Study On Social Recommendation By Joint Learning Of Matrix Factorization And Network Embedding

The recommendation system is currently the main means to alleviate the problem of information overload.It filters out a large amount of redundant and irrelevant data for users and selects valuable information from it.This not only greatly improves the user experience,but also significantly enhances the business benefits of the enterprise.Collaborative filtering is a key technology for building a personalized recommendation system.It focuses on inferring user preferences for items through collective wisdom and experience.The main challenge facing this technology is the sparse data of the user-item interaction.Nowadays,social media is developing rapidly,and a large amount of information related to users' preferences circulates in social networks.Therefore,user social links have become a key supplement to sparse interaction data,and a social recommendation model using social information can achieve more accurate and interpretable recommendations.The traditional social recommendation algorithm only considers the user's first-order friends,and ignores the information dissemination between the user and his high-level neighbors.The popularity of Network Embedding(NE)technology provides new ideas for social recommendation systems,which can ensure that high-order neighbors with similar structures in the network also have similar feature representations in the hidden space.Recently,some social recommendation methods combining NE and Matrix factorization(MF)model have emerged.Most of these methods first pre-train NE models and then feed the information to the downstream MF,which leads to the model's inability to capture network characteristics suitable for the recommended task.Furthermore,they only focus on using social networks to enhance the user's feature representation learning,and ignore the feature representation of items.Aiming at the above problems,this paper proposes two social recommendation schemes of joint learning MF and NE.The first scheme integrates MF and NE models into a unified optimization framework through feature representation alignment.In the process of joint learning,the two tasks enhance each other.The supervision of rating data makes the network features learned by the NE model more suitable for the recommendation task,and MF will learn more discriminative embeddings when fusing multiple types of information.In addition,explicit user-user networks and implicit itemitem networks are used collaboratively to enhance users' and items' representation learning.A large number of comparative experiments on three public datasets show that the method of joint learning through feature representation alignment and dual-network collaborative embedding can greatly improve the recommendation performance.In order to model the interaction between the user and the item feature representation in a deeper level,and completely retain the network structure information and the user preferences and item characteristics in the rating data,this paper proposes the second scheme.While retaining the network characteristics suitable for the recommendation task,this scheme retains the complete user preferences and item characteristics by separately training part of the feature representation using rating data.In addition,we use Multilayer Perceptron to impart nonlinearity to the model to learn the interaction between user and item feature representations.Finally,by comparing the most advanced social recommendation methods,the effectiveness of the joint learning model considering the above two points is fully verified.