Unified Embedding Model Over Heterogeneous Information Network For Recommendation

With the explosive growth of information on the Internet,we have entered an era of "information overload".It is difficult for people to find items they need effectively,so recommendation system emerges as the times require.It helps people to discover the item they may be interested in by modeling users and items.The early recommendation system only considered the user's historical behavior,which may mislead user and item modeling and affect the recommendation results.In recent years,some auxiliary information of users and items has been used to model users and items more accurately.In the real recommendation system,users,items and their auxiliary information can be regarded as different types of objects.And there are different types of relationships among these objects(such as the purchasing relationship between users and items,the ownership relationship between items and their attributes).These different types of objects and relationships form a heterogeneous information network.Recommendation algorithms based on heterogeneous information network(HIN)attract much attention due to its advantage of modeling complex information naturally and effectively.Most algorithms based on heterogeneous information network firstly use multiple metapaths to extract rich semantic information between users and items on heterogeneous network from different perspectives,and then make recommendations.Although existing methods have achieved some improvement,most of them have one of the following problems: 1)they model users and items in isolation under each meta-paths,which may lead to information extraction misled under some paths on which data is sparse and noisy;2)in the process of exploring HINs,the goal of modeling users and items is only to mine structural features of HINs,which may lead to irreversible loss of some information important to recommendation but less important in the network structure.This loss of information affects recommendations.3)When fusing the information from different meta-paths,the personalized weights of metapaths connot be learned effectively.In order to solve the problem of existing methods,this paper proposes a recommendation algorithm based on unified representation learning in heterogeneous information network,called HueRec.Specifically,in view of the above three problems,the contributions of this paper can be summarized as three parts:1)This paper presents a unified representation learning method for users and items based on different meta-paths,by which we can model users and items more accurately.We assume there exist common characteristics of users and items on different meta-paths,and use data from all meta-paths to learn unified users' and items' representations.So the interrelation between meta-paths are utilized to alleviate problems such as data sparsity and data noise that may exist on a single meta-path.2)This paper proposes an end-to-end HIN information extraction and exploitation method to avoid the loss of effective information in the process of network information extraction.In this paper,the information extraction part and the recommendation prediction part are combined into an end-to-end training process.The objectives of modelling users and items include not only the mining of the structural information on the HIN,but also the prediction of users' preferences for items.Therefore,the proposed method can effectively extract network structural information useful for recommendation.3)Based on the attention mechanism,this paper effectively learns the personalized weight of meta-paths under different users,which improves the performance of personalized recommendation.In this paper,users,items and meta-paths are embedded into the relevant latent semantic space,and then users' preferences for different meta-paths can be effectively measured based on attention mechanism.Finally,experiments were conducted on three real data sets.The experimental results showed that the proposed HueRec method always outperforms state-of-the-art methods in the top-n recommendation problem,and confirmed the three contributions mentioned above.