Research On Personalized Recommendation Method With Memory Networks

With the rapid development of network and information technology,people are already in an era of information explosion.On the one hand,people have discovered many advantages that come with the advancement of science and technology.On the other hand,information overload has become a major challenge in the Internet era.The rapid increase of information resources makes it very difficult for users to find what they need in the complex network resources.Recommendation system can effectively alleviate the above-mentioned problem of "information overload",provide personalized services for users and improve the user experience,so it has important research value.In recent years,more and more researchers have applied deep learning techniques to deal with collaborative filtering recommendation based on implicit feedback.However,traditional deep learning models usually use hidden states or weights with limited memory abilities as memory functions,which cannot store long-term information and is difficult to perform memory operations.For example,embedding all of a user's history interactions into a single hidden state vector would fail to capture the correlation between an individual user's history interaction and user preferences.Memory network can store long-term information and perform memory operations,so we consider designing memory networks to act as memory functions.In addition,traditional deep learning models generally profile both users and items directly,while neglecting the importance of collaborative information between users' and items' neighborhoods.The users' neighborhoods refer to the user sets composed of all users who have visited the same item.Users in the same user's neighborhood are neighbors to each other.Similarly,the items' neighborhoods refer to the item sets composed of all items visited by the same user,and the items in the same item's neighborhood are neighbors to each other.We guess that the collaborative information contained in the neighborhoods can improve the performance of the recommendation to a certain extent.For these reasons,this paper has launched a research on personalized recommendation method with memory networks.The accuracy of recommendation is improved by designing memory networks as the memory functions and considering the collaborative information between neighborhoods.The main contributions of this paper are as follows:(1)This paper proposes a Collaborative Memory Network based on Items' Neighborhoods(CMN-IN).This model firstly designs a memory network and combines the attention mechanism to construct the item neighborhood component to store the long-term information of the item.Then use the item neighborhood memory network to capture the relations between the target item and its neighbors,and through the iterative processing of multi-layer item neighborhood components,to achieve long-term memory of the higher order item neighborhood information.Finally,the user and item memories and the item neighborhood information are combined to encode the complex user-item relations,and the prediction score is obtained.On three real datasets,we compare the model with the best collaborative filtering methods and the deep learning-based models,and verify the superiority of the model in the evaluation indexes such as HR@5,HR@10,NDCG@5 and NDCG@10.(2)Based on the model proposed in(1),this paper further proposes a Neighborhood Attentional Collaborative Memory Networks(NAMN).In order to learn the potential similarity between users and their neighbors in the user neighborhood,the model also designs the user neighborhood component to store long-term user information.Then using the associative addressing scheme to encode the relations between the target user and its neighbors,and then stack the multi-layer user neighborhood memory network to get a deeper structure,and further mine higher order user neighborhood information.Finally,the user and item memories with the user and item neighborhood information are combined in the output module to obtain the ranking score.The experimental results show that compared with the state-of-the-art benchmark methods and the model proposed in(1),this model has a certain improvement,which verifies the effectiveness of this model.In addition,we also analyze the parameter sensitivity of the model,and explore the influence of individual neighborhood memory network component on the model through ablation experiments.