Abstract
As a vast number of ingredients exist in the culinary world, there are countless food ingredient pairings, but only a small number of pairings have been adopted by chefs and studied by food researchers. In this work, we propose KitcheNette which is a model that predicts food ingredient pairing scores and recommends optimal ingredient pairings. KitcheNette employs Siamese neural networks and is trained on our annotated dataset containing 300K scores of pairings generated from numerous ingredients in food recipes. As the results demonstrate, our model not only outperforms other baseline models but also can recommend complementary food pairings and discover novel ingredient pairings.
Introduction
The culinary world is vast, yet documented knowledge of food pairings remains narrow. Professional chefs rely on intuition, and food scientists often use expensive chemical analysis to find shared flavor compounds. However, these methods are either subjective or limited in scale. There is a massive "long tail" of ingredient combinations that have never been formally studied. KitcheNette addresses this by treating food pairing as a machine learning problem: by learning patterns from millions of existing recipes, the model can predict the success of entirely new pairings.
Methods
The KitcheNette model is built on two primary components: Siamese Neural Networks and a Wide & Deep architecture.
Siamese Neural Networks
To handle the symmetric nature of ingredient pairs (e.g., Gin & Tonic is the same as Tonic & Gin), the model uses two identical Multi-Layer Perceptrons (MLPs) with shared weights. These MLPs transform raw ingredient embeddings—pre-trained via word2vec—into a learned representation space.
Wide & Deep Architecture
The model combines two distinct pathways to capture different types of relationships:
- Deep Component: This part concatenates the learned representations and passes them through additional layers to capture complex, non-linear semantic relationships.
- Wide Component: This calculates the outer product of the two ingredient vectors, capturing direct, sparse interactions between specific features.
Datasets
The researchers introduced a new benchmark using:
- Recipe1M: A collection of ~1 million recipes.
- KitcheNette Dataset: A "silver standard" created using Normalized Point-wise Mutual Information (NPMI). The team analyzed 3,567 unique ingredients to find 300k "known" pairings (appearing at least 5 times). NPMI scores range from -1 to 1, where high scores indicate ingredients appearing together much more often than chance (e.g., wasabi & nori).
Key Results
KitcheNette significantly outperformed traditional machine learning baselines such as Random Forests (RF) and Gradient Boosting (GB). It achieved an $R^2$ of 0.8583 and a correlation of 0.9266, nearly doubling the performance of the best non-neural baseline.
The model also demonstrated strong qualitative performance, identifying "unknown" but culturally significant pairings like lime & nopales (Mexican) and gin & aquavit (Nordic) with high scores. Furthermore, KitcheNette’s recommendations were found to be more consistent with expert knowledge in The Flavor Bible than existing chemical-based databases.
Conclusion and Future Work
KitcheNette demonstrates that statistical co-occurrence in a massive recipe corpus is a highly effective proxy for culinary compatibility. While the current model focuses on pairwise interactions, future work will explore Graph Neural Networks to model multi-ingredient relationships and incorporate molecular data to further refine ingredient embeddings.
Summary updated