Cross-domain Transfer Recommendation Model Based On User Genre Preference

With the development of the Internet and related technologies,while enjoying the convenience brought by the Internet,people are also facing a serious problem of information overload.In the face of massive information,it is not only difficult for users to obtain useful information,but also make them uncomfortable.The recommendation system is an effective way to solve the problem of information overload.It can learn the user’s preference information from the user’s historical records,and based on this,predict the user’s future needs and make personalized recommendations.The recommendation system not only greatly improves the efficiency of users to obtain effective information,improves the user experience,but also brings benefits to the enterprises that apply the recommendation system.Due to the structural problems of recommendation tasks,recommendation systems often face data sparse issues and cold start problems.Therefore,a lot of researches on recommendation algorithms have focused on alleviating these two problems.Cross-domain recommendation is an effective way to alleviate the problem of data sparseness and cold start users.This type of algorithm is based on the fact that users often have interaction records in multiple domains,and such interaction records can reflect the user’s preference information.In order to improve the recommendation performance in the target domain,these algorithms learn knowledge from other related domains and transfer to the target domain as a supplement to the user-item interaction information in the target domain.However,the existing cross-domain recommendation algorithms such as the collective matrix factorization algorithm and the codebook transfer algorithm have weak expression capabilities and cannot learn deep and complex user-item interaction modes for transferring.CoNet uses deep networks to learn complex user-item interaction function and transfer between domains,but only interaction information is used,and rich auxiliary information of items is not included in the model.This article proposes a Genre-based Collaborate Cross Network(GCCN)model based on user genre preferences for cross-domain recommendation of literature products(such as books,movies,and TV series)in completely overlapping user scenarios,to improve the recommendation performance in these domains.Specifically,this article integrates the user’s preference in literary genre into a cross-domain recommendation deep network based on deep learning technology.By introducing a cross-network structure,the transferring matrix of each hidden layer is used to combine two separate deep networks.This integration realizes dual transfer of knowledge between the two domains,and automatically learns the migration mode from the data to give different weights to transferred knowledge and filter out noise in other domains.Compared with the single-domain recommendation model and the cross-domain transfer model CoNet that uses only user-item interaction information,the GCCN model has a significant improvement.It has been improved by 1.6 and 2.7 percentage on the two indicators of HR@10 and NDCG@10 respectively compared to the state-of-the-art cross-domain recommendation model CoNet.This effective improvement can be contributed to the following points:1.We found the special nature of the recommendation domains.In general,different sub-domains in the literature commodity domain have similar genre attributes,and users’ genre preference in different sub-domains are often shared.That is to say,the user’s genre preference is independent of the literature commodity sub-domains.We extracted such user preference through a genre classification model,and used the fastText model to extract the genre characteristics of the product based on its description as a supplement to the user characteristics;2.We improved the cross-domain collaborative cross-network(CoNet)from the perspective of genre preferences.To better model individual literary genre preferences,we map the product description information in different domains to the same field,and feed into the deep cross-network with the user-item interaction features.So that the model can learn not only the complex user-item interaction mode,but also the user-genre interactive features,which can be transferred to improve the recommendation performance of the target domain.Through comparative experiments on the Amazon dataset,we evaluated the performance of the GCCN model.Experiment results show that our proposed GCCN model performs better than other models,proving that the genre preference characteristics of the product are helpful for modeling user preferences.At the same time,the genre preference characteristics of users learned from different domains can effectively alleviate the data sparse issue in the domain,thereby improving the performance of the recommendation system.