Research On Recommendation Algorithms Based On Feature Learning In Social Networks

As an effective information filtering technology in social networks,the recommendation system has received extensive attention from researchers in recent years,which analyzes the behavior data of users to provide them with personalized information and services.On the one hand,although the existing algorithms can achieve better results,the recommendation system still faces many challenges,such as cold start problems,implicit feedback problems,and the geographical characteristics of items.On the other hand,feature learning as an effective tool for modeling auxiliary information such as social networks and geographic locations is helpful to alleviate data sparsity and improves the accuracy of recommendation.How to combine feature learning with recommendation system is one of the hottest topics in current researches.Accordingly,this work takes the recommendation algorithm in social network as the research object,and focuses on how to further improve the accuracy of the rating prediction by mining social network features;and how to enhance the Point-of-Interest(POI)recommendation performance on the consideration of implicit feedback and neighbor features.The main contents and innovations are as follows:We propose a recommendation method based on embedding features of social networks.In social network,we always turn to our friends for recommendations and there is a correlation between items.Although many studies have taken features of social relationships and item correlations into account to enhance recommender systems,most of them regard these features as regularization terms,and the deep structure features hidden in social networks and rating patterns have not been fully explored.Motivated by the above observations,we utilize the network representation learning technology and propose an embedding feature based recommendation method,which is composed of the network features embedding model and the collaborative filtering model.Specifically,a neural network based embedding model is first pre-trained,where the external user and item features are extracted.Then,we incorporate these extracted factors into a collaborative filtering model,where our method not only can leverage the network features to enhance recommendation,but also can exploit the advantage of collaborative filtering techniques.We propose a neighbor features based POI recommendation method.Different from traditional recommendation problem,users in location-based social network(LBSN)always express their interest only by checking in different POIs,which is a kind of implicit feedback.Moreover,although the geographical factor has been proven to be beneficial for improving POI recommendation accuracy,previous work mainly model them from the user perspective instead of location perspective.Intuitively,neighboring POIs tend to be visited bysimilar users,which implies that modeling neighbor features from a location perspective can simulate users' behavior more reasonably.Based on the above observations,this work concentrates on exploiting the neighbor features from a location perspective for improving recommendation performance.To be specific,the weighted probabilistic matrix factorization(WMF)that can deal with implicit feedback is first introduced as our basic method.Then,we incorporate the neighbor features into the WMF from a location perspective.Finally,we conduct several experiments to evaluate our method on two real-world datasets,and the experimental results indicate that the importance of these external extracted features and the effectiveness of our proposed approaches than other related methods.