Collaborative Recommendation Model Based On Heterogeneous Information Network

At present,among the existing recommendation,the challenges to be solved urgently include data sparsity and cold start matter.In order to alleviate this issue,the recommendation system is constantly improved and optimized under the research of academic circles.the traditional Collaborative filtering(CF)method only learns the potential factors from user-item rating,which cannot improve the data sparsity and cold start matter.Some improved CF methods enrich the prior information of latent factors by introducing auxiliary information.In addition,cross-domain recommendation has recently been proposed to solve these two types of problems,and the key to this approach is to learn more comprehensive and accurate user representations in both source and target domains.However,because most methods use a relatively single type of auxiliary information,the potential factors of learning may not be very effective.Aiming at the above problems,the heterogeneous information network(HIN)can effectively represent and utilize different data information due to its flexibility in modeling the heterogeneity of data.This dissertation proposes two collaborative recommendation models based on heterogeneous information networks.One is A Heterogeneous Information Network based Tightly Coupled Recommendation Model(HTCRec),which makes use of heterogeneous information network embedding and tight coupled collaborative filtering framework to make personalized recommendation.Specifically,the meta-paths and their corresponding path instances in the HIN are aggregated first,then the auxiliary information of the target user or project is represented by the embedding of their aggregated meta-paths using attention mechanism,and finally the meta-paths are explicitly merged into the tightly coupled interaction model to complete the personalized recommendation.The other is A Heterogeneous Information Network based Cross Domain Recommendation model(HCDRec).First by learning more effective user and project potential factors in the target and source domains.Specifically,for the source domain,a gated recurrent unit(GRU)is used to model the dynamic preference of users.For the target domain,a specific heterogeneous information network is constructed for the intricacy and sparsity of data,and then relational neighbor-based and semantic-based attention aggregation are used to obtain the representation of users and items respectively.After the potential factors of overlapping users are obtained,the feature mapping between two domains is learned by using multi-layer perceptron(MLP).The experimental results on relevant data sets in this paper show that HTCRec and HCDRec has better performance than other recommended models.