Research On Recommender Systems Based On Heterogeneous Information Networks

With the explosive growth of online information,various network platforms are facing with problems such as information overload.The recommender systems can accurately capture preferences of users and suggest items that they may be interested in,thus improving users’ satisfaction with the platform,and has attracted wide attention.Heterogeneous information network models complex objects and connections into different types of nodes and edges respectively,and different node sequences(meta-paths)represent different semantic information.Based on the metapath,the relationship between nodes can be accurately captured.Therefore,the meta-path has rich semantic and structural information.Under this background,many recommender systems based on heterogeneous information network have emerged.However,this kind of models still has many problems,such as not fully considering the time dimension information and multi-embedding fusion mode,etc.In order to further enhance the performance of the recommender systems and meet the high demand for information matching in the current era,this thesis studies the recommender systems based on heterogeneous information networks.The main contents and innovations of this thesis are given as follows:(1)This thesis introduces the development of recommender systems and heterogeneous information networks,as well as the related contents of classic models,including Herec,HAN,NIRec,and MCRec.It focuses on the network structure,principles,and advantages and disadvantages of these models.(2)In view of the problem that most existing recommender systems based on heterogeneous information network only pay attention to structural information while ignoring time dimension information,a temporal enhancement for recommendation(TMRec)is proposed.The model introduces temporal convolutional network,long short-term memory and self-attention mechanism.The whole model can be divided into three parts: user,item and meta-path embeddings learning.The temporal convolutional network module is introduced to capture the time relationship between different types of nodes in the sequence.In addition,L2 normalization is used to handle the embedding of a single metapath in this module to avoid over-fitting.In addition,when learning user and item embedding,long short-term memory and self-attention mechanism modules are used to highlight the time relationship and interaction of nodes of the same type on the path,thus enhancing their embedding representation.Ablation and comparison experiments are conducted on Last FM,Movielens and Yelp datasets,and the results verify the superiority of the proposed TMRec model.(3)Most existing recommender systems based on heterogeneous information network do not fully consider the fusion mode of multiple embedding and the local feature amplification and interaction of each node in the process of embedding learning.A multi-embeddings fusion for recommendation(MFRec)is proposed.Firstly,to solve the problem that the mode of multiembedding fusion is not fully considered,a new multi-embedding fusion module,namely feature importance and bilinear feature interaction network,is introduced into the MFRec model.The module uses the squeeze-excitation network to dynamically learn the importance of the features.A bilinear interaction function is used to better model the cross features.The interactions between the embeddings are located at a finer level.And the final embedding representation is enhanced.In addition,in view of the problem that local feature amplification and interaction are not emphasized at each node in the learning process of each embedding,the MFRec model uses an object-contextual representation network module in the meta-path embedding learning part to enhance contextual amplification and interaction of similar nodes.Then the the co-attention mechanism is used to improve the embedding.In addition,the proposed model also introduces dialed convolution and spatial pyramid pooling modules to obtain multi-scale information and enhance the representation of the meta-path.Ablation and comparison experiments are conducted on Last FM,Movielens and Yelp datasets,and the results verify the superiority of the proposed MFRec model.