Research On Cross-domain Recommendation Algorithm Based On Heterogeneous Network Graph

Cross-domain recommendation aims to integrate user preference features from multiple domains to achieve personalized intelligent perception,in order to more accurately meet users’ personalized needs and thereby enhance the accuracy and diversity of recommendation results in the target domain.In practical scenarios,cross-domain recommendation,where users and items do not overlap across domains,has been a topic of significant interest,driven by considerations such as privacy protection and business competition.Knowledge transfer based on semantic relationships is the primary means to address this research problem.However,in prior academic research,many models based on semantic relationship representations often assume a homogeneous network graph as input,failing to effectively integrate diverse information from heterogeneous network graphs across domains and neglecting node representations under different semantic paths,thereby overlooking the potential roles of meta-paths in user-item interactions and the insufficient exploration of implicit features of users and items.Consequently,integrating node feature representations from multiple layers of meta-paths with collaborative filtering models is crucial in cross-domain recommendation research.Furthermore,different domains possess distinct user profiles,recommendation items,and user behavior data.The fusion of multi-source information objects across domains represents the greatest challenge in cross-domain recommendation.Additionally,when dealing with dense interaction networks,effectively extracting network structural features and network label information and more fully integrating both into the recommendation model are key to improving recommendation accuracy and computational efficiency.To address these challenges in cross-domain recommendation,especially in scenarios where users and items do not overlap across domains,this dissertation delves into the issues mentioned above from the perspective of heterogeneous network graphs.The specific contributions are as follows:(1)Cross-domain Information Fusion Problem When Users and Items Do Not Overlap Across Domains.This dissertation presents a cross-domain recommendation model called Graph Community Enable(GCE)based on meta-paths and community partitioning of heterogeneous network graphs.The model facilitates cross-domain recommendation by constructing heterogeneous network graphs that integrate semantic relationships between different domain features.The recommendation process consists of three phases:(1)Cross-domain Heterogeneous Network Graph Construction Phase: Addressing the challenge of establishing connections between domains with different associated attribute representations,the dissertation first calculates the similarity of associated attributes using a text similarity algorithm for each domain.Then,it employs the Infomap algorithm to partition the associated attributes between the two domains into communities,enhancing attribute relevance.(2)Meta-Path Design Phase: The cross-domain recommendation model developed in this dissertation is primarily applied in the field of course and job information recommendation.In the course recommendation task,three types of users are analyzed,while in the job recommendation task,two user types are considered.Different meta-paths are designed for different user types.(3)Collaborative Recommendation Phase: Leveraging the meta-paths designed in the previous stage,the dissertation employs a random walk algorithm to obtain the best path sequences that are most similar to user demands,and it retrieves the top-K items.Finally,a collaborative filtering approach is used for recommendation.The model’s effectiveness is validated through evaluation on two real datasets.(2)Problem of Weighted Fusion Representation of Network Nodes Under Different Meta-Paths in Cross-Domain Heterogeneous Network Graphs.To address the issue of inadequate exploration of cross-domain user and item features,this dissertation introduces the Cross-Domain Recommendation Model called Graph Community Attention Enable(GCAE),which is based on an attention mechanism for cross-domain meta-path weighted fusion representation.Firstly,it establishes connections between the target domain and auxiliary domains through associated attributes to create a cross-domain heterogeneous network graph.It designs meta-paths at both the intra-domain and cross-domain levels,generates node sequences using a random walk strategy,and employs the Metapath2vec++ algorithm to extract original node information representations.Secondly,concerning information integration,it introduces an attention mechanism to obtain weights for different meta-paths.These weights are used to learn weighted representations of nodes based on intra-domain and cross-domain requirements.Subsequently,an extended matrix factorization model is employed to train a score predictor for cross-domain recommendation.Lastly,the proposed model is tested,and experimental results indicate that by fusing node information from different meta-paths,GCAE outperforms traditional classical algorithms in exploring user preferences and item features,thus validating the effectiveness of the proposed model.(3)The Challenge of Comprehensive Integration of Heterogeneous Network Graph Structural Information with Recommendation Models.To address the issue of fusing network structure information with recommendation models in the context of dense interaction networks,we propose the Cross-Domain Recommendation Model known as Graph Community Tag Enable(GCTE),which is based on score label similarity and overlapping community partitioning.First,we initiate the process from both intra-domain and cross-domain meta-paths.Through a random walk algorithm,we generate node sequences,filtering nodes to retain only user nodes.This enables us to construct a homogeneous network graph of cross-domain users in tensor format.Next,we employ the skip-gram model to learn user node representations and use an attention mechanism to determine weights for different meta-paths.Subsequently,we perform weighted fusion of node degrees,extracting seed nodes.We then proceed with network node overlapping community partitioning through an improved seed expansion and label propagation method.Following this,by incorporating a regularization term based on label similarity among community neighbors,we employ a matrix factorization model to combine graph structure information and score label similarity information for cross-domain recommendation.Finally,we conduct parameter analysis experiments and comparative analysis experiments.The results indicate that the proposed model is more effective in avoiding the impact of noisy data in high-density recommendation environments compared to other classical network graph algorithms.This study addresses several cross-domain recommendation challenges when users and items do not overlap across domains.It employs semantic relationship modeling to construct heterogeneous network graphs and utilizes various models to investigate how user preferences and item attributes can be represented in different domains.The aim is to enhance the precision,accuracy,and interpretability of cross-domain recommendation systems while mitigating issues related to data scarcity and cold starts.This research enriches and refines the theoretical framework for cross-domain recommendation systems in scenarios where users and items do not overlap in different domains.It offers valuable theoretical guidance for the application of cross-platform recommendation systems.Additionally,it also helps enhance the platform’s ability to understand user preferences,improve user experience,and increase user engagement with the platform..