CONTEXT-AWARE HOARDING OF MULTIMEDIA CONTENT IN A

LARGE-SCALE TOUR GUIDE SCENARIO

A Case Study on Scaling Issues of a Multimedia Tour Guide

J. K¨opke, R. Tusch, H. Hellwagner and L. Bszrmenyi

Department of Information Technology, Alpen Adria University Klagenfurt, Universittsstrae 65-67, Klagenfurt, Austria

Keywords:

Multimedia tour guide, context awareness, hoarding, multimedia data distribution.

Abstract:

This paper discusses scaling issues of a mobile multimedia tour guide. Making tourist-information available

in a substantially large geographical area (e.g. a federal state in Austria) raises new questions, compared to

providing similar information in a limited area (such as a museum). First, we have to assume a heterogeneous

network infrastructure containing high and low bandwidth links and even total network loss. Video streaming

is therefore not possible at any place. Secondly, the total amount of data grows linearly to the number of

Points of Interest (POIs) which are augmented by the tour guide. Therefore, a preloading of all data onto a

device with limited storage is not possible. A possible solution to these problems is hoarding, i.e. preloading

an ”appropriate” subset of data. The crucial question is to ﬁnd the proper subset in dependence of the actual

context. The paper discusses the questions of (1) what kind of context information should be considered

and (2) what kind of usage patterns can be assumed. Based on these considerations hoarding strategies are

developed for the tour guide. The strategies are ﬁnally evaluated with real-world data from a federal state wide

tourist-card system.

1 INTRODUCTION

Single site mobile tour guides have been a research

topic for the last years (Kray and Baus, 2003). Proto-

type as well as productive systems have shown their

abilities to assist tourists in different scenarios. Such

tour guides can (depending on their architecture) be

easily scaled from a single site or a city with good in-

frastructure to a larger region if they do not depend

on multimedia material and the total amount of re-

quired data is therefore low. If a tour guide is intended

to heavily use multimedia data and should be scaled

over a large area, such as a whole federal state, current

mobile storage systems as well as today’s available

network technologies do not allow a straight forward

scalability of such a system.

An example of a single site tour guide with intensive

usage of video and audio material was introduced in

the MultiMundus project (Kropfberger et al., 2007).

The current work utilizes the results of the Multi-

Mundus project and focuses on scaling such a tour

guide from a single site to the whole federal state of

Carinthia. On the network part the main problems

arise due to networks with much too low bandwidths

like GSM and GPRS as well as total lack of network

access in rural areas. Because Carinthia is located in

the Alps, network connectivity is even lower than in

other areas due to the alpine shape of the landscape.

Points of interests (POIs) for tourists are often located

in alpine environment. Therefore the tour guide must

be usable at such places as well. Moreover, often

many people (e.g. a group of tourists) use the tour

guide in one place at the same time. A naive solution

would result in the demand of very high bandwidths

in areas where cell-phone networks are intended to be

used by few people.

A more sophisticated solution should transfer the

needed multimedia data onto the device before the

user enters an area with low bandwidth or even un-

availability of a network. This could be realized by a

sufﬁciently large storage in each mobile device. If it is

possible to store all data in advance no network access

will be required at all. This is a really feasible solution

for a smaller scenario like a single open-air museum

or leisure park. The feasibility relies on the obser-

vation that updates on multimedia data are not very

frequent, since video and audio material is usually

produced by a production company once per season.

However,if storage capacity is limited - i.e. the device

cannot store all data of the region - the multimedia

Köpke J., Tusch R., Hellwagner H. and Bszrmenyi L. (2008).

CONTEXT-AWARE HOARDING OF MULTIMEDIA CONTENT IN A LARGE-SCALE TOUR GUIDE SCENARIO - A Case Study on Scaling Issues of a

Multimedia Tour Guide.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 15-22

DOI: 10.5220/0001933200150022

 SciTePress

material must be intelligently distributed to the mo-

bile clients. This means that at times when the users

do have an access to a high-bandwidth network (e.g.

at a WLAN hot spot), the system must ensure that the

most probably required data is stored on the device.

Such data should be transferred in advance. This tech-

nique is known as hoarding (Tait et al., 1995).

In addition to multimedia data describing the objects

of interest, other data like information about events

and weather forecasts are also needed. Such data can

nearly always be transferred over cellular networks

and this issue is therefore out of scope.

A tour guide can be seen as a classical context aware

system. This leads to the idea that the data which

should be present on the mobile device depends on

the usage context (i.e. the user’s current location

and interests). Therefore, we introduce and eval-

uate context aware hoarding strategies to overcome

the limitations of network and storage of current mo-

bile devices to realize a state-wide tour guide in

Carinthia. The hoarding strategies are evaluated with

real tourist movement patterns derived from log data

of the Carinthian ”tourist card” operator.

The remainder of this work is organized as fol-

lows. In Section 2 we present the single-location mul-

timedia tour guide MultiMundus. In Section 3 the

problems of scaling the tour guide are discussed in

depth. Section 4 discusses related work. In Section 5

hypotheses for a state-wide tour guide are presented.

Section 6 introduces the hoarding model used for the

tour guide. In section 7 the evaluation environment is

presented and the results of the evaluation are shown.

Finally, in Section 8 the results of the evaluation are

used to make a recommendation for the multimedia

data distribution of the tour guide.

2 The M3-Guide

The M3-Guide (Kropfberger et al., 2007) is a

multimedia-based guidance system for various con-

sumer devices. The system has a reference implemen-

tation called MultiMundus in the leisure park Min-

imundus

. Minimundus is a park with miniatures

of famous buildings all over the world. It is also

called ”The small world at Lake W¨orthersee”. One

can ﬁnd miniatures of the Eiffel Tower, the Sydney

Opera house and many more. The M3-Guide pro-

vides video and audio information about these minia-

tures. The system is context aware in the dimensions

User-Proﬁle, Device-Proﬁle, and Location. The pre-

sentation is rendered according the user’s language,

Minimundus GmbH - http:// www.minimundus.at

the device’s network, and screen properties, as well

as the user’s current location. The location is sensed

by a location aware middleware (Santner et al., 2006)

enabling the system to use different location sensing

technologies as GPS and Bluetooth transparently.

2.1 Transcoding Media Cache

The leisure park Minimundus is equipped with

WLAN allowing the tour guide to get all data over

the network. A major problem exists in parts of the

park, where the network connectivity is insufﬁcient

for video streaming. To solve this issue a transcoding

multimedia proxy (TMC) was introduced. The proxy

allows a device-speciﬁc transcoding and caching of

video and audio material on the servers and caching

only at the clients. This means that in normal oper-

ation the clients play all videos from the local mem-

ory. In case of a cache miss the required content is

transcoded if necessary - in real time - and transferred

to the client.

2.2 Ofﬂine Operation

In normal operation the M3-Guide system works in a

hybrid way: Textual data of the HTML-presentation

are transferred via WLAN; video and audio data are

transferred via the TMC. In case of enough storage,

they are actually completely loaded from the local

storage card of the mobile device. In addition to this

hybrid version, an ofﬂine version is available where

all textual data are also stored on the local memory

card. This conﬁguration is the most robust one and

needs no investment in network infrastructure.

3 SCALING ISSUES OF A

STATE-WIDE TOUR GUIDE

A state wide multimedia tour guide changes the

operating environment compared to the MultiMundus

scenario completely. While the tour guide is still

intended to be used to get detailed information about

exhibited objects at points of interest (POIs) the

technical basis changes totally. First of all there is

no single network technology available with which

all data can be transferred at any time. Rather, the

situation is shown in Figure 1 where many different

network types may or may not be available depending

on the user’s location. In contrast to the local tour

guide this is not only a technical problem with regard

to video content distribution over low-bandwidth

links, it is also a question of economics because in

networks which are not operated by the tour guide

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

Figure 1: Network situation of the state-wide tour guide.

provider high trafﬁc dependent usage fees (Mundt,

2004) must be paid. In addition, the usage patterns of

a multimedia tour guide will produce trafﬁc patterns

in cellular networks which they are not built for. In

alpine and rural environments usually very big cell

radiuses are used (Buddendick et al., 2003). This

does not cause a problem when only few people

transfer data in these areas. But what happens if a full

bus of tourists begins a hike and use their personal

tour guides to stream different video items more or

less at the same time? This scenario is a challenge

to today’s available public cellular networks in rural,

alpine environments.

A different approach to solve the scaling issues

is the preloading of all required data in advance.

This technique outperformed other solutions in the

case of the MultiMundus scenario, due to its high

reliability and moderate costs. Unfortunately, the

same technique cannot be applied to the state wide

tour guide either because there is no memory card or

widely available standard that allows memory sizes

bigger than 2-32GB. 2GB is the current limitation

of standard SD-cards while 32GB is the limit of

the emerging High Capacity SD-cards (SD-Group,

2006). Today’s practically available 2GB and the

future 32GB are deﬁnitely more than the card sizes

of the MultiMundus project of 512MB. But they are

still far too small to operate a single tour guide of all

Carinthian tourist attractions.

The limitations of network connections and local

memory can be solved if the tour guide system does

not deliver all content in advance (ofﬂine version) but

hoards the potentially required content in advance

at WLAN hot spots operated by the tour guide

provider. Which data should be hoarded depends

on the context, i.e., the user’s location, the user’s

preferences, and her/his history.

4 RELATED WORK

The basis of systems which support continued opera-

tion when leaving the network were already discussed

in the 1990th when practically no wireless networks

existed. The idea of hoarding (Tait et al., 1995) is that

the data a mobile computer needs from the network

should be stored on the mobile device in advance. The

data that should be present on the mobile device when

leaving the network is called Hoard-Set. The problem

of computing the Hoard-Set is called Hoarding Prob-

lem. The work in (Kuenning and Popek, 1997) shows

a scenario where the Hoarding Problem is solved by

the observation of ﬁle access patterns. The authors

introduce a semantic distance measure between ﬁles.

This means that ﬁles with a low distance to each other

should be in the Hoard-Set if one of the ﬁles was

accessed before leaving the network. In (Kuenning

et al., 2002) it was shown that in comparison to auto-

matic hoarding with semantic distance measure even

simple LRU hoarding can be nearly equally beneﬁ-

cial. In (Huizinga and Sherman, 1998) and (Kistler

and Satyanarayanan, 1992) extensions to current ﬁle

systems and distributed ﬁle systems with hoarding ca-

pabilities are presented that support hoarding based

on more or less predeﬁned Hoard-Sets.

The presented work focuses on discontinued opera-

tions for classical applications. These results are not

fully suitable for a context-aware tour guide because

the context reﬂected in these earlier applications is

limited to the observations of ﬁle access patterns. In a

mobile tour guide the usage context is much broader.

Therefore, it can be beneﬁcial to use a broad range

of context information to compute the hoarding deci-

sion. In (Kubach and Rothermel, 2001) a system is

presented which is designed to realize a tour guide in

a city. It is based on Wireless LAN hot spots. Each

hot spot is responsible for a speciﬁed area. Based

on the past visits, the system calculates which items

are needed when the user reaches a hot spot. This

means that dependent on the current location of the

user visit probabilities for each POI are calculated

based on past visits from other users. There are two

versions: one with basic visit probabilities and one

with more detailed context information provided by

external knowledge (e.g., street maps or predeﬁned

routes).

While this system is designed for a tour guide in

a single city the system in (Kirchner et al., 2004) is

designed to provide up to date information for boat

drivers on European water ways. Therefore the geo-

graphic distribution is much broader. The system uses

the usage context to determine the needed data. In the

case of boating this is mainly the type of boat, the

CONTEXT-AWARE HOARDING OF MULTIMEDIA CONTENT IN A LARGE-SCALE TOUR GUIDE SCENARIO -

A Case Study on Scaling Issues of a Multimedia Tour Guide

current position, the direction, and the speed. The fu-

ture location of a boat can easily be computed when

knowing the actual position, the target place, and type

of boat (e.g., max. speed). In addition to this mobile

hoarding on the boat, a Web platform is used to pre-

deﬁne the context in form of interests, boat type and

target. After deﬁning the context via the Web plat-

form the initial Hoard-Set containing the relevant data

is produced. On the boat therefore only changes and

new data must be transferred. The system is restricted

to boating and is therefore not suitable for our sce-

nario. In contrast to the previous system implement-

ing an actual application, (Feng et al., 2006) shows

a general context model which can be used to solve

the hoarding problem for a tour guide in a more gen-

eral way. It is based on an XML context description

combined with speciﬁed context weights to compute

which items are relevant in dependence of actual con-

text dimensions. The paper divides the usage context

into the groups: deﬁned, derived, and sensed. Sensed

context can be sensed via a sensor (e.g. location by a

GPS device), derived context is a context which is de-

rived by other context and deﬁned context is deﬁned

by the user. Deﬁned context must be queried from the

user or it can be deﬁned globally. When the context is

deﬁned by the user it must be considered that the user

is willing to deﬁne the context (e.g. in form of a ques-

tionnaire about interests). In (Kramer et al., 2005) a

user context driven tour guide was evaluated and it

was shown that users are often not ready to enter de-

tailed context information. Therefore we argue that in

our usage scenario detailed context information can

be queried from the user to take full advantage of the

system from (Feng et al., 2006).

5 CONTEXT CONSIDERATIONS

FOR A STATE-WIDE TOUR

GUIDE

The evaluation in (Kramer et al., 2005) shows that it

is not sure that a user wants to input detailed context

information. Therefore, a hoarding strategy based on

a very detailed user proﬁle does not necessarily result

in a better hit rate. In addition, in cases of hoarding it

is always possible to deﬁne the Hoard-Set manually.

Therefore, the time a user needs to deﬁne his/her con-

text (e.g., interests) must always be shorter than the

time required for a manual deﬁnition of the POIs a

user wants to visit. If a hoarding strategy would for

example rely on a predeﬁned route the hoarding deci-

sion should be made manually by the user. Therefore

we try to use context information which can be sensed

or queried with limited user input. Only if this con-

text information can produce good hoarding results it

can be beneﬁcially used for automatic hoarding in the

tour guide.

5.1 Static Context

We deﬁne static context as context which does not

usually change during the visit of the user. There-

fore it can be predeﬁned when the user gets her/his

tour guide for the ﬁrst time. This context data should

be used to determine what content should be stored

on the device before the ﬁrst usage. We refer to this

scenario as initial hoarding.

• Holiday Location. Tourists usually stay in a spe-

ciﬁc accommodation (e.g., hotel) in one of 16 re-

gions

. We suppose that the region has a strong

inﬂuence on the visit probabilities of the POIs the

user will visit.

• Season. We believe that the visit probability of a

user changes according to the season.

• User Group. If there are stereotypes of users

like family visitors, senior visitors, singles etc., we

suppose that this can be used to deﬁne the Hoard-

Set as well.

• Number of Travelers. The number of travelers in

a group and the number of children is also static

context information.

5.2 Dynamic Context

In contrast to the static context the dynamic context

of a user changes during the visit. Therefore, such

context items should be used to ensure that the right

items are on the device when the user moves around.

This scenario is referred to as mobile hoarding. A

hoarding decision must always be computed when a

user enters a POI which provides a network connec-

tion. Therefore data of relevant other POIs without a

network connection must be transferred to the mobile

device.

• Current Location. We suppose that the current

location has a strong inﬂuence on the user’s be-

havior. Therefore, we suppose that people will

visit geographically near objects or objects which

are often visited in a sequence.

• User History. We think that the history of a user

has a strong inﬂuence on the usage behavior. We

assume that visit patterns can be used to determine

knowledge about the user’s preferences.

Source: http://www.kaernten.at

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

• Weather. We suppose that the weather conditions

have a strong inﬂuence on the user’s behavior.

People usually do not go to the beach in case of

cold weather.

6 HOARDING MODEL

The total amount of data transferred over wireless

links should be reduced to a minimum. Therefore we

suppose a two-modes hoarding model. First of all,

one hoarding decision should be made to choose what

data (e.g., multimedia ﬁles) should be present on the

mobile device when it is used by a speciﬁc user for

the ﬁrst time (initial hoarding).

Second, a hoarding decision should be made when-

ever a user enters a POI with good wireless network

access (mobile hoarding). In this case the system

must ensure that all relevant data of other POIs -

where no network connections can be expected - are

located on the device. The relevance is inﬂuenced

by the context. Our proposed strategies are based on

the two major concepts visit probabilities and air dis-

tance.

6.1 Hoarding by Visit Probabilities

Comparable to the system of (Kubach and Rothermel,

2001) the Hoard-Set is computed by calculating the

visit probability of each POI. Therefore the hoarding

decision is based on the calculation of the probabil-

ity that a user visits a POI with regard to a speciﬁc

context item and value. The calculation of the visit

probability can be done straight forward using a sim-

ple formula for each POI.

t(c) = Total number of visits at context item c

v(c, POI) = Number o f visits of a POI at context item c

p(POI, c) =

v(c, POI)

t(c)

After the visit probability is computed for each

POI the items can be loaded onto the device in de-

scending order of their visit probability. The same

formula can be used for initial hoarding as well as for

mobile hoarding. The mobile hoarding of new items

results in the need for replacement strategies. The re-

placement strategy of items can be realized by using

the reverse order of the calculated visit probabilities.

6.2 Hoarding by Air Distance

The hoarding by visit probabilities requires the track-

ing of all user sessions. In case of mobile hoarding we

suppose that a simple hoarding by air distance might

result in high hit rates as well. Therefore, the n near-

est neighbors (derived from GPS coordinates) of an

object should be loaded onto the device. This can ad-

ditionally be combined with the visit probabilities of

the POI.

7 EVALUATION

We believe that the movement patterns of the tourists

strongly depend on the actual usage context. There-

fore, randomized movement patterns do not provide

a basis for an adequate evaluation test bed. We de-

cided to use real data of a tourist-card operator in

Carinthia called ”K¨arnten Card”. Tourists who pay

a comparably low price for the card can visit all afﬁl-

iated POIs for free or at a reduced price. The system

is technically realized as a chip card and terminals at

the POIs. For our evaluation we used a log ﬁle con-

taining all card usages of one season with about one

million card usages. We extracted all sessions from

the log ﬁle and stored them in a database. We divided

the sessions into 20% test sessions and 80% training-

sessions. The training sessions were used to compute

the visit probabilities with regard to different context

items like weather, current location, holiday location.

All hoarding algorithms are realized as database

queries. As the log ﬁle did not contain any further in-

formation about the card holders we could not evalu-

ate any predeﬁned interests of the users. We therefore

created stereotypes of users via a cluster analysis of

the sessions using SPSS (B¨uhl and Z¨afel, 2005). To

compute visit probabilities with regard to the actual

weather we used the weather service ”wunderground”

) to get the mean temperature of every day of the sea-

son.

The evaluation was ﬁnally realized by replaying the

test sessions from the log ﬁle using different hoarding

strategies and different cache sizes.

7.1 Mobile Hoarding

To observe the behavior of mobile hoarding we ﬁrst

tested some simple mobile hoarding algorithms with

randomized initialization. Therefore the cache is

ﬁlled randomly when the tour guide is ﬁrst used. The

result are shown in Figure 2. The reference line

(marked by a rhombus) shows the hit rate and cache

size when no hoarding is performed. This means that

the only way new items get into the cache is that

The Weather Underground, Inc - http://

www.wunderground.com

CONTEXT-AWARE HOARDING OF MULTIMEDIA CONTENT IN A LARGE-SCALE TOUR GUIDE SCENARIO -

A Case Study on Scaling Issues of a Multimedia Tour Guide

Figure 2: Simple hoarding with randomized initialization.

a cache miss occurs at one of the 20 POIs with a

hot spot out of the total 107 POIs. Replacement is

done according to the air distance (GPS). Items with

the maximum distance to the current location are re-

moved. The line marked by a square shows the behav-

ior of mobile hoarding using simple air distance. This

means that when entering a hot spot POI the hoard-

ing algorithm checks that at least the 10 nearest other

POIs are also stored on the mobile device. This strat-

egy already results in a major improvement of the hit

rate of up to 10%.

The air distance is only a heuristics and does not re-

ﬂect the known visit probabilities in any way. There-

fore, we also tested the behavior of a combined strat-

egy using the global visit probability of an object

combined with the air distance. The results of this

strategy are shown in the line marked by a triangle.

This strategy is up to 15% more beneﬁcial than no

hoarding. It is very interesting that it comes very close

to the much more expensive strategy shown in the line

marked by a cross. This strategy uses the transition

probability. This means that all visits must be tracked

and the probability of a visit of each POI after the

visit of each other POI needs to be calculated based

on historical sessions.

In Figure 3 more advanced strategies are shown.

In addition to the simple strategies they use more

knowledge about the user context. In real world sce-

narios this context should be built from user proﬁles.

As no user proﬁle were available, artiﬁcial proﬁles are

created by building clusters based on the visit patterns

of each session. Thus visit predictions are made on a

per group basis. Instead of calculating the probabil-

ity of a transition from one POI to another globally,

only sessions of the same group are used for the com-

putation. The best strategy from Figure 2 is used as

a reference. It is marked by a rhombus. Against this

strategy a transition probability strategy with 16 and

64 clusters is evaluated. The cluster-based strategies

have the main drawback that the clusters cannot be

computed a priori in real world scenarios. Therefore,

we also tested a method which computes comparable

sessions dynamically using a bit vector containing all

visits of all sessions. The columns contain bits that

indicate the visit of a speciﬁc POI. The rows repre-

sent a speciﬁc session. When a prediction is made

only sessions which have visited at least 70% of the

POIs visited by the session in question are used for the

prediction. We also believe that the visit probability

strongly depends on the weather (e.g. indoor muse-

ums have many visitors at bad weather). Therefore,

we evaluated a strategy which uses the visit probabil-

ity of different weather conditions. The actual hoard-

ing was therefore realized by the air distance and the

visit probability at the current temperature.

Figure 3 shows clearly the surprising result that none

Figure 3: Extended mobile hoarding strategies.

of the advancedstrategies achieved any signiﬁcant ad-

vantage over simple hoarding by global visit probabil-

ity and air distance.

Figure 4: Initial hoarding only.

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

7.2 Initial Hoarding

In Figure 4 the results of different initialization strate-

gies without any mobile hoarding at WLAN POIs are

shown. The simplest approach is to hoard all data of

the POIs according to their global visit probability:

p(POI) =

v(POI)

t(all)

The results are shown in the line marked by a rhom-

bus. A more ﬁne grained strategy (marked by a rect-

angle) preloads all data of POIs which are mostly vis-

ited by guests staying in their region:

p(POI, region) =

v(POI, region)

t(region)

It is interesting to see that this approach does not pro-

duce signiﬁcantly higher hit rates. Therefore, we must

conclude that people drive all around the state and

that they are not bound to the region where they stay.

In the next step we computed the visit probability of

each POI for each group of users. The groups where

simulated by a cluster analysis. We build 2, 16, and 64

clusters. It is interesting to see that even two clusters

perform better than the (16) regions.

p(POI, group) =

v(POI, group)

t(group)

The best hit rate can be achieved when using 64 clus-

ters. This pretty high number of clusters might be a

problem in a real world scenario. Therefore, it must

be noticed that even 16 clusters (which should be

practically usable) result in good hit rates. A Hoard-

Set size of 40 items results in a hit rate of around 95%.

This is about 30% better than the results of the best

mobile hoarding strategy with randomized preload.

In contrast to the advanced mobile hoarding strate-

gies which did not achieve any signiﬁcant advantage,

more complex initial hoarding strategies can be used

beneﬁcially.

7.3 Initial Hoarding and Mobile

Hoarding

For an implementation of the tour guide we assumed

that best hit rates should be reachable by combining

intelligent initialization and additional mobile hoard-

ing. Figure 5 shows the results of different combined

approaches. The reference line (marked by a rhom-

bus) shows the results of a pure initialization by global

visit probability without mobile hoarding. The line

marked by a rectangle shows the results of a com-

bined approach of hoarding by global visit probability

and additional mobile hoardingby the air distance and

Figure 5: Initial hoarding and mobile hoarding.

global visit probability. The - expensive - transition-

probability approach performs slightly better than the

air distance and global probability approach. The hit

rates get a bit better when using clusters. It is inter-

esting that the hit rates are only slightly better than

without mobile hoarding. The second reference line

(marked by *) shows pure initial hoarding with 64

clusters. This performs much better. A combination

of initial hoarding and mobile hoarding does not re-

sult in a signiﬁcantly better hit rate than initial hoard-

ing only.

8 CONCLUSIONS AND FUTURE

WORK

This work presented the scaling issues of a multime-

dia tour guide, expanding from a single location to

the whole federal state of Carinthia. Current network

and storage technologies do not allow a straight for-

ward scaling for such a tour guide system. Hoard-

ing is a technology which can help to overcome these

limitations. In a tour guide scenario automatic hoard-

ing can be based on the user’s context. The time

a user needs to enter his/her context must not take

longer than the time a user would spend to enter the

Hoard-Set (POIs she/he wants to visit) manually. We

therefore evaluated if it is possible to make a hoard-

ing decision based on context information which can

be sensed automatically or requires very limited user

interaction. The evaluation was based on real world

data provided by the Carinthian tourist card opera-

tor. The evaluation has shown that with the evalu-

ated data-set, context data and the described hoard-

ing strategies no signiﬁcant gain for mobile hoarding

could be achieved. Suggestions about visitor behavior

and corresponding hoarding strategies are of limited

value. No strong geographical dependencies could be

recognized in the visit patterns. People tend to drive

CONTEXT-AWARE HOARDING OF MULTIMEDIA CONTENT IN A LARGE-SCALE TOUR GUIDE SCENARIO -

A Case Study on Scaling Issues of a Multimedia Tour Guide

Figure 6: Distribution of visit probabilities.

fairly freely throughout the whole state. Even weather

conditionsdo not signiﬁcantly change the visitors’ be-

havior. It is possible that this surprising behavior is

caused by the fact that the ”K¨arnten Card” is a bonus

card with free entry. People might tend to visit all ma-

jor and most expensive POIs they can. This behavior

can also be seen in the total distribution of visits as

shown in Figure 6. This ﬁgure does also explain why

a simple hoarding by global visit probability already

produced good results. People tend to visit the top

20 POIs. What else of the other 90 POIs is visited

cannot be predicted with the observed context. The

good news is that initial hoarding for the tour guide

can achieve good hit rates. In contrast to our initial as-

sumptions the region where the user stays for vacation

has a comparably low inﬂuence on the visit probabil-

ities and thus cannot be used for automatic hoarding

in our scenario. We therefore suggest the implemen-

tation of a Web platform on which the user can take

the hoarding decision manually. This has the addi-

tional advantage that no mobile hoarding infrastruc-

ture (WLAN hot spots) needs to be installed. Such

a Web platform could be integrated into the ofﬁcial

tourist web-site of a large area (such as Carinthia in

our case). It should be used to identify the relevant

POIs for the user and to acquire the user and device

context to build a customized tour guide.

REFERENCES

B¨uhl, A. and Z¨afel, P. (2005). SPSS 12 Einf¨uhrung in

die moderne Datenanalyse unter Windows. Pearson

Studium, M¨unchen/Germany.

Buddendick, H., Weber, A., and Tangemann, M. (2003).

Comparison of Data Throughput Performance in

GPRS, EGPRS, and UMTS. In Proc. of the World

Wireless Congress 2003, San Francisco (CA).

Feng, W., Zhang, L., Jin, B., and Fan, Z. (2006). Context-

aware Caching for Wireless Internet Applications. In

ICEBE ’06: Proc. of the IEEE International Con-

ference on e-Business Engineering, pages 450–455,

Washington, DC, USA. IEEE Computer Society.

Huizinga, D. M. and Sherman, H. (1998). File Hoarding

under NFS and Linux. In SAC ’98: Proc. of the 1998

ACM Symposium on Applied Computing, pages 409–

415, New York, NY, USA. ACM.

Kirchner, H., Krummenacher, R., Risse, T., and Edwards-

May, D. (2004). A Location-aware Prefetching Mech-

anism. In Proc. of 4th International Network Confer-

ence (INC 2004), pages 453–460, Plymouth, UK.

Kistler, J. J. and Satyanarayanan, M. (1992). Disconnected

Operation in the Coda File System. ACM Trans. Com-

put. Syst., 10(1):3–25.

Kramer, R., Modsching, M., and ten Hagen, K. (MARCH

2005). Development and Evaluation of a Context-

Driven, Mobile Tourist Guide. j. pervasive comput.

& comm., 1(1).

Kray, C. and Baus, J. (2003). A survey of mobile guides.

In Workshop HCI in mobile guides at Mobile HCI,

Udine, Italy.

Kropfberger, M., Tusch, R., Jakab, M., K¨opke, J., Ofner,

M., Hellwagner, H., and B¨osz¨ormenyi, L. (2007). A

Multimedia-Based Guidance System for Various Con-

sumer Devices. In Proc. of the 3rd International Con-

ference on Web Information Systems and Technologies

(WEBIST ’07), pages 83–90. INSTICC Press.

Kubach, U. and Rothermel, K. (2001). Exploiting Location

Information for Infostation-based Hoarding. In Mobi-

Com ’01: Proc. of the 7th Annual International Con-

ference on Mobile Computing and Networking, pages

15–27, New York, NY, USA. ACM.

Kuenning, G. H., Ma, W., Reiher, P., and Popek, G. J.

(2002). Simplifying Automated Hoarding Methods.

In MSWiM ’02: Proc. of the 5th ACM International

Workshop on Modeling Analysis and Simulation of

Wireless and Mobile Systems, pages 15–21, New

York, NY, USA. ACM.

Kuenning, G. H. and Popek, G. J. (1997). Automated

Hoarding for Mobile Computers. In SOSP ’97: Proc.

of the Sixteenth ACM Symposium on Operating Sys-

tems Principles, pages 264–275, New York, NY, USA.

ACM.

Mundt, T. (2004). How much is a Byte?: A Survey of Costs

for Mobile Data Transmission. In WISICT ’04: Proc.

of the Winter International Symposium on Information

and Communication Technologies, pages 1–6. Trinity

College Dublin.

Santner, M., Tusch, R., Kropfberger, M., Hellwagner, H.,

and B¨osz¨ormenyi, L. (2006). Ein Ortserkennungssys-

tem f¨ur mobile Touristenf¨uhrer. In DACH Mobility

2006, pages 84–98, Ottobrunn, M¨unchen, Deutsch-

land. IT Security & IT Management.

SD-Group (2006). SD Speciﬁcations Part 1 Physical Layer

Simpliﬁed Speciﬁcation. Standard, SD Card Associa-

tion.

Tait, C., Lei, H., Acharya, S., and Chang, H. (1995). Intel-

ligent File Hoarding for Mobile Computers. In Mobi-

Com ’95: Proc. of the 1st Annual International Con-

ference on Mobile Computing and Networking, pages

119–125, New York, NY, USA. ACM.

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications