Research On Multi-Domain Resources Recommendation Methods Based On Neural Network

With the advancement of education informatization,online education platforms have developed rapidly.There are massive learning resources on the platform which enriches the choice of learner but also leads to information overload problem.Personalized recommendation can filter information to provide learners with resources of interest and value automatically and efficiently,as one of the key technologies of adaptive teaching service.Performance of personalized recommendation has been suffered from data sparsity problem.Compared with the huge amount of all resources and users,records of user-resource interaction are fewer.In traditional personalized recommendations,resources to be recommended is of same type,which is called single-domain recommendation.In addition,learners have different types of learning content at different stages,and may use different online education platforms.Therefore,to alleviate the problem of data sparseness and fit learners' real learning situations,this paper proposes a multi-domain recommendation strategy.Multi-domain recommendation can supplement relevant information that can be shared between domains for a single domain.It refers to the joint use of interaction records on resources of multiple domains to improve the recommendation effect within each domain.This paper mainly carried out the following two tasks about multi-domain recommendation:(1)Multi-Domain Neural Network Recommender(MDNNR) is proposed.Different domains refer to different kinds of resources in the same system where has same users.Based on usage records of users on resources in different domains,a novel multi-branch neural network is used to mine the preferences shared by users among domains and the unique preferences in each domain to construct a collaborative filtering recommender.The user-sharing feature of the model is not the same value between the domains,but the same transformation,The model was validated on real datasets,which demonstrated the validity of the proposed recommendation method.(2)Based on the idea of MDNNR,a Multi-Domain Deep Learning Recommender(MDDLR)is proposed.Different domains refer to different heterogeneous systems with different resource types and no corresponding users.By pre-trained word vectors,the convolutional neural network is used to process short text content information such as titles and tags,which helps multi-branch neural networks to establish semantic bridges between domains.Combined with usage records,shared features and domain-specific-features are get by decomposition of deep learning to accomplish multi-domain recommendations.Experiments were conducted on the public datasets to verify the validity of the proposed recommendation model.The two multi-domain recommendation models proposed in this paper are designed for recommendation of educational resources,but can also be applied to the recommendation of other types of item systems.Due to the lack of real data from online education platforms,this paper uses real datasets of similar structure to test model.