Research On Personalized Food Recommendation Method Based On Knowledge Graph

With the development of the national economy,the variety of food is becoming more and more diverse,and people's demand for food continues to expand,it is important to accurately and intuitively present dishes that meet their diverse needs from a large amount of food information.Recommendation systems can help users spend a small amount of time to find their favorite dishes,but in practical applications,most recommendation systems use collaborative filtering recommendation algorithms that only use user rating data for recommendation.There is a data sparsity problem,which limits the recommendation effect to a certain extent.The knowledge graph contains rich semantic information of items,which can be used as a very useful auxiliary data to enrich the feature description of users or items,and integrating it into the recommendation system will help improve the effect of traditional recommendation methods.This paper focuses on the problems in the traditional food recommendation algorithm based on collaborative filtering,studies the recommendation method based on the knowledge graph,and designs a personalized dish recommendation model based on the food knowledge graph,including the construction of food knowledge graph,user interest modeling based on knowledge graph,and collaborative filtering dish recommendation algorithm fused with knowledge graph.The main research content includes the following three points:(1)Based on real dish data from websites such as Dianping,constructed a food knowledge graph.During the construction process,the knowledge extraction method was researched,and the rule-based knowledge extraction method was designed according to the characteristics of structured and semi-structured data.For text data,a named entity recognition method based on the BERT-Bi LSTM-CRF model in the gourmet field was designed to solve the problem of multiple aliases in named entity recognition,and use Bootstrapping to extract the relationship between entities.(2)In order to make full use of the rich knowledge contained in the knowledge graph,an improved user preference vectorization method based on entity attributes is proposed.Based on the vectorization of the food knowledge food using the Trans D method,the attribute weight is calculated,and the user's preference for the dish is calculated.The interest is turned into a focus on multiple attributes to achieve a more accurate and fine-grained expression of user interest.(3)Aiming at the data sparsity problem of traditional recommendation algorithms,a collaborative filtering personalized dish recommendation model(Trans D-MF)that integrates food knowledge graph is designed,by calculating the similarity of users' dietary preferences,and introducing it into the matrix factorization recommendation model based on implicit feedback,collaborative model training,combined with the rich knowledge of the domain knowledge graph to enhance the recommendation.Experiments were carried out on the real food data of Dianping.com.The experimental results show that the Trans D-MF model is better than the related mainstream recommendation algorithms in the Precision and Recall of Top K dish recommendation tasks,which verifies the effectiveness of the model.