=Paper=
{{Paper
|id=Vol-1438/paper9
|storemode=property
|title=Interaction Design in a Mobile Food Recommender System
|pdfUrl=https://ceur-ws.org/Vol-1438/paper9.pdf
|volume=Vol-1438
|dblpUrl=https://dblp.org/rec/conf/recsys/ElahiGRFBM15
}}
==Interaction Design in a Mobile Food Recommender System==
https://ceur-ws.org/Vol-1438/paper9.pdf
Interaction Design in a Mobile Food Recommender System
Mehdi Elahi Mouzhi Ge Francesco Ricci
Politecnico di Milano, Italy Free University of Free University of
mehdi.elahi@polimi.it Bozen-Bolzano, Italy Bozen-Bolzano, Italy
mouzhi.ge@unibz.it fricci@unibz.it
Ignacio Shlomo Berkovsky Massimo David
Fernández-Tobías CSIRO, Australia Free University of
Universidad Autónoma de shlomo.berkovsky@csiro.au Bozen-Bolzano, Italy
Madrid, Spain david.massimo@stud-inf.unibz.it
ignacio.fernandezt@uam.es
ABSTRACT 18, 20, 4, 11]. There is a broad spectrum of available in-
One of the most important steps in building a recommender formation about food, such as recipe data and cooking in-
system is the interaction design process, which defines how structions. Thus, some applications and websites already
the recommender system interacts with a user. It also shapes provide support functions allowing users to browse recipes
the experience the user gets, from the point she registers and related information. However, most applications only
and provides her preferences to the system, to the point offer generic and non-personalized recipe catalogue browsing
she receives recommendations generated by the system. A support, without tailoring it to the tastes and preferences of
proper interaction design may improve user experience and individual users.
hence may result in higher usability of the system, as well User preference elicitation is a fundamental and necessary
as, in higher satisfaction. step to go beyond this generic support and generate person-
In this paper, we focus on the interaction design of a mo- alized recipe recommendations. More importantly than in
bile food recommender system that, through a novel interac- other application domains, such as movies or books, recipe
tion process, elicits users’ long-term and short-term prefer- recommendations should not only be based on user’s long-
ences for recipes. User’s long-term preferences are captured term tastes, but also fit their ephemeral preferences, such as
by asking the user to rate and tag familiar recipes, while for the available ingredients or current cooking constraints.
collecting the short-term preferences, the user is asked to In this paper we address this problem by proposing a pref-
select the ingredients she would like to include in the recipe erence elicitation approach for food recommender systems
to be prepared. Based on the combined exploitation of both that obtains user preferences through a novel and effective
types of preferences, a set of personalized recommendations interaction design. First, it exploits an integrated Active
is generated. We conducted a user study measuring the us- Learning algorithm [5, 6] for selecting the recipes to rate
ability of the proposed interaction. The results of the study and tag that are estimated to be the most useful for the rec-
show that the majority of users rates the quality of the rec- ommender. The active learning algorithm scores a recipe ac-
ommendations high and the system achieves usability scores cording to its predicted its rating (using transformed matrix
above the standard benchmark. of user-recipe) and then selects the highest scoring recipes.
This reveals the users’ long-term preferences, i.e., what they
usually like to eat or cook. Second, when requested to gener-
1. INTRODUCTION ate recommendations, the system acquires short-term pref-
Recommender systems are decision support tools that proac- erences referring to ingredients the user wants to cook or to
tively identify and suggest items, which are expected to be include in the meal. The acquired preferences are used by a
interesting for the users. Recommendations are based on Matrix Factorization (MF) rating prediction model designed
the users’ previous interactions with the system and the ex- to take into account both tags and ratings [11, 13, 7].
plicitly provided users’ preferences [15]. One important and In a real user study, we evaluated the proposed prefer-
new application domain for recommender systems is food. ence elicitation interaction and observed that the users have
This application has recently drawn much attention in the scored the usability of the system between “good” and “ex-
research community due to its potential to improve eating cellent” and assessed the presented recommendations, which
behaviour of users and positively influencing their lives [8, are generated on the basis of the elicited preferences, to be
of high quality.
Thus, the main contributions of our paper are: (a) a novel
interaction design that is used to elicit long-term (general)
and short-term (session-based) user preferences; and (b) an
effective preference elicitation method that exploits active
learning in the food recommendation domain.
.
Figure 1: (a) user instructions, (b) browsing food categories, and (c) selecting eaten or cooked recipes.
2. RELATED WORK building a real-world recommender system. To address this
Several recommender systems for the food domain have limitation, this paper focuses on the interaction design, mainly
recently been developed [9, 18, 19, 20]. For example, Freyne for the preference elicitation: long-term and session-based.
and Berkovsky [9] proposed a food recommender that, through
an easy-to-use interface, elicits user preferences and provides 3. USER-RECOMMENDER INTERACTION
personalized recommendations Their system transferred the
recipe ratings collected by the system to ingredient ratings We designed a complete human-computer interaction for
and then aggregated the ratings of the ingredients used in a collecting user preferences, in the form of recipe ratings and
recipe to generate rating predictions. tags [4]. An Android-based prototype was developed, in or-
Elahi et al. [4] proposed a food recommendation model der to implement this interaction. The first step is a general
that combines the predicted value of a recipe along differ- preference elicitation, aimed at collecting the long-term (sta-
ent dimensions (user food preferences, nutritional indicators, ble) user preferences, i.e., what she generally likes to cook
and ingredients costs) to compute a single utility measure (or eat). This step includes two stages: (1) the system asks
of a recipe. The goal is to consider factors influencing the the user to specify the recipes she cooks at home and, (2) the
user’s food decisions in order to produce more useful and user assigns ratings and tags to the recipes she experienced.
valuable recommendations. In a follow-up work [11], the au- Upon logging in the system, the user can browse the full
thors conducted an offline evaluation of the rating prediction catalogue of recipes and mark those that she has eaten be-
algorithm, which extends MF by using, in addition to rat- fore (see Figure 1). Users can navigate through the recipe
ings, the users’ tags assigned to recipes. It was shown that categories and sub-categories in order to find the desired
this additional source of information about the user pref- recipe, e.g., ‘Beef’ → ‘Roasted Beef’ → ‘Roasted Beef with
erences allowed the proposed method to outperform other Salad’. Inside each category there is a list of recipes mapped
state-of-the-art algorithms, e.g., those proposed in [10]. to this category. When the user finds one of them she can
In general, the user’s preferences that are collected and mark it as ‘Eaten’or ‘Cooked’ by clicking the check box.
used by a recommender can be either long-term (general After that, a selection of the recipes that the user marked
preferences) or short-term (session-based and ephemeral). as eaten or cooked, is presented to the user for rating and
While obtaining both preference types is crucial, many rec- tagging. This allows the system to acquire knowledge about
ommender systems do not distinguish between the two. In the general user preferences. However, the system also needs
fact, there are few studies that taken this consideration into to deeper explore the user’s preferences and it presents addi-
account. Ricci and Nguyen proposed in [14] a mobile rec- tional recipes for the user to rate and tag. These are found
ommender system in travel domain, which elicits both gen- by predicting what the user might have eaten, but did not
eral long-term preferences (e.g., explicitly defined by users) mark in the first step. In order to find such recipes, we use
and short-term preferences in the form of critiques express- active learning. For this, the rating dataset is transformed
ing more detailed session-based preferences. More recently, into a binary format indicating only whether the user rated
short term preferences were found to depend on the rec- an item: null entries are mapped to 0, and not null entries
ommendation context and many context-aware approaches to 1. Then, using a factor model, predictions are computed
have been proposed to better suit the needs of the users [1]. for all the values mapped to 0, and for each user the items
It is worth noting that RSs research often focused on the with the highest prediction are shown to the user [5, 6].
improvement of the prediction model, by assuming that the Figure 2-a shows the rating and tagging interface. This
preference elicitation process is completed. Hence, they interface uses the classical 5-star Likert scale. The users
ignore the complete user-system interaction, required for are also requested to “explain” the core motivations for their
ratings by assigning tags to recipes. Users can either tag a
Figure 2: (a) general preference elicitation, (b) session-based preference elicitation, and (c) recommendation.
recipe with the suggested tags or add their own tags. At the
recommendation time, session-specific preference are elicited
(see Figure 2-b). The user enters the core ingredient she
wants to include in the recipe. This is done by selecting
a keyword from the list of suggestions derived from food
ingredients and popular tags assigned by other users.
Then, the recommendations leverage both types of the col-
lected user preferences, long-term and session-specific. The
long term preferences are exploited by a custom MF rating
prediction model [13], which uses the tagging information
Figure 3: SUS results.
[7]. Each user is associated with a vector that models her
latent features and each recipe is modeled by a vector that
contains its latent features. Then, the rating of a user for mark value is 68, which is the average SUS score computed
an item is predicted by computing the inner product of the over 500 usability studies [16].
user and item vectors. To exploit the short-term model, In our experiment 20 subjects used the system and com-
the system post-filters the recommendations according to pleted the questionnaire. They were either computer science
the current user preferences. The recipes with the high- researchers or non-academic people. 60% of subjects were
est rating are presented to the user one by one. When the male and 40% were female, the age range was 23 to 50, and
user selects a recommended recipe, the system presents the the ethnical background varied across the subjects (Italy,
required ingredients and detailed cooking instructions (see France, USA, Germany, China, and more).
Figure 2-c). We first present the perceived recommendation quality re-
sults. The survey measures the recommendation quality us-
4. USER STUDY ing 7 questions on a Likert scale from 0 to 4, where 4 is
the highest score. Thus, the maximum overall quality score
The main goal of the evaluation was to assess whether the
is 28. The average perceived recommendation quality score
system can effectively assist users in finding recipes that suit
across the 20 subjects was 19 and the median was 19 (see
their preferences. For the user study, we designed a usage
[3] for more details on the calculation). We observed that
scenario a task that was formulated as follows: “You want to
the maximal recommendation quality score was 26, and the
avoid everyday routine meals. You can use this application
minimal was 12. Thus, we can conclude that, on average,
to discover new recipes that suit your taste”.
the users agreed that the recommendations were well-chosen
The users were asked to use the mobile application and
and suited their preferences.
complete a questionnaire referring to two performance in-
For the SUS usability score, we observed that for 75% of
dicators: perceived quality of recommendations quality and
subjects the SUS score was higher than the 68 point bench-
system usability. The first part of the questionnaire mea-
mark (see Figure 3). The system achieved overall average
sured the level of user satisfaction with the recommenda-
SUS score of 75.50 and the median was 73.75, which is well
tions. We used a validated instrument based on a set of
above the benchmark. We observed that the minimal us-
questions developed by Knijnenburg et al. [12]. The second
ability score of 55, and the maximal was 95. According to
part of the questionnaire aimed at collecting the users’ im-
these results we can conclude that the system usability was
pression of the usability of the system. Here, we exploited
considered between “good” and “excellent” [2].
the System Usability Scale (SUS) questionnaire [17]. The
overall usability scores range from 0 to 100 and the bench-
We have computed the average replies for all the SUS [6] M. Elahi, F. Ricci, and N. Rubens. Active learning in
statements and observing the statements with the highest collaborative filtering recommender systems. In
average values, we can report that the users have evaluated E-Commerce and Web Technologies, pages 113–124.
the system easy to learn and easy to use. They also believe Springer International Publishing, 2014.
that various components were well-integrated into the sys- [7] I. Fernández-Tobı́as and I. Cantador. Exploiting social
tem. On the other hand, by observing the statements with tags in matrix factorization models for cross-domain
the lowest values we can state that users think that they collaborative filtering. In Proceedings of the 1st
have to learn a lot before they can use the system properly Workshop on New Trends in Content-based
and they may need technical person for that. Our explana- Recommender Systems, Foster City, California, USA,
tion for this result is that we need to improve further the pages 34–41, 2014.
interface and provide more explanations, so that users can [8] J. Freyne and S. Berkovsky. Intelligent food planning:
better learn and understand the usage of the components in personalized recipe recommendation. In IUI, pages
the system. 321–324. ACM, 2010.
[9] J. Freyne and S. Berkovsky. Intelligent food planning:
5. CONCLUSION AND FUTURE WORK personalized recipe recommendation. In IUI, pages
In this paper, we illustrated the preference elicitation pro- 321–324. ACM, 2010.
cess of a novel food recommender system [11]. Our system [10] J. Freyne and S. Berkovsky. Evaluating recommender
generates recommendations by exploiting tags and ratings systems for supportive technologies. In User Modeling
in a MF algorithm. In our study, we collected user evalu- and Adaptation for Daily Routines, pages 195–217.
ations of the recommendation quality and system usability. Springer, 2013.
Both measurements were found to be positive. This means [11] M. Ge, M. Elahi, I. Fernaández-Tobı́as, F. Ricci, and
that the proposed preference elicitation process and system D. Massimo. Using tags and latent factors in a food
interaction are liked by users. recommender system. In Proceedings of the 5th
Considering that this is a preliminary study, this paper International Conference on Digital Health 2015,
has several limitations. First, the evaluation is performed pages 105–112. ACM, 2015.
on the whole system rather than on preference elicitation. [12] B. P. Knijnenburg, M. C. Willemsen, Z. Gantner,
Since the prediction model was already tested in another H. Soncu, and C. Newell. Explaining the user
study [11], this work mostly focuses on preference elicita- experience of recommender systems. User Modeling
tion as the main component of user interaction. Second, and User-Adapted Interaction, 22(4-5):441–504, 2012.
we have not compared our system with alternative prefer- [13] Y. Koren and R. Bell. Advances in collaborative
ence elicitation processes. Our current result mostly reflects filtering. In F. Ricci, L. Rokach, B. Shapira, and
the users’ direct perception of their interaction with the sys- P. Kantor, editors, Recommender Systems Handbook,
tem. Third, we admit the limited number of subjects in the pages 145–186. Springer Verlag, 2011.
user study. In the future, we plan to increase the number [14] F. Ricci and Q. N. Nguyen. Acquiring and revising
of participants in the study. Also, we plan to extend the preferences in a critique-based mobile recommender
recommendation model by considering nutritional factors, system. Intelligent Systems, IEEE, 22(3):22–29, 2007.
e.g., the required calories and proteins, in order to build a [15] F. Ricci, L. Rokach, and B. Shapira. Introduction to
health-aware recommender system. recommender systems handbook. In F. Ricci,
L. Rokach, B. Shapira, and P. Kantor, editors,
6. REFERENCES Recommender Systems Handbook, pages 1–35.
[1] G. Adomavicius and A. Tuzhilin. Context-aware Springer Verlag, 2011.
recommender systems. In Recommender systems [16] J. Sauro. Measuring usability with the system
handbook, pages 217–253. Springer, 2011. usability scale (sus).
[2] A. Bangor, P. Kortum, and J. Miller. Determining http://www.measuringusability.com/sus.php.
what individual sus scores mean: Adding an adjective Accessed: 2013-01-15.
rating scale. Journal of usability studies, 4(3), 2009. [17] J. Swarbrooke and S. Horner. Consumer behaviour in
[3] M. Braunhofer, M. Elahi, F. Ricci, and T. Schievenin. tourism. Routledge, 2007.
Context-aware points of interest suggestion with [18] C.-Y. Teng, Y.-R. Lin, and L. A. Adamic. Recipe
dynamic weather data management. In Information recommendation using ingredient networks. In
and Communication Technologies in Tourism 2014, Proceedings of the 4th Annual ACM Web Science
pages 87–100. Springer International Publishing, 2014. Conference, pages 298–307. ACM, 2012.
[4] M. Elahi, M. Ge, F. Ricci, D. Massimo, and [19] M. Trevisiol, L. Chiarandini, and R. Baeza-Yates.
S. Berkovsky. Interactive food recommendation for Buon appetito: recommending personalized menus. In
groups. In Poster Proceedings of the 8th ACM Proceedings of the 25th ACM conference on Hypertext
Conference on Recommender Systems, RecSys 2014, and social media, pages 327–329. ACM, 2014.
Foster City, Silicon Valley, CA, USA, October 6-10, [20] R. West, R. W. White, and E. Horvitz. From cookies
2014. 2014. to cooks: Insights on dietary patterns via analysis of
[5] M. Elahi, F. Ricci, and N. Rubens. Active learning web usage logs. In Proceedings of the 22nd
strategies for rating elicitation in collaborative international conference on World Wide Web, pages
filtering: A system-wide perspective. ACM 1399–1410. International World Wide Web
Transactions on Intelligent Systems and Technology Conferences Steering Committee, 2013.
(TIST), 5(1):13, 2013.