Discovering novel ingredient pairings in molecular gastronomy using network analysis

Molecular gastronomy is a distinct sub-discipline of food science that takes an active role in examining chemical and physical properties of ingredients and as such lends itself to more scientific approaches to finding novel ingredient pairings. With thousands of ingredients and molecules, which participate in the creation of each ingredient’s flavour, it can be difficult to find compatible flavours in an efficient manner. Existing literature is focused mainly on analysis of already established cuisine based on the flavour profile of its ingredients, but fails to consider the potential in finding flavour compatibility for use in creation of completely new recipes. Expressing relationships between ingredients and their molecular structure as a bipartite network opens up this problem to effective analysis with methods from network science. We describe a series of experiments on a database of food using network analysis, which produce a set of compatible ingredients that can be used in creation of new recipes. We expect this approach and its results to dramatically simplify the creation of new recipes with previously unseen and fresh combinations of ingredients. Introduction – Essential part of any great dish is harnessing compatible flavours from its ingredients. This includes, but is not limited to, the knowledge of which spices, herbs and other flavourings accentuate particular ingredients best. The tried and tested method of trial and error has been the go to method for finding such compatibility throughout the history. While this approach produced numerous timeless combinations—such as basil with tomatoes and mozzarella cheese or apples with cinnamon—and over time resulted in classic cuisines, its one flaw is that it depends on the imagination of the chef and his willingness to try as many random combinations as possible. There are many possibilities that make for good combinations, but would probably never get tried due to how different they seem—white chocolate with caviar, oysters with passion fruit, etc. Today, in a modern kitchen, it is possible to utilize the scientific method in finding compatible flavours without relying on a single person’s taste. Once we discovered enough about volatile compounds of ingredients, it became clear that flavour compatibility is based on molecular similarity of different ingredients [2, 10]. A single ingredient can be composed of hundreds, sometimes even more than a thousand flavour compounds. This makes it difficult to efficiently analyse and compare large numbers of ingredients and limits the creative landscape of chefs looking to invent new recipes. One approach that proved successful in analysis and comparison of ingredients in network analysis is construction of a flavour network [1]. With this approach ingredients and volatile compounds are presented as nodes in a graph, while edges connect ingredients with compounds they contain. We use this approach as basis for our work. We expand on it by looking for groups of complimentary ingredients by joining nodes that have similar molecular profiles into clusters [13]. Furthermore, we use partially labelled data in conjunction with semi supervised community detection [15] to account for nuances that can’t be described by shared molecular profile alone, to improve the accuracy of our algorithm and the method used in Ahn’s approach [1]. Since this field of food science is relatively unexplored, data on compatible flavour parings are scarce and limited. With our approach we show that finding complimentary pairings can be relatively easy to implement and execute on any food database containing enough data. Related work – One of the pioneers in finding novel ingredient pairings was H. Blumenthal, who popularized molecular gastronomy through his restaurant and recipe books [2]. Ahn et al. [1] uses the flavour network to compare and contrast Western and Eastern cuisine and to see whether we prefer to use ingredients that share more p-1 ar X iv :1 60 2. 03 71 9v 1 [ cs .S I] 1 1 Fe b 20 16 A. Ključevšek, L. Krapić flavour compounds. Their findings confirm their hypothesis, but only for Western cuisine, while Eastern cuisine displays preference for ingredients with different flavour base. An extension of this technique is further used to analyse specific local cuisines. Bogojeska et al. [3] analyse Macedonian cuisine and its flavours, while Jain et al. [7] seek to determine whether cuisines of different regions of India follow the Eastern cuisine pattern discovered by Ahn et al.. Caporaso et al. [4] look closely at how using different types of vegetable oils affects chemical properties and volatile profile of a traditional Neapolitan pizza. Interactions between food and drugs has been well documented, but Jovanovik et al. [8] show how different drug categories have different distributions of negative effects in different parts of the world due to differences in regional cuisines. In a more narrow analysis, Ruiz et al. [11] look at molecular constitution as one of the methods in creating novel chocolates. Using Ahn’s approach [1], Tackx et al. [12] generalize it in trying to develop new metrics for studying intricate patterns observed in real networks, which standard metrics do not account for. Data – All data about food and food components comes from FooDB [13]. The ingredients and compounds from this database form two disjoint sets of a bipartite graph—a type of graph with two disjoint sets of nodes inside which no two nodes share a link—while edges link ingredients with compounds if the former contains the latter. Such representation allows us to efficiently express information about similarities of ingredients based on the number of shared components. The graph consists of 16,269 nodes, with 856 of these nodes representing ingredients, while the rest represent the flavour molecules and 106,113 edges. Degree distribution of this graph projected into ingredient space is represented in Figure 1. Most ingredients share at least one component, with average degree < k >= 767.533 making the graph very dense (densityis0.898) and in need of filtering, which is described in the next section. Fig. 1: Degree distribution of projected graph—ingredient network. Fig. 2: Degree distribution of ingredient network after filtration. In order to verify our findings, we use information about known compatible ingredient pairings in Western and Eastern cuisine. Data for Western cuisine is extracted from one of the most popular recipe websites Epicurious [5], while data for Eastern cuisine is extracted from Rasa Malaysia [6] in order to avoid Western interpretation of Eastern cuisine. We collect 1,000 highly rated recipes (rating 4/4) from Epicurious, which are chosen at random from all the different types of recipes, e.g., appetizers, lunches, dinners, sides, desserts, etc., and 507 recipes from Rasa Malaysia, also from various categories. Ingredients are then parsed and matched to ingredients in our database. Parsed recipes for Western cuisine contain on average 7.62 distinct ingredients, while recipes for Eastern cuisine contain on average 10.77 unique ingredients. In order to make both sets easier to compare, we reduce the size of the larger one to the size of the smaller one, so that they contain 507 recipes each. These data form the body of knowledge used for training and testing our classification algorithm. Methods – In order to get a more compact representation of our original bipartite graph, we project it into ingredient space. This projection produces a simple weighted graph where weights of edges correspond to number of shared components between two adjacent nodes. Resulting flavour network is essentially similar to the one constructed by Ahn et al. [1]. Our initial projected graph consists of 856 nodes and 328,504 links, and is dense to the point of a complete graph. Consequently some filtration to reduce this is necessary. The filtration process is done as follows: we define similarity of two nodes based on the weight of the link they share. Furthermore, we assume every node is similar to at least one of its neighbours. Filtration is done locally for every node. Filtering with factor p ∈ [0, 1] means removing all links below value p ∗wmax for a given node, where wmax is maximum link weight, i.e., the largest similarity it shares with any other ingredient.