Research On Recommendation Algorithms Based On Network Representation Learning

The booming development of Internet technology has provided a variety of services for human life,but the problem of information overload has also followed.Personalized recommendation system can filter information from a large amount of data and select the information that meets the needs of current users,so it is regarded as an effective solution to the problem of information overload.In recent years,the rapid development of network representation learning technology has opened up new ideas for the research of recommendation algorithms.Aiming at the recommendation problem in the two scenarios of static homogeneous network and dynamic binary network,this paper proposes corresponding algorithm schemes by using network representation learning technology respectively.The main work is as follows.To solve the recommendation problem of static homogeneous networks,this paper proposes Graph Embedding based on Random Walk Associated Text(Gerate).Specifically,in the stage of random walk,the algorithm taking into account the similarity between nodes’ text,combines structure and text information to filter the next node.Then the context information is fused in the network representation learning part and the attention matrix is introduced to represent the vector in the text information matrix in a weighted way.Finally,the generated node vector is applied to the recommendation task,and the comparison experiments are carried out on multiple data sets respectively,which proves that the proposed algorithm can effectively improve the recommendation accuracy.In order to solve the recommendation problem of the dynamic binary network,the algorithm named Dynamic Bipartite Network Embedding based on Long Short-term Memory(Bi-DYNE)is proposed in this paper,considering that the interests of users in the binary network usually change over time.Specifically,the algorithm first decomposes the binary network into the user homogeneous network and the item homogeneous network at different points in time,and uses the Gerate algorithm to get the preliminary feature matrix which integrates the text attribute information.Then the large-scale LSTM model is used to learn the network information in the long span time neighborhood.Finally,the fusion feature matrix of the user and the item is input into the MLP model for training,and the interaction probability of the user and the item is predicted to achieve the recommendation task.The effectiveness of the proposed algorithm is verified by comparative experiments on multiple data sets.