Research On Cross Domain Recommendation Based On Knowledge Graph

Information overload is a critical challenge in big data environments,and information recommendation is increasingly recognized as an effective approach to mitigate this problem.Traditional information recommendation methods rely on separate or limited data domains as data sources,which often suffer from issues such as data sparsity and cold start.Furthermore,integrating rich multi-domain knowledge is a challenging task.The research on cross domain recommendation has added knowledge dimensions to address the above issues.Cross domain recommendation can enrich data in target domains by obtaining effective user preferences or item characteristics from other domains,effectively improving recommendation performance.However,research on cross domain recommendation is still in its infancy and has certain limitations,mainly including: the lack of universality and portability of existing methods,most of which only focus on knowledge transfer between two or a few domains;With the overlap of data domains,constructing recommendation models becomes increasingly complex and challenging;The knowledge association between multidimensional data has not been fully mined.In recent years,the rapid development of knowledge graph technology has opened up new research avenues for cross domain recommendation.Knowledge atlas has the advantages of knowledge structure,semantic relevance,intelligent reasoning,and interpretability.Implementing cross domain recommendation based on knowledge graph technology is an important development direction of recommendation research in the future.This thesis proposes a cross domain recommendation model focuses on knowledge fusion.Specifically,this research constructs a cross domain knowledge graph using the Data Central Platform idea and integrates and expands multi-source user interest knowledge,then reconstructs the fusion knowledge graph of interest knowledge,application knowledge,extended knowledge,and social knowledge using path search and transformation methods.It achieves cross domain recommendations by leveraging the advantages of knowledge graph and attention network.This research explores the following areas:(1)Building of cross domain knowledge graph based on the idea of Data Central Platform.This data system of cross domain knowledge graph consists of 7 layers from bottom to top,including the source data layer,domain data layer,graph schema layer,graph data layer,graph tag layer,graph calculation layer,and graph application layer.It includes a series of steps such as cross domain data processing,cross domain knowledge graph design,graph data processing and services,knowledge graph calculation,and cross domain application.In designing the knowledge graph,various types of knowledge are classified into Knowledge Space of User,Knowledge Space of Item,Knowledge Space of Application,Knowledge Space of Semantic,Knowledge Space of Encyclopedic,Knowledge Space of Field and Knowledge Space of Social Network.A lightweight unified process ontology construction method is employed to establish a cross domain knowledge graph ontology model.Relying on the Data Central Platform to process cross domain knowledge graph data,this knowledge graph establishes a basic association between multi-domain users and projects through the core ontology,and leverages the extension ontology to provide more complex extension recommendation support.The experimental results demonstrate that cross domain knowledge graph can effectively integrate knowledge from various data domains,with a higher user coverage rate than single domain knowledge graph.(2)Cross domain interest fusion model based on user tag subgraphs.To tackle the challenges of interest sparsity,data authority,content quality,and architectural differentiation of user interest knowledge across multiple domains in the knowledge graph,this model establishes a user tag subgraph that utilizes user nodes,interest tag nodes,and their associated relationships within the cross domain knowledge graph.This research labels user interests across multiple domains using user tags with a uniform specification and rich semantic information as carriers to store and represent user interests.To obtain the user entity-tag matrix of multiple domains,this model utilizes cross domain user identification and tag weight normalization methods.Finally,it fuses the tags using domain weight influence coefficients to construct a user fusion interest set with compound weights.The experimental results demonstrate that the cross domain interest fusion model can effectively enhance the tag user coverage effect while maintaining a high tag user accuracy rate across fused domains.(3)User interest extension model based on knowledge graph path association.The issue of data sparsity and long tail in the big data environment remains a challenge,and even after cross domain interest fusion,sparse interests may persist.To overcome this challenge,this research proposes an extended user interest modeling approach that incorporates extended knowledge acquisition and fusion,as well as the integration of semantic and social network associations among interest nodes.This model builds an extended interest subgraph that leverages the associations between interest tag nodes.The approach involves mining the relationships between interest tag nodes and superordinate word nodes,encyclopedia tag nodes,and social network user nodes.Then,it calculates the semantic association degree of interest tags and social network association degree to generate compound association weights between interest tag nodes.By reconstructing the derivative relationships between interests,this model achieves user interest expansion.The experiments verification demonstrate that this model can effectively fuse and expand different types of interest association knowledge,leading to improved coverage and accuracy of user interests compared to single domain data.(4)Cross domain recommendation model fusion knowledge graph and attention network.The knowledge information available in the big data environment is complex and diverse.Data extracted from the cross domain knowledge graph typically has a large number of high-order connections and noise interference,making it challenging and complex to use directly for recommendation calculations.In order to tackle this issue,path transformation is employed to integrate application knowledge,extended knowledge,and social knowledge.The resulting fused knowledge of diverse types is expressed and stored through the use of core knowledge subgraph.By utilizing flexible screening and combination techniques for multiple auxiliary domain knowledge,virtual graphs and embedding representations can be generated.This knowledge graph are then leveraged to facilitate neighborhood information propagation,higher-order relationship aggregation,and cross domain recommendation calculations based on Attention Network.To further enhance cross domain recommendation performance,domain knowledge evaluation and preference are taken into account.Experimental results demonstrate demonstrated that the utilization of fused knowledge graph in the proposed model improves the recall and accuracy of cross domain recommendation.The model effectively enhances cross domain recommendation performance by leveraging the advantages of knowledge fusion and graph attention networks in a complementary manner.