Researches On Dynamic Knowledge Graph Embedding And Knowledge-aware Recommendation

Recently,Knowledge Graph(KG)has attracted extensive attention because it is able to describe entities and their relations in the objective world in a structured form and express the information of the Internet in a form closer to the cognitive world of human beings so as to help to organize,manage,understand and utilize the massive information of the Internet.Though it is effective to represent inevitably structured data,the underlying symbolic nature usually makes it hard to manipulate KGs and apply KGs to downstream tasks.To deal with this issue,a new research direction known as Knowledge Graph Embedding(KGE)has been proposed and quickly gained massive attention.The main idea of KGE is to embed entities and relations of a KG into low dimensional spaces so as to simplify the manipulation while preserving the inherent structure of the KG,which sets basis for KG's application,such as alleviating data sparsity faced by Recommendation Systems(RS).Based on the existing research of Knowledge Graph,the present essay seeks to contribute to the study of KGE and KG's application on RS.The main work of the present essay includes the following two aspects.Firstly,most existing KGE approaches ignore the historical changes of structural information involved in dynamic knowledge graphs(DKGs),which greatly affects the learning of entities and relations.To deal with this issue,this paper integrate Gated Recurrent Units(GRU)into KGE and presents a Timespan-aware Dynamic knowledge Graph Embedding Evolution(TDG2E)method.Specifically,TDG2 E fragments a DKG into multiple static sub-KGs with each sub-KG corresponding to a specific time bin.Then a GRU-based model is utilized to deal with the dependency among sub-KGs that is inevitably involved in the learning process of the dynamic knowledge graph embedding.Meanwhile,to further deal with the time-span unbalance issue underlying the DKGs,a Timespan Gate is designed in GRU.It makes TDG2 E possible to model the temporal evolving process of DKGs more effectively by incorporating the timespan between adjacent sub-KGs.Extensive experiments on two large temporal datasets(i.e.,YAGO11 k and Wikidata12k)extracted from real-world KGs validate that the proposed TDG2 E significantly outperforms traditional KGE methods in terms of Mean Rank and Hit Rate.Secondly,most existing knowledge-aware recommendation methods largely ignores the importance of different entities when incorporating them from KG into RS,which may lead to suboptimal performance.To deal with this issue,in this paper,we incorporate Attention mechanism into knowledge-aware recommendation and present an Attention-enhanced Knowledgeaware User Preference Model for Recommendation(AKUPM).AKUPM first incorporates a bulk of entities directly or indirectly associated to users by automatically and iteratively extending a user's history clicks along links in the knowledge graph.Then,AKUPM adopts Attention mechanism and Self-attention mechanism to capture the relationship among different entities,on the basis of which we lay different importance on entities so as to highlight entities related to users' preference and then improve the accuracy recommendation.In order to verify the effectiveness of the proposed model,the present essay designs and conducts extensive experiments based on two mainstream datasets(i.e.,Movie Lens-1M and Book-Crossing).The experimental results show that AKUPM is superior to the state-of-the-art methods in terms of AUC,ACC,and Recall@top-K.