=Paper=
{{Paper
|id=None
|storemode=property
|title=User's Food Preference Extraction for Cooking Recipe Recommendation
|pdfUrl=https://ceur-ws.org/Vol-781/paper11.pdf
|volume=Vol-781
|dblpUrl=https://dblp.org/rec/conf/semweb/UedaTN11
}}
==User's Food Preference Extraction for Cooking Recipe Recommendation==
https://ceur-ws.org/Vol-781/paper11.pdf
User’s Food Preference Extraction for
Personalized Cooking Recipe Recommendation
Mayumi Ueda*, Mari Takahata**, and Shinsuke Nakajima***
Kyoto University*
Yoshida Nihonmatsu-cho, Sakyo-ku, Kyoto, Kyoto 606–8501, Japan
Kyoto Sangyo University**
Motoyama, Kamigamo, Kita-Ku, Kyoto-City 603–8555, Japan
mayumi@mm.media.kyoto-u.ac.jp,* g0846741@cc.kyoto-su.ac.jp,**
nakajima@cse.kyoto-su.ac.jp***
Abstract. There are many websites and researches that involve cooking
recipe recommendation. However, these websites present cooking recipes
on the basis of entry date, access frequency, or the recipe’s user rat-
ings. They do not reflect the user’s personal preferences. We have pro-
posed a personalized recipe recommendation method that is based on the
user’s food preferences. For extracting the user’s food preferences, we use
his/her recipe browsing and cooking history. In this paper, we present
a method for extracting the user’s preferences. In the experimental re-
sults, extracting the user’s favorite ingredients were detected with a 60 to
83% of precision. And extracting the unfavorite ingredients were detected
with 14.7% of precision, and 58% of recall. Furthermore, the F-measure
value for extraction of favorite ingredients was 60.8% when we focused
on the top 20 ingredients.
Key words: user’s food preferences, preference extraction, recipe rec-
ommendation, cooking and browsing history
1 Introduction
As a result of the lifestyle-related disease epidemic, dietary life is now attracting
attention. Good eating habits are important for maintaining a healthy life. How-
ever, menu planning requires one to take various factors into consideration, such
as the nutritional value, food in stock, food preferences, and cost. Thus, people
need to expand a lot of effort toward planning their daily menu. Against this
background, a number of cooking websites comprising various food recipes have
recently been launched, such as Cookpad[1] and Yahoo! Recipe[2]. Many people
refer to these websites when planning their menu. Cookpad contains 900,000
recipes and has 10,000,000 monthly users[3]. This data reflects the high demand
for recipe-providing services. However, these websites do not reflect user’s pref-
erences and conditions, although these two factors need to be considered if the
goal is to provide high-satisfactory recipes.
Furthermore, several researches on cooking recipe recommendation for menu
planning support have been conducted in the past. Mino et al. propose a method
98
that takes the user’s schedule into consideration[4]. This paper defines the eval-
uation value of either the intake or consumption calories that are assigned to
each event in the user’s schedules. Karikome et al. propose a system that helps
users plan nutritionally balanced menus and visualize their dietary habits[5].
Their system calculates the nutritional value of each dish, and records this in-
formation in the form of a dietary log. Next, the system recommends recipes
foster sound nutrition. Freyne et al. show the results of their investigation in
which they compare three recommendation strategies: content-based, collabora-
tive, and hybrid[6].
In these circumstances, we have proposed a recipe recommendation method
based on the user’s food preferences[7]. Our method breaks recipes down into
their ingredients, and scores them on the basis of the frequency of use and speci-
ficity of the ingredients. Furthermore, our proposed system does not recommend
dishes that are similar to the food the users have eaten over the past few days on
the grounds that people do not want to eat similar dishes iteratively. Moreover,
our system does not require any particular action on the user’s past to reflect
his/her food preferences: it estimates the user’s food preferences automatically
through his/hers recipe browsing and cooking history. In this paper, we present
a method for extracting the user’s preferences.
This paper is structured as follows. Section 2 describes the method of scoring
recipes and extracting user’s preferences. Section 3 shows experimental results,
using precision and recall. Section 4 shows concludes the paper.
2 Scoring Recipes and Extracting User’s Preferences
In the recent years, concern over various health issues, such as lifestyle-related
diseases and diets, has been growing. It has also been noted that picky eating
is one of the main reasons causing these health issues. However, people do not
want to eat food that they dislike even if it perfectly addresses their nutritional
needs. They hope to derive essential nutrition solely from their favorite foods.
We conducted questionnaire to the 20 men and women in their 20s to 40s to
survey the key considerations for menu planning. According to the results of the
questionnaire, people consider the following elements (1) food preferences, (2)
nutritional balance and calories, (3) ingredients they have in stock or they can
procure easily, (4) easily to cook, and (5) mood. Therefore, in this paper, we
focus on (1) food preferences and try to extract user’s food preferences.
2.1 Preferences for Ingredients
We express the user’s food preferences Ik by using in the form of the following
Eq.(1).
Ik = Ik+ + Ik− (1)
99
Fig. 1. Extracting the favorite ingredients using cooking history.
User’s favorite ingredients Fig.1 shows the key idea behind estimating user’s
favorite ingredients by his/her cooking history. Our method considers the ingre-
dient that the user eats repeatedly as his/her favorite ingredients. It breaks
recipes down into their ingredient as the outset and calculates the score of in-
gredients Ik+ by incorporating the frequency of use of the ingredients in the
dishes that the target user has eaten(F Fk : Foodstuff Frequency) as well as the
specificity of ingredients(IRFk : Inverted Recipe Frequency) into Eq.(2). This
equation is based on the idea of TF-IDF.
Ik+ = F Fk × IRFk (2)
For estimating the user’s favorite ingredients by using the frequency of use of
ingredient k (F Fk ), we utilize the simple frequency of use of ingredient k (Fk )
during a definite period D, as shown in Eq.(3).
Fk
F Fk = (3)
D
Then, we calculate —it the specificity of ingredient k (IRFk ) using the total
number of recipe (M ) and the number of recipes that contain ingredient k (Mk ),
as shown in Eq.(4).
M
IRFk = log (4)
Mk
User’s disliked ingredients We consider that user’s food preferences are also
influenced by his/her disliked ingredients. We estimate the user’s disliked in-
gredients, by considering the ingredients in the recipes that he/she has never
100
Fig. 2. Extracting the disliked ingredients using browsing and cooking history.
cooked, even if he/she has browsed the recipe details. Fig.2 shows the estimat-
ing method for user’s disliked ingredients through the user’s recipe browsing and
cooking history. N corresponds to the set of ingredients in the recipes that the
user has not browse. C corresponds to the set of ingredients in the recipes that
the user has cooked over the past few days. U corresponds to the set of ingredi-
ents in the recipes that the user has not cooked, even if he/she has browse them
completely. For example, “shrimp” in Fig.2 corresponds to the user’s disliked
ingredient. We calculate the score of disliked ingredient k (Ik− ) in Eq.(5).
{
k|
− 0 (0 < 2|U
|Ak | 5 0.5)
Ik (x) = k| k|
(5)
( 2|U
|Ak | − 1)
x
(0.5 < 2|U
|Ak | 5 1)
|Uk | denotes the presence of ingredient k in U and |Ak | denotes the presence of
ingredient k in the recipe database. We should investigate the ratio or frequency
of the user’s avoidance of the ingredient that he/she dislikes, because he/she will
use the ingredients that he/she does not like. x denotes the ratio or frequency
of avoiding the ingredients, and we plan to verify x through some preliminary
experiments.
2.2 Recipe Scoring
Our method scores cooking recipes in accordance with the estimation results
regarding favorite/disliked ingredients, and then provides recipes in decreasing
order of the scores. In general, people do not like eating dishes similar to those
they have eaten in the past few days. Therefore, our method weights recipes to
avoid the repetition of similar dishes. The score of cooking recipes are defined as
shown in Eq.(6).
∑ ∑
Score(R) = Ik − α (wd · sim (R, Rd )) (6)
k∈R d=1
101
d denotes the weight for avoiding repeating similar dishes iteratively. sim (R, Rd )
denotes the similarities between the considered recipe R and the recipe of the
dish eaten d days ago Rd . The weight wd for avoiding similar dishes eaten d days
ago is defined as shown in Eq.(7).
d−1
wd = 1 − (1 5 d 5 7) (7)
7
3 Evaluation of the Accuracy of Extracting User’s Food
Preference
3.1 Experimental Condition
In order to verify the extracting accuracy of the user’s food preferences, we
conducted simple experiments. We used 100 recipes extracted from Cookpad[1]
that is a most popular recipe search website in Japan. We used randomly selected
recipes categorized as main dish.
We conducted experiment as follows.
1. We present a list of 10 recipe titles to subjects.
2. He/She chooses recipes which he/she would like to browse completely, such
as ingredients, procedures, and so on.
3. He/She chooses one recipe that he/she would like to cook.
4. Repeat this sequence(Step 1 to 3) 10 times.
We gathered each user’s browsed recipes and recipes he/she would like to cook
through the above procedure. 6 men and women in their 20s to 40s participated
in this experiment as subjects.
We calculated the specificity of ingredient k (IRFk ) in the target 100 recipes,
using Eq.(4). Table 1 shows the examples of IRFk . Furthermore, we collected
the labeled data as the user’s preferences via questionnaire. Responses are coded
on a 6-point scale, ranging from “love” to “hate”.
3.2 Evaluation of Extracting the Favorite Ingredients
We evaluated the extraction accuracy of the user’s favorite ingredients. We ex-
tracted individual user’s favorite ingredients by calculating Ik+ , using Eq.(2). For
evaluating the accuracy, we calculated precision, recall, and F-measure for top
N ingredients which sorted by Ik+ . The results are shown in Table 2 and Fig.3.
The precision, recall, and F-measure were calculated by the number of extracted
user’s favorite ingredients in the top N (E) and the number of user’s favorite
ingredients via questionnaire (Q), as follows.
E
P recision =
N
E
Recall =
Q
102
Table 1. Examples of the specificity of ingredient k (IRFk )
ingredient IRFk ingredient IRFk ingredient IRFk
pumpkin 1.70 oyster 1.52 white wine 1.10
cucumber 1.70 eggplant 1.40 mayonnaise 1.01
burdock root 1.70 parsley 1.40 tomato 1.00
konjac 1.70 honey 1.40 sesame oil 0.96
snow crab 1.70 beef 1.40 carrot 0.92
green pepper 1.70 shrimp 1.30 butter 0.92
yellowtail 1.70 bacon 1.30 milk 0.92
lettuce 1.70 Japanese radish 1.22 mushroom 0.89
yam 1.70 minced meat 1.22 egg 0.77
soy milk 1.70 potato 1.15 pork 0.74
pickled plum 1.70 miso 1.15 ginger 0.68
chinese chive 1.52 lemon 1.15 garlic 0.57
bean sprout 1.52 tofu 1.15 cibol 0.57
salmon 1.52 cabbage 1.10 chicken 0.54
Colocasia esculenta 1.52 cheese 1.10 onion 0.44
Table 2. Elicitation accuracy of user’s food preferences
N Recall Precision F-measure
1 0.045 0.833 0.086
3 0.124 0.778 0.213
5 0.196 0.733 0.309
10 0.410 0.767 0.535
15 0.521 0.656 0.580
20 0.609 0.607 0.608
2 · P recision · Recall
F -measure =
P recision + Recall
(N = 1, 3, 5, 10, 15, 20) (8)
As shown in Table 2, when we focused on the only top one ingredient (N = 1),
our method extracted with precision of 83.3%. However, at the same time, the
recall was only 4.5%. When we focused on the top 20 ingredients (N = 20),
the recall was increased to 61%. Since the average number of individual user’s
favorite ingredients was 19.2, the recall was not enough value when we focused on
a few ingredients such as N = 1, 3, 5. It was found from the result that even if our
system focused on the top 20 ingredients (N = 20), the precision did not reduce
very much. Furthermore, the highest value of F-measure, at this experiment,
was 60.8% when we focused on the top 20 ingredients (Table 2). Therefore, the
results show that our system should focus on the top 20 ingredients sorted by
Ik+ for recipe recommendation.
103
Fig. 3. The precision and recall for extracting the favorite ingredients.
3.3 Evaluation of Extracting the Disliked Ingredients
We evaluated the accuracy of extracting user’s disliked ingredients. We extracted
individual user’s disliked ingredients by calculating Ik− , using Eq.(5). The preci-
sion of the extracting disliked ingredients was 14.7%, and the recall was 58.3%.
In this experiment, we considered that the ingredients which were contained only
in U , were the user’s disliked ingredients. In other words, in this experiment, our
method estimated the user’s disliked ingredients which were in the recipes that
he/she has never cooked, even if he/she has browsed the complete recipe.
disliked ingredients = {U ∩ C} (9)
In this experiment, we could not extract satisfactory accuracy for disliked ingre-
dients, because of the lack of the number of experiments. Since the number of
experiments was not enough, our method determined that the ingredient which
happened to include in U was his/her disliked ingredient. We consider that our
method can improve the accuracy of extracting favorite ingredients, by increas-
ing the number of experiments.
4 Conclusion
In this paper, we presented a method for extracting the user’s food preferences for
recipe recommendation. Our method estimates a user’s preferences from his/her
past actions, such as through their recipe browsing and menu planning history.
For extracting the preferences, our method breaks recipes down into their ingre-
dients and scores the recipes using the frequency and specificity of ingredients.
Since our method can estimate the preferences through their browsing and cook-
ing history, the user convey his/her preferences to the system without having
to carry out any particular operation. Furthermore, the user can convey the
changes in his/her preferences to the system on a daily basis.
104
In order to verify the extracting accuracy of the user’s food preferences,
we conducted simple experiments. In the experimental results, extracting the
user’s favorite ingredients were detected with a 60 to 83% of precision. And
the F-measure was 60.8% when we focused on the top 20 ingredients. Since the
average number of user’s favorite ingredients was 19.2, our system should focus
on the top 20 ingredients sorted by Ik+ to score recipes for recommendation. And
extracting the disliked ingredient were detected with 14.7% of precision, and 58%
of recall. In this time, we could not extract satisfactory accuracy. However, we
consider that our method can improve the accuracy of extraction, by increasing
the number of experiments.
As future work, we plan to consider the ingredient ontology for estimating
favorite/disliked ingredients. Furthermore, we plan to investigate the various
weights. For example, we plan to verify x in Eq.(5), the ratio or frequency of
avoiding the ingredients, through experiments. Moreover, we want to investigate
the length of time for which the system should avoid recommending similar
dishes for d in Eq.(6),(7), and the degree of similarity that the system should
consider while refraining from recommending similar dishes for α in Eq.(6).
Acknowledgments
This work is supported in part by Kyoto University Global COE Program: In-
formatics Education and Research Center for Knowledge-Circulating Society
(MEXT Global COE Program) , and the MEXT Grant-in-Aid for Scientific
Research(C) (#23500140).
References
1. Cookpad. http://cookpad.com/, (Accessed 12 August 2011)
2. Yahoo! Recipe. http://recipe.gourmet.yahoo.co.jp/, (Accessed 12 August 2011)
3. Choose dishes using websites of cooking. (in Japanese), http://otona.-
yomiuri.co.jp/news/news110118 03.htm, (Accessed 12 August 2011)
4. Yoko Mino and Ichiro Kobayashi. Recipe Recommendation for a Diet Consid-
ering a User’s Schedule and the Balance of Nourishment. Proceedings of IEEE
International Conference on Intelligent Computing and Intelligent Systems 2009,
pp.383-387. 2009.
5. Shihono Karikome and Atsushi Fujii. A System for Supporting Dietary Habits:
Planning Menus and Visualizing Nutritional Intake Balance. Proceedings of the
4th International Conference on Ubiquitous Information Management and Com-
munication, pp.386-391. 2009.
6. Jill Freye and Shlomo Berkovsky. Intelligent Food Planning: Personalized Recipe
Recommendation. In IUI(2010) 321-324.
7. Mayumi Ueda, Mari Takahata and Shinsuke Nakajima. Recipe Recommendation
Method Based on User’s Food Preferences. Proceedings of the IADIS International
Conference on e-Society 2011, pp.591-594. 2011.
105