Research On Recommendation Algorithms Based On Graph Neural Networks

In the current Internet era,from the traditional news,films and television sites to the recent mobile phone short video platforms,they generate large amounts of information and data.In the face of millions of data,users are unable to get the interesting contents in the expected time,so this inevitably result in information overload.By analyzing the historical interactive data of users and items on the platform and the personal profile of users,recommender systems can quickly discover the preference laws of users to push personalized contents to them.Therefore,the recommender systems play an important role in alleviating the problem of information overload.Recently,Deep Neural Networks(DNN)have made revolutionary progress and been widely used in many research fields.Due to its powerful nonlinear feature capture ability,DNN can capture the personalized preference characteristics of users and the category characteristics of items accurately,so it also brings a new opportunity for the development of the recommender systems.Learning the representation vectors of users and items is the core of recommender systems,In the traditional recommendation algorithms,the potential relationship between users and items are often ignored when modeling the representation vectors of users and items.There is a natural complex graph structure relationship between users and items in Internet platform.Each user and item in the graph is abstracted as a node,so high order connectivity between nodes can be extracted by Graph Neural Networks(GNN).The neighborhood information of any order can be flexibly integrated into the representation vectors of nodes.This paper mainly improves the traditional GNN and incorporates other models to improve the recommendation accuracy.The main work of this paper includes following three parts:1.Traditional GNN ignores the information loss when users and items nodes extract neighborhood information.In view of this problem,this paper proposes a Convolutional Memory Unit-based Graph Collaborative Filtering for Recommender Systems(CMGCF).In the process of generating representation vectors of users and items,the information transmission between target node and its neighbors will inevitably cause the problem of signal attenuation.GNN recursive process can be regarded as the sequence data process,which is similar to the Recurrent Neural Networks(RNN),so we introduce a Gated Recurrent Unit(GRU)to solve the long-term dependence problem in sequence data process.In this way,the target node can obtain more accurate characteristic information from its neighbors.Experiments on 4 public datasets show that the performance of our algorithm are significantly improved compared with the baseline algorithms.2.Traditional Matrix Factorization(MF)algorithm is easy to use and expand,so it is widely used in recommendation field.In the contemporary big data environment,the traditional MF algorithm cannot solve the problem of data sparsity in an effective way.Aiming at this problem,we propose an Attentional Factorization Machines Based on Graph Neural Networks for Recommender Systems(GAFM).We use Attentional Factorization Machines(AFM)to perform second-order interactions on different neighborhood information and the attention mechanisms can perform weighted redistribution of second-order interactive features.Experiments on 4 public datasets show that the performance of our algorithm are significantly improved compared with the baseline algorithms.3.Factorization Machines(FM)is an effective way to solve the problem of data sparsity,it can only model the second order interactive features between neighborhood information of nodes.However,it cannot capture the high-order nonlinear relationship of node neighborhood and ignore the correlation between the dimension of users and items representation vectors.To solve this problem,this paper proposes a Convolutional Factorization Machines Based on Graph Neural Networks for Recommender Systems(GCFM).Firstly,we use GNN to extract the features of users and items in different neighborhoods,then use tensor product equation to fuse the features between dimensions to generate a series of feature maps.These feature maps contain the relationship information between feature vector dimensions.Then the feature maps are stacked into a feature cube,which can retain the sequential information between features.Finally,we use 3D Convolutional Neural Networks(CNN)to learn the high order neighborhood information interaction between users and items.At the same time,the model can also capture the dynamic information between feature maps.Experiments on 4 public datasets show that the performance of our algorithm are significantly improved compared with the baseline algorithms.