The impact of recipe features, social cues and demographics on estimating the healthiness of online recipes

Abstract

Apart from taste, people increasingly consider nutritional facts when choosing a recipe or a meal. Researchers and companies alike aim to support health-conscious choices by providing estimates of a meal's calories and levels of sugar, fat, protein and salt. Features that are typically considered for these automatic estimates include a recipe's title, ingredients and cooking directions, as well as photographic material. Little is known based on which features users estimate the healthiness of online recipes themselves. Making use of data derived from a large online food community, and data collected via crowdsourcing, we compare the performance of algorithmic nutritional estimation with the performance of human-provided estimates, and analyze the most influential features used by humans and machines. Our results indicate that simple models already outperform human raters. Basic features such as title, ingredients and cooking directions were more informative than pictures of a recipe or user comments. For human estimates, we observed effects due to age and gender, but not due to dietary preferences or cooking habits. Our quantitative and qualitative results provide guidance for the development and evaluation of methods for nutrition estimation, and give insight in which features are most useful for nudging people into making healthier diet choices.