| With China’s rapid economic development,people’s demand for a better diet is growing.Personalized and diversified recipe recommendations can better meet people’s dietary needs,which in turn stimulates food consumption and contributes to the large domestic circulation,domestic and international double circulation and economic development.Therefore,designing efficient access to personalized recipe recommendations has become a topic of practical significance.This study focuses on personalized recipe recommendation based on knowledge graphs,and the main contributions are as follows:(1)A dataset enhancement method using knowledge graphs to diffuse preferences is proposed.To address the problem of sparse user-recipe interaction records in the homemade recipe dataset,we use the Ripple Net preference propagation idea to enhance the user-recipe interaction records by putting the high-frequency shared-attribute entity samples on the preference diffusion knowledge graph into positive samples and the non-shared-attribute entity samples into negative samples.(2)An improved Ripplenet-CA model using shared-attribute sampling is proposed.The Ripple Net-CA model is improved in terms of sampling methods and evaluation metrics to address the problems of limited depth of Ripple Net interest propagation and insufficient diversity of recommendation results.Random sampling of shared attribute frequencies is used to retain the shared attribute frequency information in features to enhance recommendation diversity and accuracy.Next,error thresholds and diversity metrics are designed to control the sampling effort to verify the impact of the improved model on recommendation diversity.(3)A Ripple Net-Trans R model with joint Trans R loss function for training is proposed.The Trans R knowledge graph loss function is added to address the problem that the original knowledge graph loss function Rescal of Ripple Net model has poor ability to portray the interactions between the head and tail entities and inter-relational bands of the knowledge graph,and the Trans R loss function aims to optimize the interaction perception ability of the head and tail nodes of the knowledge graph of the model,so that the trained knowledge graph embedding can better represent the recipe knowledge graph structure of the recipe knowledge graph.This improvement allows the model to predict user preferences more accurately and to perform better recommendation ranking.At the same time,the Trans R loss function can prevent the model from overfitting and improve the generalization ability of the model.(4)A recipe knowledge graph is constructed using Neo4 j and personalized recipe recommendation is completed.To address the problem that relational databases are weak in processing graph models,Neo4 j is used to build a knowledge graph based on the graph model and perform recipe recommendation on the recipe knowledge graph.In summary,this study obtains the recommendation list by training the embedding vector of user preferences and recipes to be recommended,calculating the similarity between preferences and recipes and ranking them.Among them,the dataset enhancement method reduces the sparsity of the recipe dataset;the Ripple Net-CA model improves the diversity of recommendation results;the Ripple Net-Trans R model improves the ability of the Ripple Net model to perceive different relationships;the Neo4j-based personalized recipe recommendation improves the database query efficiency and enhances the interpretability of the recommendation results.These improvements can help recommendations better meet user needs and enhance user experience,thus stimulating people’s food consumption. |
