The Design And Implementation Of Recommender System Based On Graph Embedding And Item Hierarchical Structure

With the rapid development of Internet information technology,recommender systems help users quickly and accurately find the items they need and are interested in,and gradually penetrate into people's clothing,food,housing and transportation.Items such as music,commodities,and laws in the recommender systems and their categories often show hierarchical structures,where items are subordinate to categories,and subcategories are subordinate to parent categories.In order to design and implement a recommender system for items with such characteristics,hierarchical structure information of items in the practical application scenarios of similar recommendation and personalized recommendation is fully explored in this thesis.The main work content is as follows:(1)In the similar item recommendation scenario,in view of the problem that the existing attribute graph embedding models cannot make full use of the item's hierarchical structure,Hierarchical Attributed Graph Embedding for Item Recommendation(HAGE)is proposed.First,in order to explore the similarity of nodes under the same category,an attributed graph embedding based on the item hierarchical structure is proposed.The graph is first constructed for the sub-nodes under each category node,and then the graph constructed by each category is used to generate random walk sequences.Secondly,a method of fusing the similarity of co-occurring items in the user behavior sequences and the similarity of items under the same category is proposed,which shares the common parameters of the two Skip-gram networks for alternate optimization,thereby generates a unified item embedding.Experiments was conducted on the Amazon e-commerce dataset to verify the effectiveness of HAGE.(2)In the personalized recommendation scenario,in view of the problem that the existing interaction graph embedding models cannot make full use of the collaborative similarity including high-order interactions,Collaborative Similarity Interaction Graph Embedding for Personalized Recommendation(CIGE)is proposed.First,the collaborative similarity based on attribute induced edge lists is introduced to mine the second-order neighbor proximity between two users or two items connected by the same discrete attribute.Secondly,by introducing the first-order neighbor information and attribute information of the node into the Skip-gram networks,and using the attention mechanism to weight the aggregation vector of the node,it can mine the high-order implicit connection of the attribute interaction graph.Experiments was conducted on the Amazon e-commerce dataset to verify the effectiveness of CIGE.(3)In the personalized recommendation scenario,in view of the problem that CIGE cannot make full use of the hierarchical structure of items,Hierarchical Interaction Graph Embedding for Personalized Recommendation(HIGE)is proposed on the basis of CIGE.Firstly,Random Jump is proposed.The co-occurrence information of items in the user behavior sequences is mined by the movement of nodes in the same level,and the hierarchical information in the item tree organization structure is mined by jumping between nodes in different levels.Secondly,the item embeddings obtained by Random Jump is regarded as the attribute information of the item,and combined into CIGE by concatenating.Experiments was conducted on the Amazon e-commerce dataset to verify the effectiveness of HIGE.(4)We designed and implemented a recommender system for the dataset with the characteristics of the item's hierarchical structure.The recommended scenarios are distinguished by the users' login status.In the similar item recommendation scenario,HAGE's research results are used to generate recommendation lists for users.In the personalized recommendation scenario,HIGE's research results are used to generate recommendation lists for users.The API response time was optimized through distributed caching and other technologies,and the system was tested for function and performance.