Research On Group Recommendation Algorithm Based On Deep Learning

With the rapid development of information technology and computer science,the amount of data in the Internet world is more and more huge.The problem of data overload caused by this has also caused great trouble to Internet users.Recommendation systems could solve this problem effectively,which can help users to get the information they really need in a short time.Most traditional recommendation systems serve independent users,they do not have the ability to recommend items for a group of several users.However,in real life,many services are not provided to individual users.For example,movie websites recommend movies for families,food apps recommend restaurants for company employees,and travel service providers recommend scenic spots for travel groups.Due to the particularity of group recommendation and its wide application,it has gradually become a hot issue for scholars in recent years.Most of the existing group recommendation algorithms adopt a way of setting the weight of group members in the decision-making process in advance to integrate the preferences of different members of the group,so as to form a preference that can represent the whole group,and then use the group preference to carry out the next project recommendation.However,this kind of method does not have the ability to simulate the complex group decision-making process in the real situation.When facing the different projects,the weight for each members of a group should not be a fixed value,but a dynamic calculation result.On the other hand,users with social relationships in the group will also have a certain impact on their own preferences,thus affecting the preferences of the whole group.In addition,most recommendation methods based on deep learning technology are lack of robustness and generalization ability,which has a great impact on the final recommendation accuracy.In view of the above problems and challenges in the process of group recommendation research,this paper studies from the following two aspects:(1)Attention mechanism and social trust based group recommendation.In thispaper,Attention mechanism and social trust based group recommendation(ASBGR)dynamically calculates the preference weights of members in different groups for different items,and integrates social trust into the learning process of group preference representation.The main idea of ASBGR algorithm is: when the group interacts with the project,the attention mechanism is used to dynamically assign weights to the members in the group,and then the members in the group are weighted to sum their preferences,and the social trust relationship between members in the group is integrated into the preference calculation process,then the sum of the final weighted preferences is used to represent the preferences of the whole group calculation of the continued recommendation list.Using this method which combines social trust relationship with dynamic weight assignment can better reflect the different decision-making abilities of the members in the face of different projects.ASBGR are more in line with the process of group decision-making in reality,and make the final group preferences more reflective of the real situation of the group.The experimental results on real datasets show that the ASBGR algorithm improves the recommendation quality significantly compared with other existing group recommendation algorithms.(2)Adversarial training for group recommendation.There are many group recommendation methods based on in-depth learning.Thanks to the tremendous advantages of in-depth learning in feature extraction and computing,these group recommendation algorithms are significantly better than those without in-depth learning.However,recent studies indicate that most deep learning-based algorithms are not robust and have insufficient generalization capabilities,and group recommendation algorithms are no exception.Based on the existing group recommendation algorithm,the ATFGR proposed in this paper uses adversarial training to further optimize the model,to increase the robustness and generalization ability of the algorithm,so as to improve the accuracy of the recommended results.The experimental results on real datasets show that the model with adversarial training improves the recommended quality compared with the original model.