Research Of Recommendation Algorithm Integrating Knowledge Graph And Tensor Decomposition

Recently,owing to the growth of the digital economy and the persistent advancement of data mining technology,the integration of Knowledge Graph(KG)into Recommendation Systems(RS)to represent human knowledge in a structured format has captured substantial interest in both scholarly and corporate exploration,which can effectively mitigate data sparse and cold start problems,and also provide some interpretability for recommendation results.However,most of the existing knowledge graph-based algorithms do embedding representations of entities or manually design metapaths,without considering the internal connections of various attributes and the influence of users’ dynamic preferences.Based on this,the study explores the fusion of different auxiliary information and recommendation algorithms around item-side knowledge graphs to better explore the semantic associations behind various auxiliary information.This thesis undertakes two main tasks,which are detailed as follows.Firstly,user attributes,as a good reflection of the user’s inherent interests,can describe the user’s preferences intuitively,but most of the existing knowledge graph-based algorithms do not take this into account well,this causes a dearth of understanding regarding the user’s representations,so in this thesis,we use the user’s historical interaction items as the seed of preference propagation,which is the head entity of the knowledge graph,and combine the entity-to-entity links to spontaneously spread to related entities to deepen the user’s preferences.Therefore,we use users’ historical interaction items as the seed of preference propagation,i.e.,the head entity of the knowledge graph,and combine the links between entities and entities to spontaneously spread to related entities so as to dig deeper into users’ potential interests,then add users’ inherent attribute information to improve users’ representations,fuse inherent and potential preferences,and then perform CTR prediction.The experiments show that the built model can indeed improve the performance of the recommendation algorithm.Secondly,considering the dynamic nature of users’ partiality,we introduce the time characteristics,combine users,items and ratings to build a 3D tensor with temporal characteristics,and use the CP decomposition algorithm with temporal awareness to obtain the hidden semantic vectors with temporal and spatial differences with the same dimensionality after decomposition,and add them to the user representation,finally,combine the inherent characteristics of users and the existing knowledge map of the item side,increase the entity preference.The diffusion level of propagation is increased to mine underlying semantic information.Our model’s superior performance over existing benchmark models has been validated through multiple experiments conducted on datasets of varying sizes,the larger the data volume,the better the model performance.