Research On Multi-view Knowledge Representation For Cross-domain Recommendation

With the rapid development of the Big Data era,it is vital to use Big Data techniques to learn useful information from data.Recommendation systems provide a powerful support to solve information overload,but there is also a cold-start problem when new users or new items appear.Therefore,academics propose cross-domain recommendation techniques to solve the data sparsity and cold-start problems in the target domain by learning user preferences or item knowledge representations in the auxiliary domain to improve the recommendation performance of the target domain.However,current crossdomain recommendation models have much room for improvement in learning item association relationships between domains and common knowledge extraction.In recent years,graph neural networks have made good progress in structural data representation learning.The use of graph neural networks helps to learn user-item interactions and itemitem associations at a deep level,and helps to enrich the knowledge representation learning of items.In this paper,we conduct research on item representation learning for cross-domain recommendation,and propose a multi-view knowledge representation method for cross domain item clustering(MKRM)for the learning of common item knowledge between domains.A network for learning deep implicit features of items is designed,incorporating a multi-headed attention mechanism and a memory module.Based on this,a crossdomain-oriented recommendation model(MKRM-Component Learning,MKRM-CL)is finally proposed.Its main research elements are as follows.(1)To address the problem that the existing cross-domain item common knowledge learning does not make full use of item association relationships,this paper constructs a user-item-attribute heterogeneous information network for the auxiliary domain and the target domain.Considering the existence of misalignment of item attributes between domains,these different attributes are clustered using methods such as statistical calculation of document word frequencies to obtain attribute classes shared by the auxiliary domain and the target domain,and items are divided into several different subgraphs of the heterogeneous information network according to the attribute classes to form multiple views.(2)To address the problem that the knowledge representation of an item does not make full use of the multiple attributes of the item,this paper designs a multi-view-based item representation learning method,which learns higher-order graph structure information and implicit user preference information in the item representation from each view respectively,and uses graph attention networks to concatenate the features of the knowledge representation of the same item in different sub-views to form an item knowledge representation containing multiple attributes.The multi-view knowledge representation method MKRM based on item clustering is proposed and used for downstream recommendation tasks to generate better recommendation content.(3)To further learn the deep implicit features of items,this paper constructs a network for learning the deep implicit features of items after clustering items in multiple views and learning the knowledge representation of items using graph attention networks using the memory module and multi-headed attention mechanism to learn each component of items and adding user preferences to mine the deep representation information of items.Finally,the cross-domain recommendation-oriented model MKRM-CL is constructed.(4)Based on the proposed MKRM-CL model in this paper,experiments were conducted on the Amazon book and movie datasets,as well as and the Douban book and movie datasets crawled using a crawler,respectively,and compared with the latest baseline algorithm to verify the effectiveness of the model,with a 5%-10% improvement in performance.