Ingredient Embeddings Constructed by Biased Random Walk on Ingredient-Compound Graph
概要
With the recent popularity of food computing, there is a growing demand for research on creating ingredient embeddings by representation learning. The general-purpose representation obtained from a latent space of food ingredients can aid in developing various applications related to food computing. Existing methods create ingredient embeddings based on the ingredient-compound graph, and the co-occurrence in recipe data construct ingredient relationship. However, existing methods need help with the learning process for ingredient representation. When generating a path to input to the graph embedding model, the path disregards the cooccurrence information in the recipe. This method treats high-frequency and low-frequency ingredients in the same way. Hence, when using ingredient embeddings created with existing methods, the need for detailed recipe information may prevent accurate food ingredient recommendations based on the recipe (food pairing recommendation, alternative ingredient recommendation). Our study proposes a novel ingredient embedding method that can solve the abovementioned problems by constructing an ingredientcompound network expressing a containment relationship between an ingredient and its chemical compounds. Our experimental evaluation in the classification task of food ingredients indicated that our method outperforms existing methods, so our ingredient embeddings can express their features in the task.
引用情報
Naoki Yoshimaru, Kazuma Kusu, Yusuke Kimura, , Kenji Hatano, Ingredient Embeddings Constructed by Biased Random Walk on Ingredient-Compound Graph, Procedia Computer Science, Vol.225, pp.3948-3957, 2023-09, DOI: 10.1016/j.procs.2023.10.390.