Research On Knowledge-correlated Cross-domain Recommendation Algorithm

As an effective information-filtering tool,Recommender System provides userfriendly digital content services in a personalized way.It plays an important role in addressing the issue of "information overload" in today’s era of big data.However,the performance of recommendation systems in a single domain is deteriorated by problems such as cold-start,sparse data.Cross-domain recommendation provides a solution to the above problems.It achieves more accurate user behavior modeling by utilizing the richer information from auxiliary domains.It aims to guide the recommendation task in the target domain with sparse data,to enhance the comprehensive service performance.However,due to the problems of data heterogeneity and feature mismatch between different domains,the straightforward transfer of knowledge may lead to negative impacts.To address the above challenges,this thesis launches research on the knowledge-correlated cross-domain recommendation system.It mainly focuses on the transfer of cross-domain knowledge,the capture of cross-domain collaborative information,and the valid usage of multisource information.The main work is as follows:(1)Analysis of the cross-domain knowledge linking through Knowledge Graph(KG).Due to the typical problems of data structure inconsistency and knowledge heterogeneity across domains,this work employs KG as common knowledge to establish the linking of cross-domain.Further,this work conducts a complicated analysis of the information provided by KG.Finally,aiming at fully exploiting the available information of KG,this work proposes to achieve the correlation of interdomain knowledge from the perspectives of content semantic and topological connectivity.(2)Domain Adaptive-based Semantic Feature Extraction Model(DASFE).From the perspective of content semantic,DASFE utilizes Knowledge Graph covering massive information from multiple domains and adopts Domain Adaptive(DA)learning technology.In this way,the model automatically extracts transferable semantic features across domains and explores the latent semantic correlation between domains.Specifically,the model explores recommendation-related semantic information in the knowledge graph by the contextual representation-learning algorithm.Furthermore,DASFE aligns the distribution of cross-domain semantic features based on the idea of adversarial training,thus realizing the automatic organization of semantic concepts and the effective extraction of transferable semantic features across domains.Finally,accomplish the correlation of cross-domain knowledge.(3)Cross-Domain Knowledge Graph Attention Network(CD-KGAT).From the perspective of the topological connectivity of inter-domain knowledge,CD-KGAT explicitly models cross-domain high-order information in an end-to-end fashion.This work constructs a Cross-domain Collaborative Knowledge Graph(CCKG),which connects cross-domain user-item interactions using the information in the knowledge graph to achieve cross-domain data correlation.It then adopts Graph Attention Network in CCKG to refine the embedding of graph nodes(users,items,and entities in the knowledge graph).Further,this works designs three feature combination methods.The role of these methods is to incorporate the semantic content information obtained based on DA and the topological connectivity information from CCKG,aiming to generate more complete item profiles,and finally predict the matching score between user and target domain items.By conducting experiments on real datasets with the simulation of a cold-start scenario,this thesis comprehensively evaluates the recommendation performance of the proposed algorithm from the aspects of accuracy and diversity.The experimental results show that compared with other single-domain or cross-domain recommendation algorithms,the proposed algorithm has enhanced 11.41-29.76% and 0.41-2.11% on MRR and the diversity metric,respectively.It is proved that this knowledge-correlated cross-domain recommendation algorithm can not only accurately predict the user’s expressed preferences,but also explore their potential diversified interests.