Food based dietary guidelines are insuficiently followed by consumers. One of the principal explanations of this failure is that they are too general and do not take into account eating habits. Providing personalized dietary recommendations via nutrition recommender system can hence help people improve their eating habits. Understanding eating habits is a keystone in order to build a context aware recommender system that delivers personalized dietary recommendations. As a first step towards this goal, we explore food relationships on real-world data using the INCA 2 dataset, a French consumption survey. We particularly focus on extracting food substitutions, i.e food items that can replace each other. We consider that two food items can be substituted if they are consumed during similar contexts. We define the context in the nutrition field and we introduce a measure of substitutability between food items based on consumption data that encodes the context.
• Information systems → Information extraction; Recommender systems;
Nutritional quality of diets is proven to be an important factor in
health dysfunctions. The risk of developing modern chronic
diseases such as cardiovascular diseases, obesity or diabetes is linked
to unhealthy eating habits [
In order to promote healthy and sustainable diet and prevent
chronic diseases, dietary guidelines targeted to the general
population are produced by public health agencies. However, it has
been noted that the compliance to the guidelines is usually low
although the awareness concerning the food based dietary
recommendations is rather good [
One fair assumption is that people are more likely to follow recommendations if these are acceptable from their point of view. We hypothesize that the user acceptance is a prerequisite for the compliance and could be improved by producing user-tailored recommendations that take into account dietary habits. On the long term, our objective is to build a nutrition recommender system taking into account dietary habits in order to encourage people toward healthier alternatives with high compliance.
In food related recommender systems, the recommended items
are recipes [
Moreover, it has been shown that context-aware recommender
systems produce better recommendations than recommender
systems that do not take into account the context [
In this paper, we specifically investigate food substitutability. To do that, we define the concept of dietary context as the set of food items a food is consumed with and the concept of food intake context as the setting of food consumption. Our intuition is that two food items are substitutable if they are consumed in similar dietary contexts and that substitutability difers according to the food intake context.
The rest of the paper is organized as follows. Section 2 describes our methodology. Section 3 reports the results. Finally in section 4, we discuss our results and present our future perspectives. 2 2.1
Let X be the set of food items. A meal is a collection of food items consumed at the same timeframe. For instance, {cofee, bread, jam, juice} is a meal. The meal database DB is the set of all meals. Let us denote DBbr eak f ast the database of breakfasts and DBlunch the database of lunches.
Our objective is to mine food pair substitutability applied by consumers when they compose their meals. Given a database of meals, we want to extract substitutability relationships based on the way people consume food. No nutritional information is used during this process. Instead, contextual information is used in order to extract meaningful substitutability relationships. 2.2
The notion of context is quite complex and dificult to define universally. In the field of recommender systems, the context is usually defined according to the field of application of the system.
In the nutrition field, we define two types of contexts: the dietary context and the food intake context. The dietary context of a food item x is the set of food items c with which x is consumed. For instance, in the meal {cofee, bread, jam, juice} , the dietary context of {cofee} is {bread, jam, juice}. We think that the dietary context is fundamental when seeking substitutability of food items because the way people compose their meals is intrinsically dependent on the relationships between the items.
The food intake context is defined as the set of all variables
such as the type of the meal (breakfast, lunch, dinner, snack), the
location (home, workplace, restaurant), the participants (family,
friend, coworkers, alone). This corresponds to the notion of context
usually used in context-aware recommender systems [
There are three paradigms for incorporating context in
recommender systems : contextual pre-filtering, contextual post-filtering
and contextual modelling [
Our objective is to investigate substitutability among food items based on the assumption that two food items are highly substitutable if they are consumed in similar dietary contexts and in the same intake context.
Investigating all possible dietary contexts of a food item is computationally expensive because the number of possible dietary context is exponential in the number of food items and the length of the dietary context. The number of interesting contexts is actually limited by the characteristics of the available data. Instead of investigating all the dietary contexts of a food item, we decided to explore collections of meals that difer only by one item. We define the dietary context of a meal database c as the intersection of a set of meals Sm such that : len(c) = max (len(x )) − 1 x ∈Sm (1) Let us define the substitutable set Sc associated to a dietary context c as the set of food items such that the context c plus one item of Sc can be efectively consumed together. For instance, the substitutable set of the dietary context c = {bread, jam, juice } might be Sc = {co f f ee, tea, yoдurt }. 2.3
To eficiently retrieve interesting sets of dietary contexts and their substitutable set, in this paper, we propose an approach based on graph mining techniques. Let us denote the meal graph G = (V , E) where V is the set of nodes representing meals from the database and E is the set of edges such that two nodes are connected if there is at most one item that changes between them. A meal should appear at least once in the database in order to appear as a node in the graph. Figure 1 is a simple illustration of a meal network.
Designed in this way, the nodes of the substitutable set of a
dietary context are adjacent. They form a sub-graph that is completely
connected. Such an object is called a clique in graph mining. More
specifically, the nodes form a maximal clique. A maximal clique
is a clique to which another node cannot be added. In our setting,
discovering substitutable sets is similar to mining maximal cliques
in a graph. In this paper we use the algorithm of Bron-Kerbosh [
All discovered maximal cliques are not cliques that are interesting for our study. We want cliques such that the size of the intersection of the nodes is a dietary context as defined above. We denote these cliques as substitutable cliques. However, we may encounter cliques as in Figure 2. In this case, the intersection of the nodes is {A} and we cannot derive a substitutable set from this clique.
To avoid retrieving uninteresting cliques, we apply Algorithm 1 that filters out substitutable cliques.
ABC ABD
AED
For instance, when we apply our algorithm to the example of Figure 1, we get that this graph is a maximal clique and a substitutable clique more particularly. The context is {bread, butter} and the substitutable set associated to this context is {cofee, tea, milk, jam, nothing}. In this particular case, it is possible to substitute an item by nothing because {bread, butter } can be consumed as such. 2.4
Substitutability is not a binary relationship because there are diferent degrees of substitutability. Moreover, if two items are consumed together, they are less substitutable because they might be associated. Therefore, we need a function to quantify the relationship of substitutability that incorporates the possibility of associativity. Our hypothesis is that two items are highly substitutable if they are consumed in similar dietary contexts.
We want to compute a substitutability score such as : (1) Two items are highly substitutable if they are consumed in similar contexts. (2) Two items are less substitutable if they are consumed together. (3) Substitutability is a symmetrical relationship.
Let us denote, for an item x , the context set Cx as the set of dietary contexts in which x is a substitutable item. If the cardinality of Cx denoted as |Cx | is high, then x is substitutable in many dietary contexts.
For two items x and y, the condition (1) is described by the intersection of Cx and Cy . If |Cx ∩ Cy | is high, then x and y are consumed in similar contexts.
We denote Ax :y the set of contexts of x where y appears :
Ax :y = {c ∈ Cx |y ∈ c } The cardinality of Ax :y denotes how y is associated to x .
Taking into account these considerations, we propose the
substitutability score inspired by the Jaccard index [
|Cx ∩ Cy | |Cx ∪ Cy | + |Ax :y | + |Ay:x | The score equals 1 when x and y appear in exactly the same contexts and Ax :y = Ay:x = ∅. If x and y are never consumed in the same (2) (3) context then the score equals 0. The higher |Ax :y | + |Ay:x | is, the higher the association of x and y is and the lower the score is. The French dataset INCA 21 is the result of a survey conducted during 2006-2007 about individual food consumption. Individual 7-day food diaries are reported for 2624 adults and 1455 children over several months taking into account possible seasonality in eating habits. A day is composed of three main meals : breakfast, lunch and dinner. The moments in between are denoted as snacking. For the main meals, the location (home, work, school, outdoor) and the companion (family, friends, coworkers, alone) are registered.
The 1280 food entries are organized in 44 groups and 110 subgroups of food items. We chose to consider the medium level of hierarchy in order to capture substitution relationships inter-groups and intra-groups.
Only adults are considered in this paper. All meals are gathered
in a meal database DBmeals regardless of the type of meal. The
database can be split according to contextual information in order
to get better results [
Applying our algorithm on DBbr eak f ast yields 2368 contexts. Some of these and their substitutable sets are given in Table 1. Our results are coherent. For example, either bread, rusk or viennoiserie can be consumed for breakfast with cofee, sugar and water.
Context Substitutable set
bread cofee, sugar, water, butter rusk viennoiserie yogurt sugar tea/infusions, donuts jam/honey nothing Table 1: Results of context and substitutable set retrieval for breakfasts
We applied our algorithm to the three datasets. The results are reported in Table 2. We can see that we can obtain inter-group substitutions such as {potatoes ⇒ green beans} but also intra-group substitutions as {bread ⇒ rusk}.
The substitutions proposed are consistent with regards to eating habits. Substitutes of drinks are also drinks : the substitutes of cofee are tea, cocoa and chicory. It is also the case for spreadable food items : the substitutes for butter for breakfast are spreadable items. No semantic information describing how a food item can be eaten is available in the dataset and yet considering the dietary context helps us retrieving this kind of information.
Substitutions between food items of the same nutritional food groups are found. For instance, the substitutes for potatoes are pasta and rice: they all contain starches. 1https://www.data.gouv.fr/fr/datasets/donnees-de-consommations-et-habitudesalimentaires-de-letude-inca-2-3/
We proposed a score of substitutability based on consumption data that can be used in a recommender system together with other scores such as a nutritional score that takes into account the nutritional contribution of the substitution and a user preference score. The substitutability score is based on the assumption that two items are substitutable if they are consumed in similar contexts. Preliminary results on the INCA2 dataset show that this assumption helps retrieving substitutability relationships based on consumption data.
When we split the dataset according to the contextual variable "type of meal", the substitutes and the scores are diferent. Cofee can be substituted by tea, chicory and cofee for breakfast whereas for lunch, it can be substituted by sodas, yogurt and fruits. Food items are consumed diferently according to the type of meal. The relationship of substitutability is therefore diferent too.
Diference of scale in scores is noted according to the type of meal. It may be due to the fact that the diversity of food items consumed during lunch is higher than during breakfast. A rescaling factor based on the diversity of the type of meal can be introduced. As future work we plan to investigate this aspect and implement the nutritional score and the user preference related score.
This study was funded by Danone Nutricia Research.