Research On Social Recommendation Algorithm Based On Memory Enhancement And Project Popularit

Recommendation system is a strategy designed to provide users with better decisionmaking and assist them in making decisions.In recent years,with the emergence and continuous development of deep learning,the development and research of recommendation systems have also increased.A sequence based recommendation system is designed to predict the probability of users clicking on the next item or rating something to help users make decisions.However,extracting user features and accurately embedding project features has been a hot research task in recommendation systems in recent years,and it is also a challenge faced by recommendation tasks.The cold start issue is also a challenge faced by recommendation systems,resulting in the recommendation system being unable to make effective recommendations.Therefore,solving the cold start problem has become particularly important.Therefore,in order to optimize embedded representation and address the challenges of cold start,this article has done the following work:(1)Memory Enhancement Social Recommendation Based on Graph Neural Networks.Previous studies have found that when extracting user features based on deep learning methods,they often only extracted one type of interest feature,or when extracting user features,they often overlooked the sequential order of user interaction when interacting with projects,which is the time information when users click to access the project.Therefore,in addition to capturing general interests in general recommendation models,this article also believes that there are some important factors that need to be modeled: user short-term interests and user long-term interests.The shortterm interests of users can reflect their recent interest preferences well from their recent interaction behavior.The long-term interests and benefits of users can reflect the longterm dependency relationship between the projects they interact with in the early and future.Therefore,in order to address these issues,this article proposes a memory enhanced social recommendation method based on graph neural networks,namely the MEGSR method.This method utilizes graph neural network aggregation to represent users’ short-term interests.The attention mechanism and memory network are used to enhance users’ long-term memory information,so as to add time information of users’ long-term interaction and further optimize users’ embedding.In addition,this article also analyzed the connections between projects,and found that there are co-occurrence relationships between continuously interacting projects.By modeling projects,this article captures the co-occurrence relationships between projects and improves the accuracy of recommendation results.(2)Graph neural network society recommendation based on project popularity.In addition to embedding user information,embedding projects is also a challenge.Social networks can take into account users’ real-life friend information,and this article found that different projects or items have different popularity(the number of times users interact over a period of time).Therefore,in response to these issues,this article proposes a social recommendation graph neural network model that integrates project popularity,namely the GILSR method.For social relationships,by constructing a user user relationship graph,incorporating the influence of user friends on target users,and aggregating friend information through graph attention networks.In addition,for projects,similar projects have similar characteristics and popularity.This article constructs a project project relationship diagram to capture the connections between similar projects,and similarly captures information about similar projects through aggregation.Finally,the long-term and short-term interactive representations and social representations between friends are combined to predict through multi-layer perceptron.(3)Social recommendations for preheating products during cold start.Embedding technology requires data,so it is inevitable to encounter cold start issues.Especially for the issue of cold start for new projects,it is difficult to train suitable and reasonable project embedding for cold start items with limited interaction,which is a challenge faced by embedding technology.In response to this issue,this article proposes a recommendation system model method for product preheating during cold start,namely the CIPSR method.This method is based on GILSR and mainly solves the problem of fast adaptation(accelerating model fitting)between cold projects and GILSR models,as well as reducing the impact of noise on cold projects.This method mainly uses common average initialization embedding to replace random initialization embedding,and in the preheating stage,defines offset and scaling networks to preheat new projects and generate relatively stable embedding.