High-Reputation Food Formulas: A Heterogeneous Information Network Representation and Semantic Analysis Approach

In food research and development, fully capturing consumer preferences is crucial for enhancing ingredient selection and formulation. However, many decisions still rely heavily on expert judgment. To address this, we propose a semantic analysis framework that combines heterogeneous information network representation learning and lexicon-based sentiment analysis to uncover relationships between ingredients and consumer word-of-mouth. We apply our approach to 1062 breakfast products and 58,206 reviews from Amazon, extracting taste- and ingredient-related sentiment while excluding non-formulation factors (e.g., price, packaging, logistics). Through embedding-based visualization, our product space groups items according to similar formulation characteristics, while the ingredient space reveals functional clusters among ingredients. We find that high-reputation products and their related ingredients form distinctive clusters in both spaces, often emphasizing natural, organic, and gluten-free attributes, whereas items containing synthetic additives appear more frequently in low-reputation regions. This framework offers a data-driven complement to traditional R&D methods, providing a perspective distinct from conventional approaches for product formulation and ingredient selection.