Research On Recommendation Algorithms Based On Deep Learning In Social Media

As an effective key information filtering technology,recommender systems have always been a popular research field in social media.By analyzing user's historical behaviors for predicting user's preferences,recommender systems guide people in a personalized way of discovering items they might be interested.With the explosive growth of data,current models have their own limitations in dealing with data sparsity and cold-start problems.In recent years,with the rapid development of social media,recommendation resources containing user behaviors and demands information have become more abundant.More researchers focus on how to integrate these multisource auxiliary information into traditional collaborative filtering models to improve the quality of recommendation.In addition,the revolutionary advances of deep learning bring new opportunities to the research of recommender systems.How to use deep learning to extract effective features for users and items and combine them with recommender systems has been a hot topic of current research.In this paper,we focus on mining auxiliary information in social media(such as social networks and reviews)by deep learning technology to alleviate data sparsity.The main work and contributions in this paper are summarized as follows:First,a convolutional matrix factorization algorithm based on user social network and review content is proposed for rating prediction.Given that social network and review content in social media can help users filter information and express their interests,using them to learn user and item feature representations is helpful to alleviate the sparseness of scoring data.Thus,we build a united framework that models the review text of both users and items and user social network simultaneously.Specifically,we use CNN to extract contextual information in review text as user preferences and item features,and then we jointly decompose the social network matrix to model the influence of social network.Finally,the above components are incorporated into the collaborative filtering framework for common training.In this way,the model takes advantage of both collaborative filtering and neural networks to enhance the accuracy of recommendations.Second,a graph convolutional matrix factorization algorithm based on social propagation is proposed in this paper.Given that the above algorithm considered the local neighbors of each user and ignored the process that users' preferences are influenced as information propagation in the social network.Meanwhile,graph convolutional networks show great advantages in modeling graph structure data.Based on the above observations,we use GCN to imitate the propagation in social network.At first,we use the review content as node attributes for each user,and obtain user's social embedding by GCN.Next,we integrate the social embedding into MF to model the influence of propagation process on user preference.Finally,we perform experiments on realworld datasets.The experimental results show that our proposed model achieves better performance than other related methods in this paper.