Kissing Cuisines:

Exploring Worldwide Culinary Habits on the Web

Sina Sajadmanesh , Sina Jafarzadeh , Seyed Ali Ossia , Hamid R. Rabiee , Hamed Haddadi ,
Yelena Mejova? , Mirco Musolesi] , Emiliano De Cristofaro] , Gianluca Stringhini]

Sharif University of Technology

Queen Mary University of London

arXiv:1610.08469v1 [cs.CY] 26 Oct 2016

ABSTRACT

Qatar Computing Research Institute

UCL

this motivation in mind, in this paper, we set to investigate the way

in which ingredients relate to different cuisines and recipes, as well
as the geographic and health significances thereof. We use a few
datasets, including 157K recipes from over 200 cuisines crawled
from Yummly, BBC Food data, and country health statistics.
Overview & Contributions. First, we characterize different
cuisines around the world by their ingredients and flavors. Then,
we train a Support Vector Machine classifier and use deep learning
models to predict a cuisine from its ingredients. This also enables
us to discover the similarity across different cuisines based on their
ingredients e.g., Chinese and Japanese while, intuitively, they
might be considered different. We look at the diversity of ingredients in recipes from different countries and compare them to geographic and human migration statistics. We also measure the relationship between the nutrition value of the recipes vis--vis public
health statistics such as obesity and diabetes.
Paper Organization. The rest of the paper is organized as follows. In Section 2, we present the datasets used in our evaluation,
then Section 3 presents an analysis of the diversity of the ingredients around the world, looking at geographic diversity patterns of
cuisines and notable ingredients in particular ones. In Section 4,
we look at the similarity between the cuisines based on their ingredients and flavors, and use these results to train machine-learning
classifiers for ingredient-based cuisine prediction models in Section 5. In Section 6, we correlate the nutrition values of recipes for
different countries with their public health statistics. After reviewing related work in Section 7, the paper concludes in Section 8.

As food becomes an important part of modern life, recipes shared

on the web are a great indicator of civilizations and culinary attitudes in different countries. Similarly, ingredients, flavors, and
nutrition information are strong signals of the taste preferences of
individuals from various parts of the world. Yet, we do not have a
thorough understanding of these palate varieties.
In this paper, we present a large-scale study of recipes published
on the Web and their content, aiming to understand cuisines and
culinary habits around the world. Using a database of more than
157K recipes from over 200 different cuisines, we analyze ingredients, flavors, and nutritional values which distinguish dishes from
different regions, and use this knowledge to assess the predictability of recipes from different cuisines. We then use country health
statistics to understand the relation between these factors and health
indicators of different nations, such as obesity, diabetes, migration,
and health expenditure. Our results confirm the strong effects of
geographical and cultural similarities on recipes, health indicators,
and culinary preferences between countries around the world.

1.

INTRODUCTION

Nowadays, the importance of food and eating goes far beyond a

means to survive. Often regarded as a social construct, food has become an essential part of modern day life. New jargon has entered
our vocabulary as expressions like foodie, food porn, or food
tourism, hint at the buzz around the entertainment arising from
our culinary experiences. With the rise of social media, and the
proliferation of always-on always-connected devices, this gobbling
revolution is not confined to our kitchens, restaurants, and food
stalls, but naturally breaks out on the social web. Sharing pictures
of ones food has become a growing passion for both tourists and
locals [16], and dedicated food searching and sharing apps, along
with recipe websites and the ubiquitous social presence of celebrity
chefs, have all contributed to a thriving culture and passion around
food worldwide.
Around the world, different cuisines are naturally intertwined
with cultures, traditions, passions, and religion of individuals living in different countries and continents. Sushi, curry, kebab, pasta,
tacos these are just examples of foods conventionally associated
with specific countries, as are specific cuisines and ingredients.
Different dietary habits around the world are also closely related
to various health statistics, including cancer incidence [3], death
rates [13], cardiovascular complications [18], and obesity [14].
Although there are many common beliefs about cuisines,
recipes, and their ingredients, it is still unclear what types of ingredients are unique in/about different countries, what factors make
cuisines similar to each other (e.g., in terms of ingredients or flavors), and how these factors are related to individuals health. With

2.

DATASETS

Our study relies on a number of datasets, namely, a large set

of recipes collected from Yummly, a list of ingredients compiled
by BBC Food, and country health statistics. In this section, we
describe these datasets in detail.

2.1

Yummly data

Yummly is a website offering recipe recommendations based on

the users taste.1 It allows users to search for recipes, learning
which dishes the user likes and providing them with recipe suggestions. It also provides a user-friendly API, which we use to collect
recipes. First, we crawled Wikipedia for a list of cuisines2 , then, in
Summer 2016, we queried the Yummly API for recipes belonging
to each cuisine. In the end, we obtained 157,013 recipes belonging
to over 200 different cuisines. Due to API restrictions, we limited
the number of recipes per to 5,000.
Each recipe obtained from the Yummly API contains a number
of attributes. In our study, we use the following:
1
2

http://www.yummly.com
https://en.wikipedia.org/wiki/List_of_cuisines

1. Ingredients: Each recipe contains a list of the ingredients

that are required to prepare it. Since Yummly acts as a recipe
aggregator from various cooking sites, the ingredients do not
always appear with the same wording. In fact, it is very
common to see the same ingredient written with different
spellings or by using a different terminology. We overcome
these issues through a standardization process described in
Section 2.2.

we will use the diabetes prevalence estimates from World Development Indicators by The World Bank4 , the health expenditure as
a percentage of total GDP from The World Bank5 , and the obesity
prevalence from the World Health Organization6 in the countries
to which the cuisines are mapped, using the most recent available
data, which is from 2014.

2. Flavors: Recipes are identified by six flavors, specifically,

saltiness, sourness, sweetness, bitterness, savoriness, and
spiciness. These scores are on a range of 0 to 1.

In this section, we provide a characterization of the ingredients

used in dishes from all over the world. First, we investigate the
diversity of ingredients in different countries. Next, we define the
concept of complexity of a dish in terms of its ingredients and
look at how complexity changes around the world. Finally, we discuss a series of case studies of most notable and significant ingredients in some eminent cuisines.

3.

3. Rating: Users are encouraged to provide a rating, from 1 to

5, for the recipes that they try. From the Yummly API, we
retrieve the average review rating for each recipe.
4. Nutrition: Unfortunately, the Yummly search API does not
directly provide nutritional information for the recipes. As a
consequence, we designed a simple web crawler to fetch the
corresponding web page for each recipe in our dataset, and
extract information on the amount of protein, fat, saturated
fat, sodium, fiber, sugar, and carbohydrate of a recipe (per
serving), as well as calories.

3.1

1. How many different unique ingredients are used in total in

dishes of each country? In other words, what is the number
of unique ingredients the people of a country have ever used
to prepare a culinary dish? The answer to this question is
what we refer to as the global diversity.
2. How different are the dishes of an individual country relative
together in terms of their ingredients combination? In other
words, do different dishes usually share some ingredients or
their ingredients are almost different? The answer to this
question is what we call local diversity.
The local and global diversity of ingredients in a country depend on many parameters including the geographical location, climatic conditions, agricultural situation, or even the amount of immigration which directly influences the diversity of culinary cultures. The calculation of the global diversity is performed in two
steps. Since the number of recipes per different cuisines are variable, we first set a fixed number of 100 recipes per cuisine, discarding cuisines containing fewer number of recipes, and sampling
from cuisines containing more number of recipes uniformly at random, to have an equal number of recipes in all cuisines. This results
in a final set of 82 different cuisines each containing 100 recipes.
We then map the result obtained for each cuisine to its corresponding country. Here, some countries are mapped with more than one
cuisine. For such countries, we record the average result over their
associated cuisines.
To calculate the local diversity, we look at each cuisine as a probability distribution over all standard ingredients. By counting the
total number of occurrences of each ingredient in all recipes of a
particular cuisine, and then normalizing the values such that they
sum to one, we obtain the ingredient distribution for that cuisine.
We then calculate the entropy of these distributions as the local
diversity of their corresponding cuisines. The entropy of the ingredient distribution measures the unpredictability of ingredients
used in the dishes. Therefore, the higher the entropy of the ingredient distribution of a particular cuisine, the more different the
ingredients combination of its recipes, and thus the higher the local
diversity. To preserve the smoothness of the ingredient distributions, we again keep the 82 cuisines with more than 100 recipes.
After calculating the local diversity for each cuisine, we follow the

BBC Food Data

BBC Food3 is a part of the BBC website providing information

about recipes, ingredients, chefs, cuisines, and other information
related to cooking and dishes from all BBC programs. In Summer
2016, we crawled all the ingredients from the BBC Food website,
collecting about 1,000 ingredients, which we used to organize and
standardize the ingredients in the Yummly dataset. The standardization process is as follows:
(i) We extracted all the 11,000 ingredients from the Yummly
dataset and performed a preliminary data cleaning, i.e., removing measurement units (mass, volume, etc), numbers, punctuation
marks, and other symbols.
(ii) Due to the multilingualism of the Yummly data, we used the
Google Translate API to perform automatic language detection and
translation of all the Yummly ingredients to English.
(iii) We used the BBC list of ingredients as a reference, and
mapped all possible ingredients from the Yummly list to it.
(iv) As not all ingredients from the Yummly list were successfully mapped, we merged the similar ones into groups, and the ingredients in each group were manually mapped to the its representative ingredient.
Overall, this process yields about 3,000 standardized ingredients.

2.3

Country health statistics

As diet is directly related to the health of individuals, we also set

to relate Yummly statistics to real-world health data. To this end,
3

Diversity of ingredients

Aiming to investigate the diversity of ingredients in dishes of a

cuisine, we set to answer the following questions:

Although some ingredients appear in other languages (e.g., German, French, etc), the recipes presented here are mostly in English;
hence it is possible that some more authentic or niche local recipes
might be missing from our dataset. However, considering the number of recipes and a large cut-off threshold introduced later on, we
are confident this does not significantly affect our analysis. Moreover, authors of the recipes might not represent the entire population, given the fact that they are likely to be tech-savvy. This might
introduce a potential bias in the dataset, but at the same time, this
potential issue is compensated by its richness in terms of the variety of dishes from different countries available in it as structured
information.

2.2

INGREDIENTS AROUND THE WORLD

http://data.worldbank.org/indicator/SH.STA.DIAB.ZS
http://data.worldbank.org/indicator/SH.XPD.TOTL.ZS
6
http://apps.who.int/gho/data/view.main.2450A
5

http://www.bbc.co.uk/food/

180

195

210

225

240

255

270

285

4.20

4.35

4.50

(a) Global diversity

4.65

4.80

4.95

5.10

5.25

(b) Local diversity

Figure 1: Diversity of ingredients around the world

104

same procedure as for the global diversity to map the cuisine-based

results to countries.
Figure 1 shows the local diversity and the global diversity of
ingredients for different countries around the world. According to
the figure, we can see that the local and global diversities have a
meaningful correlation with each other. The countries with high
global diversity have also high local diversity, and countries with
low global diversity have low local diversity as well. This happens
because as the global diversity increases, people will have more
options to choose as the ingredients for their foods, so they can
prepare relatively different dishes.
Another interesting trend observable in Figure 1 is that countries like the United States and Australia which usually accept a
high number of immigrants, have a relatively high global diversity.
Regarding this, we hypothesized that the number of immigrants
coming to a country must have an influence on the ingredient diversity of that country. This is mainly due to immigrants bringing
their native culinary culture with themselves, which in turn makes
the cuisines of their target country richer. To investigate this fact,
we collected the net migration data from the World Bank7 which
shows the difference between the total number of immigrants and
emigrants during a time period. We correlated the global diversity
with the average net migration from 1960 to 2016. To this end, we
fitted a polynomial curve to the data points considering the global
diversity and the net migration. The result is shown in Figure 2. As
we expect, the figure indicates that an increase in the net migration
would result in an increase in the global diversity of ingredients.

2.00

Net Migration

0.00

2.00

4.00

6.00

8.00

160

200

220

240

260

280

300

Figure 2: Relation between the global ingredient diversity and the net migration
1.00

0.80

Complexity of dishes

Another interesting concept about the culinary preferences of

different countries is the complexity of dishes. The complexity of a
dish is simply the number of unique ingredients required to prepare
it. Accordingly, a cuisine is more complex than another one if its
dishes are proportionally more complex than another ones.
Formally speaking, each cuisine is associated with the complexity distribution of its dishes. For a sample cuisine, this distribution,
namely P (X = i), specifies the probability of a dish from that
cuisine to have exactly i unique ingredients. This way, the cumulative complexity distribution (CCD) will give us an insight about
the complexity of dishes in a particular cuisine.
Figure 3 shows the cumulative complexity distribution (CCD)
for Norwegian, Russian, Spanish, Tunisian, and Lao as an example. From the figure we can see that the CCD for Norwegian cuisine grows faster than the others, while for Lao, it is relatively
7

180

Global Diversity

Cumulative Probability

3.2

4.00

0.60

0.40
Norwegian
Russian
0.20

Spanish
Tunisian
Lao

0.00
0

10

15

20

25

30

# ingredients
Figure 3: Cumulative complexity distribution for a number of sample
cuisines.

slower. This means that Lao dishes are more complex than Norwegian ones. Thus for each cuisine, the area under its CCD is inversely related to its complexity. Hence, we use the reciprocal of

http://data.worldbank.org/indicator/SM.POP.NETM

4.1

Ingredient-based similarities

At first, we calculate the similarity between different cuisines

based on the ingredients used in their recipes. To this end, we
convert cuisines into vector space, representing each cuisine as a
vector where each element indicates the frequency of an specific
ingredient in that cuisine. Thereby, for each cuisine we obtain an
ingredient-based feature vector which we leverage to calculate the
similarity between different cuisines using the following metrics:
1. Jensen-Shannon divergence: If we normalize each
ingredient-based feature vector such that the elements of a
vector sum to one, then each vector will represent a probability distribution over standard ingredients. This way, we can
use the distance measures proposed for probability distributions like Jensen-Shannon divergence. The Jensen-Shannon
divergence between two probability distributions P and Q is
defined as:

0.0146 0.0148 0.0150 0.0152 0.0154 0.0156 0.0158 0.0160

Figure 4: Complexity of dishes around the world

the area under CCD as a measure of complexity for a cuisine.

Figure 4 shows the complexity of dishes for different countries
around the world. Here we have used the same approach as in
Section 3.1 to map the cuisines to countries. We can see from the
figure that except for some cases, the complexities are consistent
with the diversities. This is due to the fact that as the number of
available ingredients increases in a country (which is the result of
global diversity,) people can leverage more ingredients and prepare
more complex dishes. The exceptions here are China and India, two
countries with the most population in the world. the complexity of
dishes in these countries are relatively high, while their ingredients
diversity is low. This can be the result of overpopulation or special
culinary culture in these countries. Perhaps, these countries had
or have good chefs that could cook more complex foods with the
available ingredients!

3.3

JS(P, Q) =

2. TF-IDF similarity: Another measure we use to calculate

the similarity between two cuisines is the well-known TFIDF. Using this measure, we followed the approach described
in section 3.3 to calculate the weight of ingredients in each
cuisine. This way, each cuisine is represented as a vector
where each of its elements indicates the representative power
of the corresponding ingredient for that cuisine. The TFIDF similarity between two cuisines is then simply the cosine
similarity between their associated TF-IDF vectors.

Notable ingredients

Using the above similarity metrics, we calculated all similarities between each pair of cuisines. To assert the smoothness of
ingredient distributions needed to compute Jensen-Shannon divergence, we limited our cuisines to those 82 ones having more than
100 recipes. Figure 6 illustrates the results for different similarity
measures in a graph-based fashion. In these graphs, each node represents a cuisine and each cuisine is linked to its top-5 most similar
cuisines. Link weights are proportional to the obtained similarity
score between two endpoints. We colored each cuisine node according to the geographical region it resides in. The cuisines are
classified into 9 regions which are North America, Latin America, Africa, Western Europe, Eastern Europe, Middle East, South
Asia, East Asia, and Oceania. To visualize the graph, we have
used ForceAtlas graph drawing algorithm implemented in Gephi
tool [4]. This a force-directed algorithm which makes densely connected nodes to be grouped together [19, 24] and thus the communities become revealed.
We can see from the Figure 6 that using either of the ingredientbased similarity measures, the cuisines which reside in the same region are more similar to themselves and thus are grouped together.
For example, we clearly see the clusters formed by Eastern and

CUISINE SIMILARITY

How do we determine the similarities between Korean and

Japanese cuisine? What makes the Middle Eastern dishes seem
similar to one another? In this section we answer these questions
using a number of different methods and data.
8

(1)

where M = 12 (P + Q) and KL(P k M ) is the KullbackLeibler divergence from M to P . Since the Jensen-Shannon
divergence is a distance measure between 0 and 1, we take
1JS(P, Q) as the similarity measure between two cuisines
with their associated ingredient distributions P and Q. We
have used Jensen-Shannon divergence instead of the simpler
Kullback-Leibler divergence because KL(P k Q) goes to
infinity when for an ingredient like i, P (i) is non-zero while
Q(i) is. This case almost always happens in our data due
to the geographical locality of ingredients. Therefore, we
turned to Jensen-Shannon divergence which does not have
this drawback.

Specific cuisines are mostly associated with different sets of ingredients due to the geographical locality of the ingredients. In this
section, we study the most notable ingredient associated to some
well-known cuisines using our dataset of recipes from Yummly.
We use the Term Frequency - Inverse Document Frequency (TFIDF) calculation to find notable ingredients in each cuisine. In this
approach, each ingredient is considered as an atomic word, and the
collection of all the ingredients appeared within a cuisine is considered as a document. A TF-IDF calculation leads us to find the
weight of different ingredients in the corpus of documents. This
way, the importance of each ingredient within each cuisine is specified.
Figure 5 shows the top-50 most notable ingredients for Italian,
Indian, and Mexican cuisines as a case study. We also looked at
other similar cases, which we do not present here due to space constraints. According to the figure, the bigger the name of an ingredient, the more distinctive it is in its associated cuisine. We verified
the soundness of results using Google Trends8 service. For example, the term Mozzarella has the highest search frequency in
Italy. Similarly, Garam masala is the most popular food additive
in India according to its search volume.

4.

1
[KL(P k M ) + KL(Q k M )]
2

https://www.google.com/trends

(a) Italian

(b) Indian

(c) Mexican

Figure 5: Notable ingredients in three famous cuisines.

Southern Asian, Middle Eastern and African, Latin American, and

Western European cuisines. Furthermore, due to the similarity of
cultures in Europe and North America, and even Oceania, it can bee
seen that clusters formed by the cuisines of these regions greatly
overlap with each other.

4.2

Chinese

Indonesian
Korean

Asian
Japanese

Thai

Vietnamese
Malaysian

Hungarian
Spanish
Lao
Cantonese
Louisiana Mexican
Italian-American
Chilean
Peruvian
Mongolian
Latin-American
Indian
Cajun
Cuban
Mediterranean
Swiss
Tunisian
Caribbean
Jamaican
American
Philippine
Ethiopian
Berber
Canadian
African
Portuguese
Italian
Israeli Cornish
Australian
BasqueBrazilian
Bulgarian BritishMoroccan
English
Finnish
Ukrainian
Sicilian
Hong
Irish
Colombian
Oceanic
Belgian
Austrian
Norwegian
Lebanese ArabArmenian
Greek Dutch
German
Scottish
Croatian
Swedish
Egyptian
Iranian
Romanian
Danish
Turkish Polish
Russian
French
Welsh

Flavor-based similarities

In addition to the ingredient-based similarity, we calculate the

similarity between cuisines in terms of the flavors provided in their
recipes. This can help us understand how different cuisines are
related to each other based on the taste of their dishes. As we
mentioned in section 2.1, each recipe contains the flavor scores for
six different flavors including saltiness, sourness, sweetness, bitterness, savoriness, and spiciness. To calculate the similarity between
cuisines based on these flavors, like what we did for ingredientbased similarity, we consider each cuisine as a distribution over
different flavors. Regarding the fact that different flavors of a recipe
are correlated to each other for instance, a dish can hardly be both
sweet and spicy simultaneously and due to the continuity of flavor scores, we hypothesize that the flavor scores are sampled from
a multivariate Gaussian distribution, where each covariate corresponds a particular flavor. Considering this assumption, we fit a
multivariate Gaussian distribution to each cuisine so that each cuisine become associated by a mean vector representing the average
of flavor scores over all of its recipes, and a covariance matrix representing how flavors change relative to each other within that cuisine.
After fitting a multivariate Gaussian distribution to each cuisine
using maximum likelihood estimation, we use Kullback-Leibler
divergence to measure the distance between the distributions associated to each pair of cuisines. Since Kullback-Leibler divergence is and asymmetric distance measure, for each pair of
cuisines with P and Q as their corresponding flavor distributions,
1

as a symmetric simiwe use 21 (KL(P k Q) + KL(Q k P ))
larity measure between them.
Figure 7 shows the result of flavor-based similarity between different cuisines in a graph-based manner. We followed exactly the
same steps as in Figure 6 to draw the graph, except that we used
flavor-based similarity between cuisines. We can see from the figure that even though the flavors are not as much discriminant as ingredients, still we can observe some geographical patterns. For instance the clusters formed by Eastern Asian, Middle Eastern, Latin
American, and Northern European cuisines are clear in this case
as well. But what is obvious here is the fact that although there
is a sense of taste similarity between the dishes from neighboring
countries, the flavors are naturally shared all over the world.

Figure 7: Graph of flavor-based similarity between different cuisines. Regional colors are the same as in Figure 6.

SVM

DNN

SVM

0.75

0.75

0.70

0.70

0.65

0.65

0.60

0.60

0.55

0.55

0.50

0.50

0.45

0.45

0.40

DNN

0.40
Acc

F1

(a) Cuisine Prediction

Acc

F1

(b) Region Prediction

Figure 8: The prediction performance of different methods under different

settings.

5.

CUISINE PREDICTION

In this section we address the question of How good we can

predict a recipes cuisine, given its ingredients?. To answer this
question, We use two different classifiers, Support Vector Machine
5

North America
Africa

Latin America
South Asia

Western Europe
East Asia

Iranian
Lebanese

French

German Austrian
Scottish
Danish
Swedish
Finnish
Norwegian

Ethiopian
Berber

Cambodian
Lao

Japanese

Middle East

Welsh
Chinese
CantoneseMongolian
Latin-American
Hong
Mexican
Italian-American
Asian
Korean
Aztec Louisiana
Cuban
Taiwanese
Cajun
Japanese
Spanish
American
Colombian
Caribbean Basque Italian
Portuguese
Canadian
Vietnamese
PeruvianJamaican
Thai
Chilean
Saint-Lucian
French
BrazilianSwiss
Cambodian
Dutch
Malaysian
Lao Philippine
Sicilian Belgian
Russian
Indonesian
PolishGerman
Irish Oceanic
Maltese
Cornish
Bengali
Croatian Hungarian
Danish
Punjabi
English
Ukrainian
Sri-Lankan
Indian
Austrian
Australian British

Arab
Italian
Mediterranean
Armenian
Sicilian
Italian-American
Turkish
Syrian Egyptian
Greek
Swiss
Israeli
Hungarian
Cajun
Australian
Bulgarian
Basque
Louisiana
Moroccan
Tunisian
Cornish
British
Canadian
Spanish
Polish
Portuguese
Ethiopian
Berber
Croatian
African
Dutch
American
RomanianChilean
Icelandic
Indian
Jamaican
Peruvian
Punjabi
Saint-Lucian
Colombian
UkrainianWelsh
Bengali
Oceanic
Mexican
Caribbean
Russian
Cuban
Aztec
Sri-Lankan
BelgianEnglish
Brazilian
Latin-American
Irish
Maltese

Eastern Europe
Oceania

Malaysian
Vietnamese
Korean Philippine
Thai
Taiwanese Mongolian
Indonesian
Hong

Icelandic
African
TunisianRomanian
Mediterranean
Swedish
Scottish
Israeli
Moroccan
Norwegian
Armenian
BulgarianGreek
Finnish
Arab
Egyptian
Syrian Lebanese

Asian
Cantonese
Chinese

Turkish Iranian

(a) Jensen-Shannon divergence

(b) TF-IDF similarity

Figure 6: Graph of ingredient-based similarity between cuisines with different similarity measures.

can see in the figure, the DNN model performs about 24% better
than SVM for cuisine prediction task under accuracy and over 13%
better under F-measure. For region prediction task, since the number of classes are much fewer than cuisine prediction, both methods performed relatively better. In this case, the accuracy and Fmeasure achieved by the DNN model is about 12% and 9% better
relative those achieved by SVM, respectively.
In order to get more intuition about the similarity of recipes in
different regions, we bring the confusion matrix of the DNN model
for region predictions in Table 1. Each region name is abbreviated
in two letters. For example, LA means Latin American and AF
means African cuisines. The number of correctly classified recipes
are shown in bold and for each class, the greatest number of missclassifications is shown in red. This table clearly demonstrates that
the almost all of the miss-classifications are Western European.
This is probably due to the huge ethnic composition of Western
European countries which resulted in the diversity of culinary cultures of that region. The table shows that for some regions like
Southern and Eastern Asian, the number of miss-classified recipes
are somewhat low relative to the correctly classified ones. This result is as same as in Figure 6 in which these regions were almost
disconnected from the others. On the other hand, for some cuisines
like Oceanic, Eastern European, and Northern America, the number of miss-classifications are relatively high, mostly with Western
European. Figure 6 shows clearly that the cultures in these regions
are very similar to each other, mainly due to the common ethnics
and history.

(SVM), which is previously used in [22] for the same task, and
Deep Neural Network (DNN), which is popular nowadays for classification purposes. To extract a feature vector for each recipe, we
convert it into a boolean bag of words vector, considering each ingredient as an atomic word. Therefore, each recipe is represented
as a vector with a length equal to the total number of ingredients,
which is 3,286. The labeling of recipes are performed according to
the following settings:
1. Cuisine Prediction: Each recipe is labeled to its cuisine. we
consider 82 different cuisines having more than 100 recipes
as different classes, resulting in about 100K recipes.
2. Region Prediction: Each recipe is labeled according to one
of the 9 geographical regions where its cuisine belongs to.
The regions are considered the same as in section 4. This
results to have about 157K recipes.
For multi-class classification with SVM, we use linear kernel
with one vs. rest coding. The class imbalance problem is resolved
with adjusting the weight of each cuisine inversely proportional
to its frequency. The implementation is done using Scikit-learn
machine learning library in python [5]. For DNN, we use Keras
deep learning library [7] and create four dense hidden layers and
a softmax output layer. Each of the first two hidden layers consists of 1000 neurons, and the two last ones each have 500 neurons.
Dropout regularization [21] is used for all of the hidden layers. We
use Adadelta [26] with default parameters as the optimizer. For
both methods we take 80% of the data as training set and the remaining 20% as the test set. The prediction performance of both
methods are evaluated under accuracy and F-measure.
Figure 8 shows the obtained results with both SVM and DNN
methods. Figure 8a illustrates the results for cuisine prediction,
while Figure 8b shows the results of region prediction task. As we

6.

NUTRITION VALUES

In this section, we investigate the relation between the nutrition

values of the recipes associated with countries and their hard measures of health, including obesity rate, diabetes rate and health ex6

30.00
28.00

10.50

26.00

10.00

24.00

9.50

9.00

9.00

8.00

22.00
20.00

10.00

8.50

18.00
16.00

8.00

14.00

7.50
Carbohydrate

12.00
8.00

Fat

6.00

Protein

4.00

6.00
Carbohydrate

7.00

Calorie

10.00

7.00

Calorie

6.50

Fat

6.00

Sugar

Calorie
Fat

4.00

Protein

5.50

Sugar

5.00

2.00

Carbohydrate
5.00

Protein
3.00

Sugar

2.00
0

5 10 15 20 25 30 35 40 45 50 55 60 65

5 10 15 20 25 30 35 40 45 50 55 60 65

(a) Nutritions vs obesity

(b) Nutritions vs diabetes

10

15

20

25

30

35

40

(c) Nutritions vs health expenditure

Figure 9: Smoothed Average Vector of nutritions and health measures.

Table 1: Confusion Matrix for DNN Region Prediction

hard measures of health. As the results suggest, nutrition values

show a significant correlation with the health related measures of
countries. The dominant positive correlated nutritions are the sugar
and carbohydrate. It is intuitive because those are the main elements of snack meals like cakes, creams, etc which can contribute
to the health difficulties and the consequence expenditures eventually. On the other side, protein value shows strong negative correlation with the level of obesity and diabetes in societies. Noticeably,
the positive impact of high-protein diets on losing weight is frequently studied in the literature [11]. Figure 9 exhibits the SAV
for different nutritions and health measures. The trend of the diagrams endorses that including the countries with high average nutrition values (except protein) results in an increase in the average
health measures (e.g. average obesity). Proteins show completely
opposite patterns as expected. Including the countries with high
protein diets decrease the rate of health difficulties (e.g. obesity or
diabetes). A noticeable trait in both Table 2 and Figure 9 is that
the correlations and trends are more highlighted in the obesity results rather than the diabetes and health expenditure. The reason
is that the diabetes and health expenditure are more elaborate phenomena than the obesity. For example, in addition to consuming
foods, there are a variety of other genetic and environmental factors that may cause the diabetes. Remarkably, the genetic susceptibility of different ethnics varies so much [9]. As another example,
over intaking of proteins itself can lead to an spectrum of adverse
effects [8]. Therefore the relation of protein intaking and health
expenditures of the countries is not as clear as the relation between
obesity and proteins.

Actual Class

Prediction Outcome

LA
SA
OC
EA
AF
WE
ME
EE
NA

LA

SA

OC

EA

AF

WE

ME

EE

NA

1888
18
21
49
31
453
35
51
127

15
961
2
37
23
49
24
30
7

2
1
177
2
2
21
9
1
5

84
52
21
5211
28
660
107
58
128

25
16
3
13
704
85
72
29
22

455
40
119
342
136
9430
366
885
1045

13
17
5
26
57
165
634
41
16

33
5
6
24
7
541
78
1320
60

92
3
18
51
15
557
17
94
1508

penditure. Since the recipes are categorized by their cuisine but

the health measures are reported by countries, we need to map each
cuisine to one or more countries beforehand. To this end, we assign
each cuisine to the countries that include that cuisine geographically or semantically. For example, we mapped the Kurdish cuisine
to the countries: Iran, Syria, Iraq and Turkey because the Kurdistan
region is divided between those countries. As another instance, we
mapped the Italian-American cuisine to both Italy and USA countries because the cuisine represents both the Italian and American
food culture. This way, we consider the recipes of each country as
all of the recipes of the cuisines that belong to that country.
For each country, we calculate the average calorie, protein, fat,
carbohydrate and sugar values over its recipes. We used the following three methods to study the relation between nutritions and
health statistics:
1. Pearson Correlation: Pearson method calculates the linear
correlation between the vectors u and v, where each element
of u represents a nutrition value (e.g. sugar value) for a specific country, and each element of v represents a health measure (e.g. obesity rate) of that country.

7.

2. Kendall-Tau Correlation: Kendall-Tau correlation is used

to measure the ordinal correlation between u and v vectors.
3. Smoothed Average Vector: Given two vectors u and v, consider the vector u is sorted in ascending order as u0 . Then
v 0 would be the the reordered version of v to match the u0 .
Smoothed Average Vector for vectors
u and v is a vector s
P
i

RELATED WORK

Recently, public health has been increasingly analyzed through

the lens of the web and social media. We refer the reader to [6]
for an overview of the recent research in this area. Abbar et al. [1]
relate food mentions on Twitter conversations to the obesity and diabetes rates, using caloric values, and find a high correlation (coefficient 0.77) between caloric values of tweets and obesity values in
various states in the US. Low-obesity areas of USA have also been
shown to be more socially active on Instagram (posting comments
and likes) than those from high-obesity ones by Mejova et al. [17],
who present a large-scale analysis of pictures taken at 164K restaurants in the US.
Ahn et al. [2] study culture-specific ingredient connections, creating a flavor network from a dataset of about 56K recipes and
relating them to the geographical groupings of countries. Similar flavor-based food pairing studies are conducted on cuisines

v 0 [i]

. This measure
with the same size where s[i] = j=1i
captures the health trends with regards to the increases in
consumption of nutritions. Moreover, It is also robust against
the noisy variations of the underlying data.

Table 2 shows the correlation between different nutritions and the

7

Table 2: Countries Health Measures vs Their Recipes Nutrition Values

recipes on the Web, their ingredients, nutrition, similarities across

countries, and their relation with country health statistics. Our results have multiple implications: we found strong similarities between cuisines in neighboring countries, yet, the diversity of ingredients and flavors varies largely across the continents, mostly
affected by net migration trends. You are what you eat might be
a clich, but we did find quantitative evidence of a strong correlation between nutrition information of the recipes (e.g., in terms of
sugar intake) and obesity. Also, we demonstrated that deep learning can be used to effectively predicting cuisines from ingredients,
potentially providing possibility for fine-grained analysis of food
and dishes as well as improved recipe recommendations based on
individuals profile.
Our findings indicate that certain ingredients (e.g., mozzarella)
uniquely represent a certain cuisine (e.g., Italian) and there are
strong clusters of ingredients across neighboring countries. This
feature eases the prediction of regions (e.g., continents) from the
combination of ingredients in a cuisine. Moreover, the correlation between the ingredients and health conditions, such as diabetes, is naturally of great importance for public health experts,
where behavior nudges or recommendation of similar dishes in flavor and ingredient complexity can be utilized to improve dietary
intake [10, 20].
In future work, we plan to explore the possibility of recipe recommendation based on regional and personal tastes and user ratings. This is important as a local Chinese dish or a distinct flavor
combination may be alien to, e.g., a Western person, but of interest to a Japanese individual. We also wish to asses the ability to
model flavors with ingredients, and discover ingredients to match
a specific flavor palette. Finding answers to these questions would
provide a better understanding of the composition of flavors and
ingredients in popular dishes and provide a better recommendation
system for a healthier, tastier, and more diverse experience.

Correlation Values
Health Measure

Nutrition

Pearson

Kendall-Tao

Obesity

Calorie
Protein
Fat
Carbohydrate
Sugar

0.104
0.483
0.115
0.300
0.461

0.110
0.299
0.127
0.201
0.293

Diabetes

Calorie
Protein
Fat
Carbohydrate
Sugar

0.077
0.162
0.123
0.173
0.142

0.048
0.022
0.063
0.106
0.066

Health Expend.

Calorie
Protein
Fat
Carbohydrate
Sugar

0.098
0.083
0.197
0.064
0.134

0.110
0.022
0.141
0.015
0.069

in distinct geographical areas such as India [12]. West et al. [25]

mine logs of recipe-related queries to uncover temporal patterns in
consumption. Using Fourier transforms, they show the yearly and
weekly periodicity in food density of the searched recipes, with
different trends in Southern and Northern hemispheres, suggesting
a link between food selection and climate. A study of Austrian
recipe sites by Wagner et al. [23] also highlights differences in the
recipes of regions which are further apart. Zhu et al. [27] conduct a similar study on Chinese recipes to investigate the effect of
geographical and climatic proximities on ingredients similarity of
domestic cuisines.
Kular et al. [15] create a network of recipes using a dataset of 300
recipes from 15 different countries, and show the networks smallworld and scale-free properties. As opposed to this line of work,
we also exploit flavor and nutritional information, alongside health
statistics countries to provide a deeper analysis about the dishes,
cuisines, culinary cultures, and the impact of food on human life.
Su et al. [22] investigate underlying connections between cuisines
and ingredients via machine learning classification, with an application to predicting the cuisine by looking at recipes. Like our
analysis, theirs is based on a large-scale data collection of recipes
specifically, 226K recipes collected from food.com. However, they
only look at classifying cuisines using Support Vector Machine
(SVM), while we propose a deep neural network architecture to
capture the highly non-linear relation of a recipe cuisine and its associated ingredients. The results approve that the proposed deep
model outperforms SVM by a significant margin in terms of prediction accuracy and F-1 measure.
There are major difference between our work and the ones discussed above, in both scale and domain. An important characteristic of our work comes from the size and the quality of the various
datasets we used, which enable us to derive first-of-its-kind insight
on worldwide cuisines and their relationship to health factors. In
addition to the ingredients, we also exploited flavor and nutritional
information, alongside health and immigration statistics, allowing
us to perform a deeper analysis of the dishes, cuisines, culinary
cultures, as well as the impact of food on human life.

8.

9.

REFERENCES

[1] S. Abbar, Y. Mejova, and I. Weber. You tweet what you eat:
Studying food consumption through Twitter. In Proceedings
of the 33rd Annual ACM Conference on Human Factors in
Computing Systems, pages 31973206, 2015.
[2] Y.-Y. Ahn, S. E. Ahnert, J. P. Bagrow, and A.-L. Barabsi.
Flavor network and the principles of food pairing. Nature
Scientific reports, 2011.
[3] B. Armstrong and R. Doll. Environmental factors and cancer
incidence and mortality in different countries, with special
reference to dietary practices. International journal of
cancer, 15(4):617631, 1975.
[4] M. Bastian, S. Heymann, M. Jacomy, et al. Gephi: An Open
Source Software for Exploring and Manipulating Networks.
In International AAAI Conference on Web and Social Media,
pages 361362, 2009.
[5] L. Buitinck, G. Louppe, M. Blondel, F. Pedregosa,
A. Mueller, O. Grisel, V. Niculae, P. Prettenhofer,
A. Gramfort, J. Grobler, R. Layton, J. VanderPlas, A. Joly,
B. Holt, and G. Varoquaux. API design for machine learning
software: experiences from the scikit-learn project. In ECML
PKDD Workshop: Languages for Data Mining and Machine
Learning, pages 108122, 2013.
[6] D. Capurro, K. Cole, M. I. Echavarra, J. Joe, T. Neogi, and
A. M. Turner. The use of social networking sites for public
health practice and research: a systematic review. Journal of
medical Internet research, 16(3):e79, 2014.
[7] F. Chollet. Keras. https://github.com/fchollet/keras, 2015.

CONCLUSION
This paper presented a large-scale study of user-generated
8

[8] I. Delimaris. Adverse effects associated with protein intake

above the recommended dietary allowance for adults. ISRN
nutrition, 2013, 2013.
[9] S. C. Elbein. Genetics factors contributing to type 2 diabetes
across ethnicities. Journal of diabetes science and
technology, 3(4):685689, 2009.
[10] G. D. Foster, A. P. Makris, and B. A. Bailer. Behavioral
treatment of obesity. The American journal of clinical
nutrition, 82(1):230S235S, 2005.
[11] T. L. Halton and F. B. Hu. The effects of high protein diets
on thermogenesis, satiety and weight loss: a critical review.
Journal of the American College of Nutrition,
23(5):373385, 2004.
[12] A. Jain, N. Rakhi, and G. Bagler. Analysis of food pairing in
regional cuisines of india. PloS one, 10(10):e0139539, 2015.
[13] A. Keys, A. Mienotti, M. J. Karvonen, C. Aravanis,
H. Blackburn, R. Buzina, B. Djordjevic, A. Dontas,
F. Fidanza, M. H. Keys, et al. The diet and 15-year death rate
in the seven countries study. American journal of
epidemiology, 124(6):903915, 1986.
[14] M. Kratz, T. Baars, and S. Guyenet. The relationship
between high-fat dairy consumption and obesity,
cardiovascular, and metabolic disease. European journal of
nutrition, 52(1):124, 2013.
[15] D. K. Kular, R. Menezes, and E. Ribeiro. Using network
analysis to understand the relation between cuisine and
culture. In Network Science Workshop (NSW), 2011 IEEE,
pages 3845, 2011.
[16] Y. Mejova, S. Abbar, and H. Haddadi. Fetishizing food in
digital age:# foodporn around the world. In International
AAAI Conference on Web and Social Media (ICWSM 2016),
2016.
[17] Y. Mejova, H. Haddadi, A. Noulas, and I. Weber. #FoodPorn:
Obesity patterns in culinary interactions. In Proceedings of
the 5th International Conference on Digital Health 2015,

pages 5158, 2015.

[18] M. Michel de Lorgeril, P. Salen, J.-L. Martin, I. Monjaud,
J. Delaye, and N. Mamelle. Mediterranean diet, traditional
risk factors, and the rate of cardiovascular complications
after myocardial infarction. Heart failure, 11:6, 1999.
[19] A. Noack. Modularity clustering is force-directed layout.
Phys. Rev. E, 79:026102, Feb 2009.
[20] N. Regulating. Judging nudging: can nudging improve
population health? Bmj, 342:263, 2011.
[21] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov. Dropout: a simple way to prevent neural
networks from overfitting. Journal of Machine Learning
Research, 15(1):19291958, 2014.
[22] H. Su, T.-W. Lin, C.-T. Li, M.-K. Shan, and J. Chang.
Automatic recipe cuisine classification by ingredients. In
Proceedings of the 2014 ACM Joint Conference on Pervasive
and Ubiquitous Computing, pages 565570, 2014.
[23] C. Wagner, P. Singer, and M. Strohmaier. Spatial and
Temporal Patterns of Online Food Preferences. In
Proceedings of the Companion Publication of the 23rd
International Conference on World Wide Web Companion,
WWW Companion 14, pages 553554. International World
Wide Web Conferences Steering Committee, 2014.
[24] L. Waltman, N. J. van Eck, and E. C. Noyons. A unified
approach to mapping and clustering of bibliometric
networks. Journal of Informetrics, 4(4):629635, 2010.
[25] R. West, R. W. White, and E. Horvitz. From cookies to
cooks: insights on dietary patterns via analysis of web usage
logs. In WWW, 2013.
[26] M. D. Zeiler. Adadelta: an adaptive learning rate method.
arXiv preprint arXiv:1212.5701, 2012.
[27] Y.-X. Zhu, J. Huang, Z.-K. Zhang, Q.-M. Zhang, T. Zhou,
and Y.-Y. Ahn. Geography and similarity of regional cuisines
in china. PloS one, 8(11):e79161, 2013.