=Paper=
{{Paper
|id=Vol-1247/recsys14_poster2
|storemode=property
|title=Interactive Food Recommendation for Groups
|pdfUrl=https://ceur-ws.org/Vol-1247/recsys14_poster2.pdf
|volume=Vol-1247
|dblpUrl=https://dblp.org/rec/conf/recsys/ElahiGRMB14
}}
==Interactive Food Recommendation for Groups==
https://ceur-ws.org/Vol-1247/recsys14_poster2.pdf
Interactive Food Recommendation for Groups
Mehdi Elahi Mouzhi Ge Francesco Ricci
Free University of Free University of Free University of
Bozen-Bolzano, Italy Bozen-Bolzano, Italy Bozen-Bolzano, Italy
mehdi.elahi@unibz.it mouzhi.ge@unibz.it fricci@unibz.it
David Massimo Shlomo Berkovsky
Free University of CSIRO, Australia
Bozen-Bolzano, Italy shlomo.berkovsky@csiro.au
david.massimo@stud-inf.unibz.it
ABSTRACT [2]. This approach has the potential to be effective for het-
We present a prototype of a novel interactive food recom- erogeneous small groups, with members of different ages,
mender for groups of users that supports groups in planning having different tastes and preferences, e.g., families. Also,
their meals through a conversational process based on cri- the interactive nature of the recommendations can poten-
tiquing. The system comprises two novel elements: a user tially increase user engagement with the diet and make the
interface and interaction design based on tagging and cri- recommender more enjoyable for users.
tiquing, and a utility function incorporating healthiness and In summary, the contribution of this work is two-fold: (i)
diet compliance factors. a novel interface design that applies interactive strategies,
such as tagging and multi user critiquing, and (ii) a novel
recommendation algorithm that generates a long term diet
1. INTRODUCTION plan for users within the group.
Despite current technological progress, following a healthy
diet is still a challenge. While there are numerous informa- 2. APPROACH
tion resources such as books, websites, and mobile apps,
One of the main considerations related to food recom-
finding a suitable meal plan is not easy, mainly due to the
mendations refers to the recurrent nature of eating and food
majority of resources being generic and non-personalized.
consumption. Indeed, people eat similar times a day and ev-
As such, meals suggested by a diet may not appeal for users
ery day, and they plan their meals in a sequential manner.
and they will not stick to the diet guidelines. Food recom-
Although some works focused on sequential recommenda-
mender systems tackle this problem by generating personal-
tions [5, 6], to the best of our knowledge none of them has
ized meal plans. They exploit food and recipe data, explicit
been applied in the food domain. Likewise, the match of
and implicit food preferences (e.g., ratings and browsing be-
a recommended meal to a user cannot be measured on its
havior) to train predictive models and deliver personalized
own, but should rather be considered in the context of the
food recommendations to inform the meal plans [4].
entire meal plan, diet guidelines, and other nutritional fac-
Many of the existing works on food recommendations fo-
tors. Here, we consider several factors that affect the overall
cused on individual users. However, in real-life scenarios the
meal utility function. For user u, the utility of meal m at
food is consumed by a group of people, e.g., family, friends,
time t is quantified by
or school canteen. Moreover, it has been shown that small
groups of close users, such as families, have a profound role util(u, t, m) ∝ rat(u, m)+diet(plan(u, t1 , t2 ), m)+health(m)
and supportive effect on health promotion within the groups
[1]. In this work we propose a novel interactive mechanism where rat(u, m) denotes the rating of u for m, plan(u, t1 , t2 )
for food recommendation for groups. This elicits user prefer- is the set of meals consumed by u in a recent time window
ence for food through a conversational process incorporating [t1 , t2 ], diet(plan, m) denotes the compliance of m to the
rating and tagging [7]. The preferences are used to compute diet constraints with respect to plan, and, finally, health(m)
personalized predictions for individual users, which are ag- denotes the health score associated with m.
gregated into group-based recommendations [5]. The users Note that this computation disregards other factors that
can provide their critiques and refine the recommendations may affect util(u, t, m). Among these are cost(m) – the
estimated cost of cooking m, avail(u, m) – the availability
of the ingredients of m to u, and seq(plan, m) – the match
of m to food consumption patterns observed in plan. At
the current stage of our work, we focus primarily on rat,
highlight diet and health as two influential factors, and leave
the remaining factors for future research.
Once the utility of m is computed for every user u ∈
g in the group g, predicted score for the entire group is
Copyright is held by the author/owner(s).
RecSys 2014 Poster Proceedings, October 6-10, 2014, Foster City, Silicon quantified through aggregating
P the individual utility scores,
Valley, USA. as per util(g, t, m) ∝ u∈g util(u, t, m) [5].
Figure 1: Screenshots of the system prototype
3. PROTOTYPE 4. DISCUSSION AND FUTURE WORK
We are developing an Android prototype food recommender In this paper, we have presented the prototype of an inter-
for groups that is depicted in Figure 1. We will briefly active food recommender for groups. It features a novel util-
present the interfaces and the envisaged interactions. ity function, a tag- and rating-based preference elicitation
In the preference elicitation phase, we ask users to specify interface, group recommendation aggregation, and critique-
which meals they have ever eaten or cooked (see Figure 1- based conversational recommendations.
left). This is done by presenting a tree of meals, e.g., pasta We are now implementing the proposed system and work-
→ spaghetti → spaghetti bolognese. Users can navigate the ing on the user study design. In the future, we plan to
tree and mark familiar meals. Next, users are asked to rate apply a mechanism that considers the role of users in the
and tag meals selected from: (1) familiar meals marked by group when aggregating recommendations. This is due to
the user, (2) meals selected by an active learner [3], and (3) the higher impact of the preferences of dominant users (e.g.,
set of popular meals. The rating interface uses a 5-star Lik- mother in the family) on the decisions of other group mem-
ert scale (see Figure 1-middle). In addition, the users can bers (e.g., kids). We also plan to weigh differently the pre-
explain their ratings with tags extracted from the recipes dictive rating and critiques of the group members. A group
and online resources. The tags can be meal characteristics, member with strict diet constraints like diabetes, may can
recipe ingredients, or free user feedback [7]. Users can se- be assigned a higher importance than other members.
lect the suggested tags, add their own tags, and associate
positive or negative attitude with the selected tags. 5. REFERENCES
Given the rating and tagging input, we populate the item×
[1] N. Baghaei, S. Kimani, J. Freyne, E. Brindal,
tag matrix and incorporate the attitude into the matrix.
S. Berkovsky, and G. Smith. Engaging families in
Then, an extended matrix factorization algorithm (with con-
lifestyle changes through social networking. Int. J.
tent and tag features) is run to compute the predicted rating
Hum. Comput. Interaction, 27(10):971–990, 2011.
rat(u, m). The computation of diet(plan, m) and health(m),
as well as of other factors affecting util(u, t, m), depends on [2] L. Chen and P. Pu. Critiquing-based recommenders:
the constraints of the user’s diet and nutritional considera- survey and emerging trends. User Model. User-Adapt.
tions and is left beyond the scope of this paper. Interact., 22(1-2):125–150, 2012.
Apart from the individual rating and tagging elicitation, [3] M. Elahi, F. Ricci, and N. Rubens. Active learning
we also use group-based preference elicitation, in which we strategies for rating elicitation in collaborative filtering:
differentiate the role of the group leader, called the cook. A system-wide perspective. ACM Trans. Intell. Syst.
Based on the tags and ratings of all the group members, and Techn., 5(1):13, 2013.
the system delivers to the cook the meal recommendation [4] J. Freyne and S. Berkovsky. Intelligent food planning:
using the group utility score (Figure 1-right). The cook can personalized recipe recommendation. In IUI, pages
either accept the recommendation, reject it outright (e.g., 321–324. ACM, 2010.
due to unavailability of ingredients), or criticize it using the [5] J. Masthoff. Group modeling: Selecting a sequence of
tags (e.g., “less spicy”). Similarly, each group member can television items to suit a group of viewers. User Model.
criticize the meal accepted by the cook through a series of and User-Adapt. Interact., 14(1):37–85, 2004.
interactions with the recommender. The system then aggre- [6] A. Piliponyte, F. Ricci, and J. Koschwitz. Sequential
gates individual critiques into a compound critique by ap- music recommendations for groups by balancing user
plying a voting mechanism and resolving possible conflicts, satisfaction. In UMAP Workshops, 2013.
and recommends alternative meals [2]. The critiquing cycle [7] J. Vig, S. Sen, and J. Riedl. The tag genome: Encoding
continues until all the group members accept the recommen- community knowledge to support novel interaction.
dation. ACM Trans. Interact. Intell. Syst., 2(3):1–44, 2012.