Reseach And Implementation Of HIN Embedding For Recommendation Algorithm

In the era of information overload,recommender systems plays a piv-otal role in various online services,which aim to match user interests with resource items.Classic recommendation methods,e.g.,matrix factorization,mainly model users' preference towards items using historical user-item inter-action records.Nowadays,various kinds of auxiliary data become available in recommender systems,which can be leveraged to improve recommendation performance.Recently,heterogeneous information network(HIN),consisting of either multiple types of nodes or links,has been proposed as a powerful modeling method to fuse complex information,and is successfully applied in many data mining tasks.Due to its flexibility in modeling data heterogeneity,HIN has also been adopted in recommender systems to characterize rich auxiliary data in recent years,and those algorithms are also called HIN based recommen-dation methods.Existing HIN based recommendation methods mainly rely on meta-path based similarities and commonly learn a linear weighting mechanism to combine the path similarities or latent factors for recommendation.These methods fail to effectively extract and exploit information for HIN recommen-dation.Network embedding,which aims to learn a low-dimensional representa-tion vector for each node,has shown its potential in structure feature extraction and has been successfully applied in many data mining tasks.Among existing network embedding methods,most of them focus on homogeneous networks,and thus they cannot directly be applied to heterogeneous networks.Therefore,attention is increasingly shifting towards heterogeneous information networks,which aims to learn semantic representations to capture complex structure and rich relationships.Although these HIN embedding methods has shown their effectiveness in some tasks,they usually focus on general node embeddings,seldom considering the specific embedding for the recommendation task.To address the above issues,in this paper we propose a series of HIN based recommendation works to leverage heterogeneous inform ation network embed-ding for recommendation,consisting of three research works and one prac-tical application.We effectively model heterogeneous information networks from different aspects,which are the representation for nodes,representation for meta-path based context,and the fusion representation for local and global information.Specifically,these works include:(1)Representation for user and item.We propose a novel heterogeneous network embedding based approach for HIN based recommendation,called HERec,for fully mining latent structure features of users and items in HIN.o embed HINs,we design a meta-path based random walk strategy to generate meaningful node sequences for network embedding.The learned node em-*beddings are first transformed by a set of fusion functions,and subsequently integrated into an extended matrix factorization(MF)model.The extended MF model together with fusion functions are jointly optimized for the rating prediction task.(2)Representation for meta-path based context.We develop a novel deep neural network with the co-attention mechanism for leveraging rich meta-path based context for top-N recommendation,called MCRec.We elaborately de-sign a three-way neural interaction model by explicitly incorporating meta-path based context.To construct the meta-path based context,we propose to use a priority based sampling technique to select high-quality path instances.Our model is able to learn effective representations for users,items and meta-path based context for implementing a powerful interaction function.The co-attention mechanism improves the representations for meta-path based context,users and items in a mutual enhancement way.(3)Representation for local and global information.We propose a uni-fied model LGRec to fuse local and global information for top-N recommen-dation in HIN.We model most informative local neighbor information with a co-attention mechanism and learn effective relation representations to capture rich information in HIN by optimizing a multi-label classification problem.Fi-nally,we combine the two parts into a unified model for top-N recommenda-tion.(4)Application.With the real datasets in Ant Credit Pay of Ant Financial Services Group,we first study the cash-out user detection and recommendation problem and propose a novel hierarchical attention mechanism based cash-out user detection and recommendation model,called HACUD.Specifically,we model different types of objects and their rich attributes and interaction rela-tions in the scenario of credit payment service with an Attributed Heterogeneous Information Network(AHIN).The HACUD model enhances feature represen-tation of obj ects through meta-path based neighbors exploiting different aspects of structure information in AHIN.Furthermore,a hierarchical attention mecha-nism is elaborately designed to model user's preferences towards attributes and meta-paths.