FOOD RECOMMENDATION BASED ON INGREDIENTS AND RECIPES USING TEXT EMBEDDING AND NETWORK ANALYSIS: APPLICATION TO KOREAN FOOD

초록

This study presents a culinary recommendation system that employs recipe similarities to suggest foods. The system proposes the foods that align with an individual’s preferred food based on comparable ingredients and cooking processes. Recipes are transformed into high-dimensional vectors and their cosine similarities are used to quantify their closeness. We use the Sentence-BERT model, which is a pre-trained large language model, for the vector transformation. To evaluate the effectiveness of this similarity measure and fine-tune the transformation model, we use the human survey for food similarity. The result shows that the proposed similarity measure is significantly aligned with the human judgement (with the Spearman correlation 0.5702). Additionally, we construct a clustered network that offers an intuitive overview of relationships among numerous foods, enabling more accurate recommendations compared to rank-based recommendations. © 2025, ICIC International. All rights reserved.

키워드