CATS: A Synchronous Approach to Collaborative Group Recommendation
Kevin McCarthy, Maria Salamo´, Lorcan Coyle, Lorraine McGinty, Barry Smyth & Paddy Nixon
Adaptive Information Cluster, School of Computer Science & Informatics
University College Dublin Belfield, Dublin 4, Ireland.
{kevin.mccarthy, maria, lorcan.coyle, lorraine.mcginty, barry.smyth, paddy.nixon}@ucd.ie
Abstract
Group recommender systems introduce a whole set of
new challenges for recommender systems research. The
notion of generating a set of recommendations that will
satisfy a group of users with potentially competing in-
terests is challenging in itself. In addition to this we
must consider how to record and combine the prefer-
ences of many different users as they engage in simulta-
neous recommendation dialogs. In this paper we intro-
duce a group recommender system, called CATS, that is
designed to provide assistance to a group of friends try-
ing the plan a skiing vacation. The system uses the Dia-
mondTouch interactive tabletop to allow up to 4 users to
simultaneously engage in parallel recommendation ses-
sions and we describe how personal and shared profiles
and interaction spaces can be managed to generate sets
of recommendations for the individual and the group.
Introduction
The paper describes a novel conversational, collaborative
group recommender system called CATS (Collaborative Ad-
visory Travel System), designed to help a group of up to 4
friends plan and arrange their skiing vacation. This system
is designed around the DiamondTouch interactive tabletop.
The CATS system is based around the notion of a shared
collaborative space for a group of users who also can access
their own personal spaces. Individual user feedback in used
to update explicit user models, on a per user basis, as well as
a global user model. In addition, recommendations for the
individual are generated in response to direct user feedback
while at the same time group recommendations are gener-
ated proactively through the shared interaction space.
Before describing the details of the CATS framework, we
begin with a background review that looks at related work
in recommender systems research as well as providing an
overview of the DiamondTouch device. Following this we
detail the CATS framework focusing on the core interface
and recommendation components and providing an example
walk through of a particular recommendation session.
Copyright c© 2006, American Association for Artificial Intelli-
gence (www.aaai.org). All rights reserved.
Research Background
In this section we discuss related research to this project in
the areas of collaborative/co-operative recommendation and
multi-user interfacing. We are especially interested in con-
versational content-based recommender systems where cri-
tiquing is appropriate as a user-feedback strategy. We also
introduce the DiamondTouch interface device used by the
CATS framework to showcase our simultaneous collabora-
tive group critiquing recommender system.
Group Recommendation
Group decision making has long been a topic of research
in distributed AI and multi-agent systems. The notion of
an agent or a multi-agent system interacting, collaborat-
ing or negotiating in an environment has been previously
addressed by several researchers in the agent community.
Within the area, the applications range from virtual envi-
ronments (Prada & Paiva 2005) to sales by action (Faratin,
Sierra, & Jennnings 2002). The vast majority of these sys-
tems assume an automated negotiation that is based on some
static preferences previously defined in the system. How-
ever, in our case, we have human users with individual (often
different) initial preferences, and usually these preferences
change as the recommendation session progresses. There-
fore, a static model of preferences is inappropriate and an al-
ternative method for modelling these preferences (and pref-
erence conflicts) is required. Moreover, preference mod-
elling needs to not only capture individual user preferences
but also an effective way of modelling the groups evolving
preferences as a whole is required.
Other research in the area of group recommendation
includes the MUSICFX System (McCarthy & Anagnost
1998). MUSICFX is a group preference arbitration sys-
tem that adjusts the selection of music playing in a fitness
center to best accommodate the musical preferences of the
people working out at any given time. The preferences
have been previously specified by the members who are cur-
rently working out. POLYLENS (O’Connor et al. 2001) is a
generalization of the MOVIELENS system that recommends
movies to group of users. In that case, the recommender
is based on collaborative filtering which uses the history of
preferences of past users in similar situations. Another ex-
ample, is the TRAVEL DECISION FORUM (Jameson 2004)
prototype which helps a group of users to agree on the de-
86

sired attributes of a vacation that they are planning to take
together. Special attention is given to support for users who
are not collocated and who can therefore not engage in face-
to-face discussions. Plua and Jameson previously worked on
a group travel recommender system (Plua & Jameson 2002)
where users can get help from others in their group about
preferences when their domain knowledge may be incom-
plete. However this system was intended for use by a group
that interact asynchronously rather than simultaneously.
Critiquing Based Recommenders
We are especially interested in a form of user feedback
called critiquing (McGinty & Smyth 2003), where a user
indicates a directional feature preference in relation to a pre-
sented recommendation. For example, in a travel vacation
recommender, a user might indicate that they are interested
in a vacation that is longer than the currently recommended
option; in this instance, longer is a critique over the du-
ration feature. Within recommender systems literature the
basic idea of critiquing can be traced back to the seminal
work of Burke et al. (Burke, Hammond, & Kozlovsky 1995;
Burke, Hammond, & Young 1997). Entre´e is a restaurant
recommender system that employs critiquing to allow users
to refine restaurant features such as price, style, atmosphere,
etc. The advantage of critiquing is that it is a very low-cost
form of feedback, in the sense that the user does not need to
provide specific feature values, while at the same time help-
ing the recommender to narrow its search focus quite signif-
icantly (McGinty & Smyth 2003). Recently, there has been
renewed interest in critiquing, as recommender systems be-
come more commonplace, and a number of enhancements
have been proposed to the basic critiquing approach. For
instance, an improved approach (Reilly et al. 2004) is to
consider a user’s critiquing history, as well as the current
critique, when making new recommendations. Given that
this approach has been shown to deliver significant improve-
ments in recommendation efficiency (McCarthy et al. 2005),
it is the assumed method of feedback by the CATS frame-
work.
However, the conventional implementation of critiquing
can pose problems in a group recommender setting because
each user represents two roles within the system at the same
time: their individual role, and implicitly their membership
to the group role. Some of the individual user preferences
may conflict with the group preferences. For example, our
user might have received a recommendation for a luxury 2-
week package in Spain for 2000. She might be interested
in something around the 1500 mark and so may indicate
that she wants a cheaper recommendation. However, the
group preference is for a more expensive recommendation
because they want a luxury package holiday. There is little
to be gained from conflicting interactions between individ-
ual and group preferences, so the coordination of both roles
becomes one of the most important challenges for a collab-
orative recommender. Another key challenge is how best
to make interacting users aware of other user’s preferences.
In this paper we propose a simple approximation to average
the interaction of the individual and the groups roles within
a group recommender. We also describe how we communi-
cate the evolving preferences of the group to each of the par-
ticipants in order to facilitate convergence on a single recom-
mendation through dynamic visualization, and a combina-
tion of proactive and reactive recommendation techniques.
We also discuss some lessons we have learned from our ini-
tial design, and discuss avenues for future research.
The DiamondTouch
The majority of collaborative applications involve sep-
arate (often distributed) workspaces, however, with
technologies such as the DiamondTouch and Mimio
(http://www.mimio.com ) it is possible to allow users to
work together co-operatively around a common workspace.
The DiamondTouch (see Figure 1) supports small group col-
laboration by providing a display interface that allows users
to discuss decisions openly whilst interacting with the dis-
play simultaneously (i.e., without having to take turns) (Di-
etz & Leigh 2001). It consists of a touch sensitive table con-
nected to a computer whose display is projected onto the
table. The table can detect and distinguish between simul-
taneous touch events, allowing the development of innova-
tive and intuitive collaborative and cooperative applications.
It enables up to four users interface with the same touch-
surface in a very simple fashion, and is capable of discrimi-
nating between multiple simultaneous interactions.
Figure 1: The DiamondTouch interactive table-top device.
Research has shown that the DiamondTouch is a more ef-
fective interface for solving certain collaborative problems
than the two-mouse and one-monitor alternative (Kobourov
et al. 2005). It is a natural interface for the collaborative
task where friends need to book a skiing holiday together.
Its ‘coffee table’ form factor, and intuitive flat orientation,
allow users to easily and co-operatively search the space of
options and at the same time understand the preferences of
the other participants. We propose to use the DiamondTouch
tabletop device to showcase our new synchronous collabo-
rative group recommender system.
Collaborative Recommendation
Our approach to group recommendation is based on a collab-
orative recommender framework that, at the interface layer,
assumes the availability of individual and group interaction
spaces and at the recommendation layer, assumes a recom-
mendation engine that is able to record and manage personal
87