Research On Recommendation System For Graph Data Reasoning

With the rapid development of the internet,the application platform of the amount of data users and products are exponential,the recommendation system is designed to understand user preferences.The performance of the traditional recommendation system,with its high accuracy,is widely used in various scenarios.Due to the lack of interpretability,the traditional recommendation system is not well recognized by users.The introduction of graph data as auxiliary information in the recommendation system can effectively solve the problem of interpretability.Graph data-oriented inference uses the relationship between entities to mine the potential knowledge in the data,and the inference process can be displayed explicitly,which has strong research significance and application value.The main work of this paper includes the following contents:Firstly,a knowledge-aware path reasoning model is proposed to infer the relationship between substitutable and complementary products in the e-commerce environment.We extract structured information from the e-commerce dataset to construct a knowledge graph.We transform the product relationship inference task into a path reasoning problem and use a dynamic policy network to make a precise decision.The experimental results show that the model can effectively use knowledge graph to predict substitutable and complementary product relationships,and achieve competitive performance compared with state-of-the-art approaches.Secondly,aiming at the problem of an attribute-based conversational recommendation system,we propose a two-stage graph data reasoning model.The whole session is modeled as a graph reasoning process.The model consists of two networks(Employer Network and Employee Network).The Employer Network is responsible for deciding the next round of action,asking the user preferences for attributes or recommending products.Employee Network consists of two modules,the inquiry module,and the recommendation module.Responsible for the completion of each task.The whole model learns the optimal interaction strategy through the collaboration of the two networks.Experimental results show that the proposed model achieves the best results on four datasets,the average number of rounds of conversations is reduced by 0.59,and the probability of successful recommendation within 15 turns of conversations is increased by 10%.Thirdly,aiming at the problem of content-based conversational recommendation system,we propose two models to complete the recommendation task and topic prediction task respectively.The model regards the entities contained in the session as the seeds of graph data inference and extends the neighbor entities connected with the seed entities based on certain rules.Experiments show that the path set obtained from the graph provides richer semantics for graph reasoning,which is beneficial to enrich dialog information and improve the accuracy of the model.