Recommender Systems Via Multi-user Representation Optimization And Fine-grained Preference Aware

The cold start and sparsity problems have been plaguing collaborative filteringbased recommendation algorithms.Knowledge graphs are able to provide structured information and rich semantic information,and therefore,they are used to improve the performance of recommender systems.In personalized recommender systems,user preference information plays an important role in knowledge graph-based recommender systems,as reflected by the fact that users have different preferences for each entity-relation pair in the knowledge graph.Existing methods do not model such fine-grained user preference features well,which affects the performance of recommender systems.Furthermore,potential connections among users also contain personalized information about them,and many studies introduce auxiliary information such as social networks to improve the performance of recommendation systems.However,many recommendation scenarios cannot get explicit relationships between users with social attributes,so only implicit relationships can be established to learn the similarity between users.Implicit relationships establish potential connections between users through the similarity of their interaction behaviors on third-party media,and introducing implicit relationships in the recommendation process can effectively improve the performance of recommendation systems.Existing approaches tend to build user-implicit relationship networks through the overlap of behaviors between two users which ignore the interaction patterns between multiple users,making the constructed implicit network structure unable to contain the interaction information of multiple users.It leads to bringing noise or omitting useful information in the learning of network features of user representation.(1)Modeling fine-grained preference features for users,we propose a deep reinforcement learning-based knowledge preference-aware network,or KPRLN for short,which learns the fine-grained preference features of each user-entity-relation by building paths between user history interaction items in the knowledge graph with a deep reinforcement learning model.A hierarchical propagation path construction method is designed to solve the problem of deep reinforcement learning for learning preference information of dangling entities and long path exploration in the knowledge graph.The method forms clusters by spreading outward from the center of the item and uses clusters to represent the starting and target states in deep reinforcement learning.With the iteration of cluster diffusion,the model can better learn the preference information of overhanging entities in the knowledge graph and explore more distant paths.After a sufficient number of iterations,the model generates a weighted knowledge graph with the user’s fine-grained preference features and aggregates higher-order representations of items containing the user’s fine-grained preference features in a graph convolutional network based on an attention mechanism for use in the recommender system.(2)Aiming at multiple implicit relational feature modeling for users,we propose a recommendation method for optimizing multiple user representation based on hypergraph motifs,referred to as HMKRec.First,the knowledge graph is used as auxiliary information to construct a user-item interaction hypergraph and map it to the user-user adjacency graph.Then,association patterns among multiple users are learned based on the multivariate structure in the hypergraph,and an implicit relationship network with weights and directions is built to explore the higher-order relationships among multiple users.Finally,to aggregate item features and user-implicit relationship features,a hierarchical graph convolution is designed to fuse the hypergraph convolutional network and the graph convolutional network to obtain higher-order representations of users for use in the recommendation system.The performances of the models are evaluated in multiple real-world recommendation scenarios using three widely used publicly available datasets,Moilens-1M,Last.FM,and Book-Crossing.Extensive experiments show that both models proposed in this thesis,KPRLN and HMKRec,outperform other state-of-theart baselines.Adequate comparison experiments are also designed to verify the effectiveness of each algorithm component.