Research On Network Representation Learning Based Recommendation Algorithm

With the rapid popularization and development of commercial platforms such as commodities,news,and social networking,the data on these platforms are showing an explosive growth trend.Internet users face a serious problem of information overload.In recent years,recommender systems,which can cope with the problem of information overload,have been widely used in commercial platforms,such as online shopping websites and social media apps.Recommender systems can help users to get the content they are interested in from massive items,which improve the user experience.At present,the research on recommender system focuses on how to improve the performance of recommender system by using a large amount of auxiliary data in commercial platforms.Aiming at the problems of inaccurate characterization of user and item features in existing recommendation systems,this thesis carried out a research on the recommendation algorithm based on network representation learning.The main work is as follows.1.In recommender systems,the rating matrix is usually not global low-rank but local low-rank.How to construct low-rank sub-matrices for matrix factorization is an important problem.In this thesis,we propose a novel local matrix factorization model based on network embedding,which can construct more meaningful sub-matrices.To alleviate the sparsity of the rating matrix,the social data and the rating data are integrated into a heterogeneous information network,which contains multiple types of objects and relations.Through the network embedding algorithm,the node representations of users and items are extracted from the heterogeneous information network.According to the correlation of the node representations,the rating matrix is divided into different sub-matrices.Finally,the matrix factorization is performed on the sub-matrices for rating prediction.Experimental results from the real-world dataset Yelp and Douban demonstrate that the proposed model can achieve better performance than the comparative method.2.The growing heterogeneous data in the Internet effectively improves the performance of recommender systems(RS).The main problem faced by the traditional matrix factorization(MF)is how to fuse more heterogeneous information data in MF to improve the performance of RS.In view of this,this thesis proposes an Extended Matrix Factorization(EMF)based on network representation learning.EMF integrates multiple types of data in Heterogeneous Information Network(HIN)to improve the accuracy of scoring predictions.Auxiliary data is useful for recommender systems to characterize users and items more accurately and alleviate the sparsity problem of the rating matrix.However,the quality of auxiliary data,especially social relations,has received much attention.In social networks,some social relations cannot be directly applied to recommender systems.We propose a social filtering algorithm to filter invalid social relations.Then a heterogeneous information network is constructed by social relations,user-item interaction data and item category data.The nodes in HIN are first mapped to a low-dimensional representation vector by network representation learning(NRL).Then the representation vector is used as the input of the EMF,the parameters are optimized by the gradient descent,and finally the prediction model is obtained.The experiments on two real data sets show the effectiveness of the EMF.Compared with the baseline algorithms,the EMF model can obtain more accurate prediction rating.3.Existing recommendation algorithms based on heterogeneous information networks cannot effectively use auxiliary data to describe users and items.In order to improve the performance of recommender systems,user social information and item attribute information should be integrated when building the prediction model,which is a hotspot and difficulty in the field of recommender systems.In this thesis,we propose an extended matrix factorization model based on network representation learning.To characterize users and items comprehensively,we construct the user relation network and the item relation network from the multi-source data.Then the representation vectors of users and items are learned from two networks respectively.Since users and items belong to different vector spaces,a matrix is used to connect user and item representation vectors when predicting ratings.To obtain the connection matrix,stochastic gradient descent is applied to minimize the errors between the predicted and observed ratings.Experimental results on two realworld datasets,demonstrate the effectiveness of our model compared to the state-of-the-art recommendation algorithms.4.Aiming at the problem that existing algorithms cannot accurately use side information in the information network to predict user behavior.We propose a prediction model based on representation learning and label propagation.First,the social relations between users are constructed by the user's mention data and the user's followers list.User preference features are extracted from the user's mention data.The strength of the relations in the social network is determined by the similarity of user preference features,and the edges in the user social network are weighted by the strength of the relations.In the process of network representation learning,the transition probability of random walk is first determined by relation strength,and a more reasonable user node sequence is sampled.The user node sequence is taken as the input of the skip-gram to train the user representation vector.Based on the user's representation vector,the label propagation probability among users in the social network is calculated.Finally,the user's labels are predicted by the label propagation algorithm.