estimating novelty and surprise for food recipe generation …

estimating novelty and surprise for food recipe generation …
24 views 3 downloads 282.92 K size 2 pages cake

Report DMCA / Copyright

Title : estimating novelty and surprise for food recipe generation …

PDF summary :

Title : casi_cc_poster
Producer : macOS Version 11.2.3 (Build 20D91) Quartz PDFContext
Author : Mykhaylo Zayats
Creator : PowerPoint
Creationdate : 1970-January-01
Moddate : 1970-January-01
Rgid : PB:351359327_AS:[email protected]

Preview content in text :

See discussions, stats, and author profiles for this publication at:
Estimating Novelty and Surprise For Food Recipe Generation Problem Link
prediction problem

Poster · May 2019

DOI: 10.13140/RG.2.2.16996.50569

CITATIONS
0

2 authors, including:

Mykhaylo Zayats

National University of Ireland, Galway

9 PUBLICATIONS   10 CITATIONS   

SEE PROFILE

READS
13

All content following this page was uploaded by Mykhaylo Zayats on 06 May 2021.

The user has requested enhancement of the downloaded file.

Estimating Novelty and Surprise For Food Recipe
Generation Problem

Mykhaylo Zayats1, Cian O’Hagan1

Accenture The Dock, Dublin, Ireland

The creativity process is often hard and time consuming, even just for generating novel combinations among existing elements, whether they are groceries,
chemical compounds, sounds or colours. We aim to build a system capable of creating those connections in a fast programmatic way.

We consider a problem of a culinary recipe generation and attempt solving it in two steps: i) building a formal model of a recipe space and ii) computing
novelty and surprise metrics for selecting interesting combinations. To this end, we propose two approaches.

Knowledge Graph Approach:

has_ing

has_comp

Recipe

Ingredient

Compound

has_type

type

n
g

i

i
r
o
_
s
a
h

origin

has_nutr

Nutrient

Knowledge graph represents data as a set of
triples (facts):



for example:

t=
t=

and maps entities to vector embeddings 𝑒!, 𝑒", 𝑒#

Skip-gram Model Approach:

Skip-gram network (classificatory) learns ingredients that
occur together with the input ingredient

- denotes ComplEx scoring function which
maps set of vectors to a number

Link prediction problem

𝑝 𝐼 ∈ 𝑅 = 𝜎

Ingredients set 𝐼$ probability model (novelty):
𝑝 𝐼$ ∈ 𝑅 = 𝜎
𝑝 𝐼$ = 𝐸 𝑝 𝐼$ ∈ 𝑅%

Surprise of a sequence 𝐼$ :

𝑠 𝐼$ =

𝐾𝐿 𝑝( 𝐼$ ), 𝑝 𝐼$ ∈ 𝑅

1
𝑁"

0
"

=

1
𝑁"

0
"

𝑝 𝐼$

log

𝑝 𝐼$
𝑝 𝐼$ ∈ 𝑅

Network input vector 𝐼’(!)* may encode a set of
ingredients 𝐼$ , so that the last layer (softmax) produces
conditional probabilities

𝑝(𝐼’| 𝐼$ )

Ingredients set 𝐼$ probability model (novelty):

(
𝑝 𝐼$ = 𝑝 𝐼& 9
’.+
𝑝 𝐼$ ∈ 𝑅 = 𝑝 𝐼$ + {𝐼"}

𝑝 𝐼’|𝐼&, … , 𝐼’/&

where 𝐼" = {𝐼: 𝐼 ∈ 𝑅}

Surprise of a sequence 𝐼$ :

𝑠 𝐼$ =

𝑝 𝐼$

log

1
𝑁"

0
"

𝑝 𝐼$
𝑝 𝐼$ + {𝐼"}

𝑝(𝐼&|𝐼’(!)*)

𝑝(𝐼+|𝐼’(!)*)

𝑝(𝐼,|𝐼’(!)*)

𝑝(𝐼-|𝐼’(!)*)

KG surprise

SG surprise

0.1489

0.1698

0.1891

0.1823

0.13439

0.13438

0.13429

0.13425

Results:

Combination

onions, salt

carrots, garlic

apple, whiskey

feta, grapefruit

View publication stats
View publication stats

Discussion:
• KG model provides more consistent results then SG model.
• KG model is faster from the computational perspective then SG model.
• Validation protocols for surprise metrics remain an open question.


Related Recipes :

premier protein shake recipes weight watchers
13 2 79.43 K 3 pages

premier protein shake recipes weight watchers

Author : Voyuteniwi Kigahune
enjoy your recipe cards from family cookie recipes
18 4 433.07 K 1 pages

enjoy your recipe cards from family cookie recipes

recipes for everyday baking taste of the south magazine
18 8 1.19 M 15 pages

recipes for everyday baking taste of the south magazine

commonly used recipes
18 5 2.35 M 25 pages

commonly used recipes

gluten sorghum recipes unl food
19 2 311.04 K 4 pages

gluten sorghum recipes unl food

Author : User
fat burning recipes the superhuman life
30 3 732.73 K 51 pages

fat burning recipes the superhuman life

Author : Kelsey Heenan
Coconut Cake   CokeSolutions
30 6 139.77 K 2 pages

Coconut Cake CokeSolutions

Coat rim of a 16-oz. glass with water; dip into toasted coconut flakes and cake crumbs. 2. Combine all ingredients in a shaker with ice, shake until.
Frosting for Italian Cream Cheese Cake
29 7 13.97 K 1 pages

Frosting for Italian Cream Cheese Cake

Italian Cream Cheese Cake-. 1 stick butter. ½ cup vegetable oil. 2 cups sugar. 5 egg whites, beaten. 1 teaspoon baking soda. 3 teaspoons vanilla extract.
german chocolate cake martini ingredients rim garnish directions
12 10 22.09 K 1 pages

german chocolate cake martini ingredients rim garnish directions