Food Pairing Unveiled: Exploring Recipe Creation Dynamics through Recommender Systems

TL;DR

This work applies state-of-the-art collaborative filtering techniques to the dataset, providing a tool that can recommend new foods to add in recipes, retrieve missing ingredients and advise against certain combinations, and confirms the existence of food pairing.

Abstract

In the early 2000s, renowned chef Heston Blumenthal formulated his "food pairing" hypothesis, positing that if foods share many flavor compounds, then they tend to taste good when eaten together. In 2011, Ahn et al. conducted a study using a dataset of recipes, ingredients, and flavor compounds, finding that, in Western cuisine, ingredients in recipes often share more flavor compounds than expected by chance, indicating a natural tendency towards food pairing. Building upon Ahn's research, our work applies state-of-the-art collaborative filtering techniques to the dataset, providing a tool that can recommend new foods to add in recipes, retrieve missing ingredients and advise against certain combinations. We create our recommender in two ways, by taking into account ingredients appearances in recipes or shared flavor compounds between foods. While our analysis confirms the existence of food pairing, the recipe-based recommender performs significantly better than the flavor-based one, leading to the conclusion that food pairing is just one of the principles to take into account when creating recipes. Furthermore, and more interestingly, we find that food pairing in data is mostly due to trivial couplings of very similar ingredients, leading to a reconsideration of its current role in recipes, from being an already existing feature to a key to open up new scenarios in gastronomy. Our flavor-based recommender can thus leverage this novel concept and provide a new tool to lead culinary innovation.

Figures (6)

• Figure 1: Schematic representation of dataset and algorithm: (a) each recipe is connected to the ingredients it is made up of, and each ingredient to its flavor compounds. Collaborative filtering can work in two different ways. (b) exploits the cooccurrences of ingredients in recipes to recommend a new one for the chosen recipe; e.g. it suggests to add blackberries to a yogurt chocolate cake, as many ingredients in this cake are often used together with blackberries. (c) recommends an ingredient according to the flavor compounds it shares with the other ones in the recipe; e.g. buttermilk is recommended since it shares many flavor compounds with yogurt and butter, that are already in the recipe.
• Figure 2: (a) Mean Average Precision (mAP) and (b) Hit Ratio at 20 (HR@20) for all the methods used, averaged over 100 runs with an error bar equal to one standard deviation. The red point accounts for the performance of random recommendations.
• Figure 3: Same plots as in \ref{['recipe_based_comparison']} but computed with flavor compounds similarity.
• Figure 4: mAP for the recommenders built for each regional cuisine, reported with one standard deviation error bars, along with random recommender in red.
• Figure 5: Boxplot of the Z-Scores between the missing ingredient and the other ingredients in the recipe (first 20 ingredients only, recipes contain no more than 32 ingredients each), ranked by similarity, for the recipe-based (a) and the flavor compound-based (b) recommendations.