Proximo, Location-Aware Collaborative Recommender

Abstract

Pervasive computing systems typically rely on a range of context data, from user preference and interaction through to the sensed environment. In Proximo we propose an approach to combine a collaborative recommendation system, with location-aware technology to provide personalised, dynamic and domain-driven paths through work and social spaces. For our purposes, context includes user preferences, current location, importance measures for digital artifacts and a community of other users' measures. We report on an ethnographic study of two art galleries, the architecture of a privacy-centric proximation and recommendation system, an exploratory user study and results from a gallery tour application.

Full text

Proximo, Location-Aware Collaborative Recommender
Eoghan Parle and Aaron Quigley
UCD School of Computer Science & Informatics
University College Dublin
Belfield, Dublin 4, Ireland.
{eoghan.parle, aquigley}@ucd.ie
Abstract. Pervasive computing systems typically rely on a range of context
data, from user preference and interaction through to the sensed environment.
In Proximo we propose an approach to combine a collaborative
recommendation system, with location-aware technology to provide
personalised, dynamic and domain-driven paths through work and social
spaces. For our purposes, context includes user preferences, current location,
importance measures for digital artifacts and a community of other users’
measures. We report on an ethnographic study of two art galleries, the
architecture of a privacy-centric proximation and recommendation system, an
exploratory user study and results from a gallery tour application.
1 Introduction
The selection of relevant information for a person is constrained by the ability of
the system to determine that person’s context. Typically we provide explicit “context”
through preferences, selection or search criteria. Examples of implicit actions
(context) include selecting a particular file or opening a specific website. Context
includes information from the sensed environment (environmental state) and
computational environment (computational state) which is provided to alter an
application's behavior in relation to the context. However, the context for ones human
computer interaction includes implicit context in addition to users interactions and
their peer’s (within a given community) interactions.
Context-aware computing aims to leverage the entire gamut of physical and
digital interactions people and their peers have with computer systems. A major factor
for systems that are context-aware is the ability to know users' location as this dictates
much for the systems understanding about current activity. However, location is but
one small part of a context-aware system. We propose Proximo; a system that merges
location-aware computing and recommendation techniques to create socially minded
applications (guides) for use within buildings such as museums, art galleries or
hospitals.

2 Eoghan Parle and Aaron Quigley
2 Background
One aspect of context aware computing is location-awareness. Location or
positioning systems operate within a clearly defined area of operability with a
resolution and performance usually defined by the characteristics of the technology
supporting the system. Location-aware applications include finding services such as
printing or telephones or tracking individuals [2,11,16]. Here we would like to focus
on systems that support context-aware use within buildings such as museums, art
galleries or hospitals.
A range of approaches to indoor location use specially designed infrastructure.
These include approaches using Infrared in the Active-Badge [8], 802.11 RADAR [5]
and RF/Ultrasound in Cricket [6]. Each approach can locate a user quite accurately
within an indoor environment but would not work outside or in any area not fitted
with a plethora of sensors. Approaches such as BlueStar [12] attempt to utilize
existing infrastructures to realize indoor location-awareness (proximation). Location
is deduced by handset/PDA-resident (mobile terminal) applications that have two
sources of information, passive sniffing of existing wireless infrastructure (Bluetooth
or 802.11b) and details of the local wireless infrastructure provided based on the
GSM knowledge of the user’s approximate location from the networks positioning
system.
2.1 Recommendation-Based Systems
Recommendation systems provide tailored suggestions to the user based on an
understanding of the content or from a view of the collective group or community
which this user fits into or even a hybrid “boosted” approach. Content-based
recommendation uses similar techniques to information retrieval. A comparison
between the user profile and the content of the objects is the basis of a
recommendation. If it can be seen that a user has rated highly a set of objects with
similar content then it may be determined that that particular user would probably be
interested in anything which contains more similar content. NewsWeeder [3] and
InfoFinder [4] are examples of purely content-based recommendation systems.
Collaborative recommendation differs from content-based recommendation in that
it recommends items based on what similar users liked rather than what the individual
user liked. These systems sometimes define a set of ‘nearest neighbour’ users for each
user based on the correlation of past likes and/or dislikes. Ratings for unseen items
are based on a combination of the ratings from these neighbours. Pure collaborative
recommendation systems such as GroupLens [10,11] know nothing about the items
themselves, only about what users think of them. Collaborative recommendation
systems address some of the problems associated with content-based systems as these
systems can cope with any kind of content whether it is a web page, a piece of art or a
DVD. This is because these systems examine user ratings rather than content. They
can recommend items that are totally dissimilar to any seen before – as long as other
users have rated them.

Proximo, Location-Aware Collaborative Recommender 3
These systems do have their own shortcomings. If a new item is added to the
database then it will not be recommended until a user rates it. This might prove
difficult, as the user might not be able to find it unless it is recommended. If there are
a large number of items and a small number of users then a lot of items could end up
not being rated. Users with unusual tastes might also have problems with this type of
system as they might not have anyone close enough to their own tastes to give good
recommendations.
3 Ethnographic & Online Studies
To help explore our research questions on socially minded electronic guides [2] we
have adopted a domain-driven approach to our study. Our research involved a small
ethnographic study involving the curator of a family-run gallery (Cherrylane Fine
Arts Ireland) and two of archivists from the National Gallery of Ireland. These people
were questioned about a number of aspects of the day-to-day activities in their
respective galleries using a questionnaire. This questionnaire focused on who, what,
when, where & why questions to help elicit details on the ways people wish to
interact with the media (i.e. paintings) beyond simply looking at them. This work was
supplemented with an online study of many major art galleries such as the Tate
Modern London and the Louvre Paris, using a similar series of measures.
Our findings show that a majority of people, from a diverse set of backgrounds,
visiting either a small or major gallery tend to want more information about some of
the paintings they are viewing. There is an interest in leaving messages in both
galleries but at the moment these must be left on message cards or a message book.
These current methods do not allow for messages to be left in the vicinity of the
paintings and can not be accessed whilst viewing a painting. Online access to galleries
affords the opportunity to plan visits in the future where half of all visitors to the Tate
Gallery website do so to plan a visit [1]. This shows that there is an interest in using
emerging technologies to help personalise visits to art galleries.
4 Proximo
Proximo consists of a PC-based recommendation system (Figure 1b) using sample
paintings and a rating system in addition to an application running on a Java-enabled
Bluetooth a mobile phone as shown in Figure 1. The recommendation system collects
& stores data from multiple user interactions that is used to create user profiles based
on a weighted nearest neighbors algorithm [1]. The items with the highest predicted
ratings are those, which will be recommended to the user. Once the user has given an
initial rating the location and other domain-specific information relevant to these
items are transferred to the handset-resident application.
The indoor positioning of BlueStar works by 'sniffing out' the fixed Bluetooth
devices or low-cost beacons deployed in the area of use [12]. This room-level
accuracy means improved precision and is accurate enough for a certain class of trail-

4 Eoghan Parle and Aaron Quigley
based applications such as a tour guide. The mapping service on the mobile handset
displays a map of the intended area of use which can be manipulated in a variety of
ways including ZoomZones, Scrolling, Zooming and toggle selection. Items provided
by the recommendation system are displayed as icons over the map of the active area
on the mobile application. This provides both location information and an interface
for the user to select between the different items. The user can scroll between the
icons to view information about the corresponding item.
Figure 1: Proximo architecture
The mobile application constantly monitors the users location and displays the
active areas of the building (the area they are in) accordingly. The paintings on the
tour are also displayed in a different colour that highlights them from the others. The
first step in taking any action relating to a particular painting is to first select it. Any
action performed now will relate to the painting which is selected. To further aid the
users of the system a small image of the item is displayed in the top-right corner of
the map. Scrolling clockwise and anti-clockwise navigates around the group of items
(paintings).
When a painting is selected there are a number of actions which a user can take.
There is a facility for the user to provide a rating for the painting. Our messaging
feature akin to Stick-e Notes [7] and GeoNotes [9] where messages can be left at a
specific location and can only be accessed from that location within a certain context.
Proximo allows users to leave and receive messages at each painting they visit on
their tour (and those not on the tour too).

Proximo, Location-Aware Collaborative Recommender 5
5 User Study
A user study of the Proximo system was conducted to determine usability, assess
acceptance of the system and to measure the effectiveness of the recommendation
system. ‘The CSI Gallery’ was set up on the ground floor of the U.C.D. School of
Computer Science and Informatics building. Seven users were asked to complete a
tour of the gallery, each visiting a number of paintings which were suggested by the
recommendation system. The tour provided for each user was created by the
recommendation system. Each user was asked to rate (between 1 & 5) ten paintings
which were displayed in the gallery. These ratings were entered into the collaborative
recommendation system, that created the tour for each user.
Users were asked to complete fully their tour of the CSI Gallery answering each
question on a given question sheet. They were also asked to leave a rating for each
painting on their tour and leave at least three messages themselves (about anything
they liked) whilst on their tour. The last two tasks were to be completed using the
application running on the mobile phone. When the tour was completed each user was
asked to fill in a questionnaire which consisted of 28 multiple-choice questions which
were split up into three sections.
5.1 Results and Evaluation
In our user study we asked users to evaluate a range of aspects of the system
including, usability, recommendation quality and workload. The recommendation
system was tested by examining the predicted rating and actual rating for the
paintings on each users tour. The absolute difference or error between these two
values is used to measure the success of the predictions. The overall success of the
system is calculated by finding the mean absolute error over all the predictions made.
In this case the mean absolute error was 1.19.
Figure 2: User Interest in Information

6 Eoghan Parle and Aaron Quigley
Contrary to expectations there was a marginally better response to seeing
messages left by all users rather than just by similar users. It was suggested by one
user (obviously before he got to the question about similar users) that there be an
option to view messages left by only similar people. One user thought it would be
“hard to say if they would be helpful or not”. In general the users thought that having
the ability to see information left by different sources was a good idea. Information
left by more learned people are considered more desirable than information left by
just anyone.
6 Conclusions
We have explored the use of Proximo, a location-aware recommendation-driven
system for use within indoor environments such as museums and art galleries. Our
user study suggests that small amounts of both explicit and implicit context data can
enhance the quality of the users experience in trails through a physical space. The
notion that collective wisdom can guide a mobile experience based on the suggestions
of others has been demonstrated.
References
1. E. Parle, “Proximo: A Location-Based Recommendation-Driven Tour Guide”, School of
Computer Science and Informatics, University College Dublin Ireland 2005
2. G. D. Abowd, C. G. Atkeson, J. Hong, S. Long, R. Kooper, M. Pinkerton, “Cyberguide: A
mobile context-aware tour guide”, Wireless Networks 3, 5, p. 421–433, 1997
3. K. Lang, “NewsWeeder - Learning to filter Netnews”, In Proceedings of ICML-95,12th
International Conference on Machine Learning (Lake Tahoe,US,1995), p. 331 –339
4. B. Krulwich, C. Burkey, “The InfoFinder Agent - Learning User Interests through
Heuristic Phrase Extraction”, 1997
5. P. Bahl, V. N. Padmanabhan, “RADAR: An In-Building RF-based User Location and
Tracking System”, In Procedings of IEEE INFOCOM (Tel-Aviv, Israel, Mar. 2000)
6. N. B. Priyantha, A. Chakraborty, H. Balakrishnan, “The Cricket Location-Support
System”, MOBICOM, 2000
7. J. Pascoe, “The Stick-e Note Architecture: Extending the Interface Beyond the User”, In
Proceedings of International Conference on Intelligent User Interfaces 1997, p. 261-64,
1997
8. R. Want, A. Hopper, V. Falcao, J. Gibbons, "The Active Badge Location System", ACM
T. on Information Systems. Vol 10, No. 1 January 1992. pp. 91-102
9. F. Espinoza, P. Persson, A. Sandin, H. Nyström, E. Cacciatore, M. Bylund, “GeoNotes:
Social and Navigational Aspects of Location-Based Information Systems”, In Proceedings
of Ubicomp, Springer-Verlag Berlin Heidelberg 2001
10. J. Herlocker, J. Konstan, J. Riedl, “Explaining Collaborative Filtering Recommendations”,
In Proceedings of the 2000 Conference on Computer Supported Cooperative Work, 241–
250, 2000
11. J. Konstan, B. Miller, D. Maltz, J. Herlocker, L. Gordon, J. Riedl, “GroupLens: Applying
Collaborative Filtering to Usenet News”, Communications of the ACM (40) 3, 1997
12. A. Quigley, D. West, “Proximation: Location-Awareness Through Sensed Proximity and
GSM Estimation”, Procedings In First International Workshop, LoCA 2005, p.365-376