Research On Recommendation Algorithm Based On Dynamic Knowledge Graph

Using algorithms and data analysis technology,recommendation system provides services for users based on their behavior logs and interest,which has been widely used.At present,taking the advantage of knowledge graph to improve the performance of recommendation has become a popular trend.Among them,GNN is widely used in recommendation algorithms to process graphs and data better.However,there are still some problems in recommendation models based on GNN.First of all,the existing models fail to transmit and integrate user preference and neighborhood information of projects in a more granularly way.In addition,they also ignores the potential connections between users and non-interactive projects,making it difficult to obtain more accurate information of user preference and feature vector representation of projects.As a result,they will lead to poor results of model recommendation.Secondly,in the dynamic recommendation scenario,the existing methods simply embed the new knowledge into the knowledge graphs,or fully update the total values of the nodal variables of the knowledge graph which has been embedded the new knowledge.Obviously,they will also lead to poor results together with long training time.In order to solve these problems,a new knowledge graph recommendation model and a dynamic update model are proposed in this thesis.The main work of this thesis is as follows:(1)Existing methods have the problem of not modeling user preference and projects in a more granularly way and ignoring the potential relationship between users and non-interacting projects.Therefore,this thesis proposes an attention recommendation model(PNGAT)that integrates user preference and projects’ neighborhood information to solve these problems.Firstly,the model uses the attention network to independently transmit and integrate user preference,to obtain the vector representation of user preference.Then,by calculating the correlation between non-interactive projects and user preferences,the target projects is found as the potential interest objects of users.Meanwhile,utilize the attention network that can independently transmit and integrate the neighborhood information of the target projects,and the vector representation of the target projects is obtained.Finally,the feature vectors of user preferences and target projects were obtained by multi-hop and integration,and input them to the DNC to get the predicted score.(2)Existing methods have the problem that the efficiency of dynamic update is not good because the new knowledge is simply embedded and totally updated.Therefore,this thesis proposes a dynamic knowledge graph recommendation model(ILDGR)based on incremental learning to solve this problem.Firstly,the model samples and plays back the local old knowledge of the knowledge graph according to the importance principle to solve the problem of forgetting in the process of incremental updating.Then,initialize the embedded vectors of new knowledge nodes and local old knowledge,and then update the embedded vectors of local nodes by means of central node transmition.Finally,input the updated embedded vector into the PNGAT model to obtain the predicted score.(3)In this thesis,the two models are experimentally analyzed using the public dataset,and the results prove the effectiveness of the proposed modles.