Graph Neighbor And Knowledge Semantic-based Personalized Recommendation

With the development of the Internet industry and social productivity,the volume of multi-media information in the social media environment increases exponentially.The scenario lays a great obstacle for users on information searching.Information overload has become a bottleneck restricting the development of related industry.To alleviate the issue,recommender systems emerge with information filtering as the core,which effectively helps users search their satisfied contents from the overloaded multimedia contents.It plays to bridge users and social media networks,attracting extensive attention from both fields of academic and industry.Collaborative filtering currently is the most widely employed recommendation algorithm in industry.It mainly relies on users’ historical interaction records to predict users’ interests and provide users personalized information recommendation,aiming at improving users’ experience.However,the extremely sparse user interactions limits the performance of collaborative filtering techniques.With the limited historical interactions,it is difficult for the recommender system to completely uncover users’ interests,especially inactive users.Therefore,recommender systems meet a great challenge from both interaction sparsity and cold start problems.In recent years,with the rapid development of deep learning techniques,the related learning techniques of graph neural network and knowledge graph provide a strong theoretical support to address the above-mentioned issues.This dissertation aims to investigate the multi-order collaborative information hidden in users’ interaction records with heterogeneous graph propagation.Besides,we explore the rich semantic in the knowledge graph to effectively guide the learning of users’ interests and alleviate the limitations for personalized recommendation.This dissertation mainly investigates the following four works.1.We propose a non-pairwise collaborative filtering model for personalized recommendation.It investigates local neighborhood structure of the heterogeneous interaction graph to mine non-pairwise collaborative signals and integrates pairwise collaborative signals and non-pairwise collaborative signals for interaction learning.The aim is to leverage non-pairwise collaborative signals hidden in local structure of the interaction graph to predict user’ interactions.For the target user-item pair,we design three manners to evaluate the collaborative intensity and adaptively aggregate the collaborative information in the non-pairwise neighbors for personalized recommendation.2.We propose a siamese graph-based dynamic matching model for collaborative filtering.It constructs a dual-channel embedding propagation mechanism to extract multi-order collaborative information from homogeneous neighbors for embedded learning.Considering the diversity and complexity of users’ interests and items’ attributes,we additionally introduce a dynamic embedding learning strategy to improve personalized recommendation performance.3.We propose a dual-channel common and special embedding-based collaborative filtering model for personalized recommendation.It explores multiorder commonality and semantic speciality between users and items to mine users’ interests in multi-view aggregation.Specifically,an embedding propagation network plays to mine the multi-order association between users and items.Besides,knowledge embedding learning performs to model the semantic speciality of users and items,respectively.The multi-order commonality and semantic speciality are aggregated to model users’ interests for personalized recommendation.4.we propose a knowledge-aware multi-space embedding learning model for personalized recommendation.It employs the rich semantic association between items in knowledge graph to alleviate the interaction sparsity issue.Embeddings of users/items are learned from multiple independent semantic structures to mine users’ interests on different semantics.We additionally build a dual-channel target-aware attention mechanism to adaptively integrate the embeddings in multiple semantic spaces,which models users’ interests in a complementary manner of multiple semantic spaces for personalized recommendation.