiliary information of the current session to extract user preferences.Finally,feature fusion is performed on the session representations of both encoders to obtain the user’s current preferences and predict the next interaction.In response to the fact that most of the previous research works only consider the influence of users’ conscious behaviors on users’ preferences,the thesis proposes another collaborative Session-based Recommendation model.The session information is enriched by fusing the Dwell time and item view of users’ conscious and unconscious behaviors,adding users’ session clicks as the input of the model,and using the co-attentive module to introduce users’ neighborhood sessions to update and enhance users’ global preferences and local preferences,and finally using the fusion selective pass gate to feature select and feature fuse the two preferences and propose a new session recommendation algorithm to obtain the final user preferences and perform personalized recommendations.In this thesis,we use Recall@20 and MRR@20 as evaluation metrics and conduct experiments on two publicly available datasets to verify the effectiveness of this study,while comparing the performance with the more commonly used baseline algorithms and using the control variables method to verify the performance of this model under different situations.The experimental results demonstrate that the algorithm achieves good performance improvement and good robustness under both metrics.