Data Science in Cooking

Stephanie Dodson

Group MeetingMarch 22, 2017

S. Dodson Data Science in Cooking March 22, 2017 1 / 17

Main Source

(2011)

S. Dodson Data Science in Cooking March 22, 2017 2 / 17

Outline

Flavor Network

Applications

– Food Pairing Hypothesis– Regional Cuisines– FPH in Medieval Times

Limitations

Conclusions

S. Dodson Data Science in Cooking March 22, 2017 3 / 17

Flavor Network

Food scientistis have linked ingredients with flavor compounds

Contains: 381 ingredients and 1,021 flavor components

S. Dodson Data Science in Cooking March 22, 2017 4 / 17

Flavor Network

S. Dodson Data Science in Cooking March 22, 2017 5 / 17

What questions can the flavor network answer?

S. Dodson Data Science in Cooking March 22, 2017 6 / 17

Question 1: Food Pairing

Food Pairing Hypothesis

Ingredients sharing flavor components are more likely to taste well together thaningredients that do not.

Uses:

– Learn why foods taste good– Food science– Search for novel food combinations:

white chocolate and caviar,chocolate and blue cheese

Reformulate with math: Are ingredient pairs in recipes strongly connected in theflavor network?

S. Dodson Data Science in Cooking March 22, 2017 7 / 17

Question 1: Food Pairing

Food Pairing Hypothesis

Ingredients sharing flavor components are more likely to taste well together thaningredients that do not.

Uses:

– Learn why foods taste good– Food science– Search for novel food combinations:

white chocolate and caviar,chocolate and blue cheese

Reformulate with math: Are ingredient pairs in recipes strongly connected in theflavor network?

S. Dodson Data Science in Cooking March 22, 2017 7 / 17

Testing the Food Pairing Hypothesis

56,498 recipes from:

– American sources epicurious.com and allrecipes.com– Korean source menupan.com

Recipes grouped into distinct regions

S. Dodson Data Science in Cooking March 22, 2017 8 / 17

Testing the Food Pairing Hypothesis

Number of shared compounds, Ci , in recipe R with nR ingredients

Ns(R) =2

nR(nR − 1)

∑i,j∈R,i 6=j

|Ci ∩ Cj |

Real examples:

Mustard Cream Pan Sauce: chicken broth, mustard, cream → Ns(R) = 0Sweet and Simple Pork Chops: pork, apples, cheddar → Ns(R) = 60

For each category, compared mean number of shared compounds in recipes (N reals )

with the mean number in 10,000 randomly constructed reciepes (N rands )

∆Ns = N reals − N rand

s

N reals =

∑R

Ns(R)/Nc

S. Dodson Data Science in Cooking March 22, 2017 9 / 17

Food Pairing Hypothesis: Results

North American recipes tend to share more compounds, East Asian share less.

S. Dodson Data Science in Cooking March 22, 2017 10 / 17

Question 2: Regional Cuisines

Are specific compounds/ingredients responsible for regional differences?

χci =

1

NC

∑i∈R

2

nR(nR − 1)

∑j 6=i(j,i∈R)

|Ci ∩ Cj |

−( 2fiNc < nR >

∑j∈c fj |Ci ∩ Cj |∑

j∈c fj

)

Positive χi values increase the number of shared compounds.

S. Dodson Data Science in Cooking March 22, 2017 11 / 17

Question 3: Flavors of Medieval Europe, 1300 - 1615

(2013)

S. Dodson Data Science in Cooking March 22, 2017 12 / 17

Question 3: Flavors of Medieval Europe, 1300 - 1615

Was the Food Pairing Hypothesis true during Medieval Times?

Recipe Database:

4,133 recipes from 1300 – 161525 source texts from England, France, Germany and ItalyManually placed ingredients into 391 equivalence groupings386 different ingredients

Compound Databases:

Volatile Compounds in Food (VCF) - more complete datasetFenaroli’s Handbook of Flavor Ingredients - sparser

FPH true with Fenaroli, not with VCF

S. Dodson Data Science in Cooking March 22, 2017 13 / 17

Question 3: Flavors of Medieval Europe, 1300 - 1615

Was the Food Pairing Hypothesis true during Medieval Times?

Recipe Database:

4,133 recipes from 1300 – 161525 source texts from England, France, Germany and ItalyManually placed ingredients into 391 equivalence groupings386 different ingredients

Compound Databases:

Volatile Compounds in Food (VCF) - more complete datasetFenaroli’s Handbook of Flavor Ingredients - sparser

FPH true with Fenaroli, not with VCF

S. Dodson Data Science in Cooking March 22, 2017 13 / 17

Flavors of Medieval Europe

Columbian Exchange (1492) resulted in variety of new foods

Source: thecolumbianexchange.weebly.com

Conjecture: Western cooks maintained FPH after Columbian Exchange

S. Dodson Data Science in Cooking March 22, 2017 14 / 17

Flavors of Medieval Europe

Columbian Exchange (1492) resulted in variety of new foods

Source: thecolumbianexchange.weebly.com

Conjecture: Western cooks maintained FPH after Columbian Exchange

S. Dodson Data Science in Cooking March 22, 2017 14 / 17

Limitations and Conclusions

Limitations:

Flavors depend on method of cooking

Does not account for texture, color, sound, or temperature

Some ingredients have structural role (eggs)

Does not include flavor compound concentration

Additional Applications:

Searching for similar recipesRecipe recommenderIngredient recommenderStudy cultural differences and mixing of culturesStudy historical culinary trends

S. Dodson Data Science in Cooking March 22, 2017 15 / 17

Limitations and Conclusions

Limitations:

Flavors depend on method of cooking

Does not account for texture, color, sound, or temperature

Some ingredients have structural role (eggs)

Does not include flavor compound concentration

Additional Applications:

Searching for similar recipesRecipe recommenderIngredient recommenderStudy cultural differences and mixing of culturesStudy historical culinary trends

S. Dodson Data Science in Cooking March 22, 2017 15 / 17

References

Ahn, Y.-Y., Ahnert, S. E., Bagrow, J. P., & Barabsi, A.-L. (2011). Flavornetwork and the principles of food pairing. Scientific Reports, 1(1), 217.http://doi.org/10.1038/srep00196

Varshney, K.R., L.R. Varshney, J. Wang, D. Myers, Flavor Pairing inMedieval Europen Cuisine: A study in Cooking with Dirty Data. InternationalJoint Conference on Artificial Intelligence Workshops, Beijing, China, August2013. arXiv:1307.7982

S. Dodson Data Science in Cooking March 22, 2017 16 / 17