Data-driven recipe completion using machine learning methods

Background: Completing recipes is a non-trivial task, as the success of ingredient combinations depends on a multitude of factors such as taste, smell and texture. Scope and approach: In this article, we illustrate that machine learning methods can be applied for this purpose. Non-negative matrix factorization and two-step regularized least squares are presented as two alternative methods and their ability to build models to complete recipes is evaluated. The former method exploits information captured in existing recipes to complete a recipe, while the latter one is able to also incorporate information on flavor profiles of ingredients. The performance of the resulting models is evaluated on real-life data. Key findings and conclusions: The two machine learning methods can be used to build models to complete a recipe. Both models are able to retrieve an eliminated ingredient from a recipe and the two-step RLS model is also capable of completing an ingredient set to create a complete recipe. By applying machine learning methods on existing recipes, it is not necessary to model the complexity of good ingredient combinations to be able to complete a recipe. (C) 2015 Elsevier Ltd. All rights reserved.