Research Of Recommendation Algorithm Combined With Network Embedding

In big data era,personalized recommendation systems have become an indispensable part of people's lives as an effective means to combat information overload.Network data widely exists in various recommended scenarios,such as bipartite networks for describing the interactions between users and items,as well as social networks composed of users,geographic information networks composed of geographic locations and so on.Network data conveys users' diverse information needs from different sides.This paper focuses on social recommendation scenario and route recommendation scenario.According to the characteristics of the two application scenarios,two recommendation methods combined with network embedding are proposed.In terms of social recommendation,how to effectively integrate social network embedding with traditional matrix factorization is a hot issue in this field.However,existing methods only use explicit or implicit social networks of users,and there are few research paradigms that make use of both networks at the same time.Meanwhile,most methods usually adopt on-by-one learning framework,which means the network embedding model is independent of the matrix factorization model.As a result,the models cannot promote each other.In view of this,this paper proposes a social recommendation method based on joint learning of dual network embedding and matrix factorization.This method jointly models the explicit and implicit social relationships of users to obtain more accurate social embedding;and solves the joint objective function of the dual network embedding model and matrix factorization model under the unified learning framework to promote the two-way interaction and collaborative optimization between models,making the users' latent representations not only contain the social characteristics helpful for the recommendation task but also retain the behavior characteristics under social influence,so that the algorithm can capture the users' interests more accurately.A large number of experimental results based on three public datasets show that the recommendation performance of proposed method is greatly improved compared to existing social recommendation algorithms.In terms of route recommendation,trajectory embedding learning based on Recurrent Neural Network(RNN)is the mainstream method in this field.It aims to combine neural network embedding technology to mine spatial relationships in geographic information networks and learn users' movement behavior preferences from historical trajectories.However,the existing methods do not consider the topological constraints in geographic scene,resulting in inefficient learning of the model;at the same time,they also ignore the difference in the importance between key locations and ordinary road sections,so that the recommendation performance of existing methods needs to be improved.In view of this,this paper proposes a neural network embedding algorithm based on attention mechanism for topological constraint sequence modeling in route recommendation task.The model first maps the user's trajectory to a sequence of vertices,and then learns the embedding representations of vertices from RNN;meanwhile,it incorporates geographic topology constraints into the representation learning process to reduce unnecessary computational overhead;in addition,we also introduce attention mechanism to emphasize the key geographic location in the sequence,so as to reduce the interference of the noise in the sequence to the model.Experimental and visual results based on large-scale trajectories data of Meituan Takeout verify the superiority of this method compared with similar methods.