CONTEXT DIMENSIONALITY REDUCTION FOR MOBILE

PERSONAL INFORMATION ACCESS

Andreas Komninos

, Athanasios Plessas

2,3

, Vassilios Stefanis

2,3

and John Garofalakis

2,3

Mobile and Ubiquitous Computing Group, Glasgow Caledonian University, Cowcaddens Road, Glasgow, U.K.

Department of Computer Engineering and Informatics, University of Patras, Patras, Greece

Computer Technology Institute “Diophantus”, Rion, Patras, Greece

Keywords: Context awareness, Dimensionality Reduction, Mobile Personal Information Access.

Abstract: We propose an application of the Fastmap algorithm that could provide a breakthrough in the efforts to

present mobile personal information to the user in context, and describe our vision for context-driven

interfaces generated by this method that will support the richness of data stored in personal devices.

1 INTRODUCTION

Advances in mobile hardware technology in terms of

storage space and data generation modalities

(efficiency-enhancing input UIs, sensor-generated

data and meta-data) allow today’s mobile device

user to quickly generate and store large volumes of

personal data. Much of this data is organized in

structured repositories (e.g. contact list, photo

gallery, message inbox), affording the user a means

of learnable, procedurised retrieval. Even though in

daily situations the user may require access to

multiple types of personal information, in most

devices the structured repositories remain mostly

“walled gardens”, requiring the user to sequentially

visit each of them in order to assemble the

information pieces she needs in one coherent

mentally held collection, relevant to their current

activity. Naturally, the cognitive load on the user

increases with the number of information items that

have to be retrieved from the repositories. But still,

even considering each “walled garden” individually,

it quickly becomes apparent that current retrieval

methods are largely inefficient, in view of the ever-

expanding storage space and richness of personal

information stored in mobile devices. For example

consider work on visualizing and retrieval from

large mobile photo galleries (Hsu et al., 2009) or

ever expanding music collections (Tolos, Tato and

Kemp, 2005) or contact lists. What can a user do

with all this data? Surely users want the data, but is

it really useful in its raw form as a singular item and

can it be used to help them achieve their goals?

Mobile devices such as smartphones, are still

primarily information access devices and

communication devices. Much of the activity on

mobile devices, especially mobile phones, is in

support of Human-to-Human interaction.

Consequently, many of the actions that someone

might perform on their mobile device involve the act

of looking up a contact, that perhaps carries at that

time some importance to the user, in order to initiate

some form of communication (call, sms, email,

comment on facebook etc.). Not forgetting that

communication is by definition the exchange of

information, information items are an important part

of it. Thus, mobile devices should support not just

the retrieval of contacts, but also of information that

is somehow relevant to these contacts. It is critical

here to underline that our work extends beyond the

notion of contact importance – in fact, we believe

that importance as a multidimensional vector can

provide a complex query “key” on the large, rich

dataset in a user’s mobile device for any type of

personal information item (e.g. a photograph, or a

set of SMS messages or stored Word documents) so

as to answer the question of who to communicate

with and what to communicate. As such our aim is

to construct an extensible, flexible approach to the

definition of importance in a k-dimensional context

space, which can be applied to any mobile

information retrieval problem (and consequently UI

design for mobile information access). We choose to

start with the concept of contacts as a problem

493

Komninos A., Plessas A., Stefanis V. and Garofalakis J..

CONTEXT DIMENSIONALITY REDUCTION FOR MOBILE PERSONAL INFORMATION ACCESS.

DOI: 10.5220/0003688504850490

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2011), pages 485-490

ISBN: 978-989-8425-79-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

domain, largely due to its importance in everyday

interaction with mobile devices.

2 BACKGROUND

Our motivation stems from the observation of user

behaviour in accessing non-contextualised

information from single repositories, namely the

mobile contact list. In recent work (Bergman et. al.,

2011), we demonstrated some interesting user

behaviours, which we believe, highlight the issue of

non-contextualisation. Firstly, we observed that our

18 young users (<25 yrs. old) from varied

backgrounds tend to keep fairly large repositories of

contacts (m=92.47 sd=56.93), which, as literature

suggests (Boardman and Sasse, 2004), are bound to

get bigger as they grow older (just one of our users

reported actively deleting unused contacts). In these

collections, a large percentage of contacts (av=39%,

sd=17%) has either not been used in over 6 months

or not at all. Viewing the collections as a whole

(n=754 contacts), over 47% of all contacts had not

been used in over 6 months or at any other point,

while just 15% were identified as frequently used

contacts, the remaining 38% having been used

between 1 and 6 months previous to our study

period. We also found a highly significant

correlation between the size of the contact list and

the number of unused contacts (r=0.81). In a

carefully designed experiment, we presented our

users with a contextualised UI which placed

“important” contacts in an alphabetic list, followed

by an alphabetic list of all other contacts and asked

them to perform several retrieval tasks on this

augmented UI and on a classic (Nokia) UI.

Importance in this experiment was set as a single-

value vector of frequency of use, with 6 months or

more being a threshold for considering a contact as

not important. We found that users provided very

positive subjective feedback to their experience

using this approach (86% found it easier to use than

the traditional UI and 64% would like to have this

feature on their next device). This was backed up by

significant performance metrics. On average users

required less button presses to successfully “find” a

contact (mean reduction=1.96, sd=3.56) and

significantly less time with the augmented UI

(m=4,424ms, sd=1,872ms) than the traditional UI

(m=5,204ms, sd=2,829ms).

While these findings assumed a naïve

classification of importance (just frequency of use),

they demonstrate quite effectively how

contextualisation on just a single dimension can

have a very positive effect on the retrievability of

personal mobile information. In the past, other

researchers have demonstrated either issues with the

usability of contact and call lists (Böcker and

Suwita, 1999) (Klockar et al., 2003) or improvement

in usability through the introduction of user activity,

proximity & state context (Oulasvirta et al., 2005),

social context (Gaur, 2008) (Rhee et al., 2006) and

temporal context (Jung et al., 2008). Further work by

Ankolekar et al. (Ankolekar et al., 2009) discusses

how a combination of contextual cues might offer

usability advantages but leaves the categorisation of

contacts to users and does not present any tangible

research into one of the most oft-used applications

of mobile devices. The volume of research in contact

list use remains low and the technology behind

contact list UIs in today’s devices is mainly based on

alphabetic lists. Currently, only the Android OS

provides a feature of presenting a contact list by use

frequency, though again this is a simplistic view of

importance, considering just one type of context (fig.

1).

Figure 1: The Android contact list UI, augmented with a

single-dimension context cue (frequency of use) [left] and

a standard Symbian contact list UI, restricted to

alphabetically ordered lists [right].

3 THE NEED FOR A

K-DIMENSIONAL APPROACH

TO CONTEXT MAPPING

Mobile devices collect a significant amount of data

and information about the user's context. Such

information includes location (absolute or relative),

the current time, whether the user of the device is on

move and their speed, the orientation of the device,

the user’s current task (e.g. on the phone,

messaging), whether the vibration or the silent mode

are enabled (Beach et al., 2010) etc. The user

considers her mobile device a "trusted device". She

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

494

usually has the device close to her, sometimes

operating 24 hours per day. Devices contain a lot of

personal information related to the user’s social

environment (Toninelli et al., 2008). These are often

generated automatically by the device (e.g. a

smartphone's phone list saves the calls that have

been made, the time of the day for each call and the

duration of each call for the past few days or even

weeks). Moreover, mobile devices store user-

generated content (e.g. SMS/MMS and audio files,

browser's history, calendar with user's events etc).

Therefore, a mobile device could also be aware of

the social environment of the user (social context).

In addition, mobile applications can take advantage

of social data from online social networks in order to

enrich a contact with more information (Bentley et

al., 2010). The combination of social and mobile

context results in a dynamically defined social

context, termed the mobile social context (Gilbert et

al., 2009). Therefore, a truly context aware mobile

information access application has to consider social

context as part of its context representation.

It is therefore quite apparent that true

contextualisation is much more complex than in our

previous experiment’s assumption, and that it

requires k-dimensional space in order to be defined.

As shown in equations (1), (2) and (3) we use a

vector model to represent the context of an

information item i in a k-dimensional space as a k-

tuple, where d

is the context atom value for

dimension n. Its importance can be considered as the

sum of the weighted distances between the vector

atoms describing current context properties and an

item’s context atoms (

(

)

),...,,()(

21 n

dddiC =

(1)

(

)()

∑

Δ−×=

iCwI

)(1

(2)

)()()( iCnCiC

GGG

−=Δ

(3)

Researchers in the past have often attempted to

combine the user's input with the user's context in

order to provide a richer user experience (context-

aware applications), e.g. (Yoon et al., 2008). Most

approaches tend to focus on narrow objectives, as

the capture of context for general use is still

considered very difficult. In literature, context has

been represented in the form of vectors, e.g. (Du and

Wang, 2008) while other researchers have adopted

an ontological approach, e.g. (Korpipaa et al., 2004),

often using naïve Bayes classifiers to solving the

issue of defining context space. Vector based

approaches carry the disadvantage of computational

complexity in similarity searches (each dimension

needs to be compared separately) and that weighed

vectors require an empirical (hence error-prone or

narrowly applicable) estimation of the weights. On

the other hand, ontologies tend to be inflexible and

too strict to be applicable to wider ranges of

problems, requiring careful, manual approaches to

their construction.

4 MAPPING CONTEXT IN

K-DIMENSIONAL SPACE

In the previous section we outlined the

disadvantages of the vector and ontological

approaches for determining context. In our view, the

vector approach can offer a more flexible solution to

context acquisition and representation, hence our

work relates to overcoming its computational

complexity and vector weighting issues. In

(Komninos and Liarokapis, 2009) we proposed four

criteria for the determination of a m-PIM (mobile

Personal Information Management) item importance

(namely contacts). From these criteria, we can derive

the following context dimensions for the contact list

problem domain, on which a contact can be mapped:

• Frequency (e.g. Frequency of use in last n

months)

• Recency (e.g. days since last use)

• Location (e.g. Geographic areas from which at

least n% of uses are made)

• Time of Day (e.g. Time segment of day in

which at least n% of uses are made)

• Task (e.g. Boolean measure of existence of a

scheduled task involving a contact within a

certain window of time [for example today] or

temporal distance between now and such

scheduled task?)

• Personal preference (e.g. scale of 1-5 of explicit

user rating of importance for a given contact)

In order to estimate a “match” signifying

importance between these types of contextual

information and the user’s current context, a typical

approach would be to measure the distance between

current context and contextual data (e.g. time of day

now vs. usual time of day of contact use) and

combine this with static context (e.g. explicit

importance rating). The derived metrics would need

to be weighted and the sum of these weighted

metrics could then be used to infer “importance” for

a single contact under any context (equations (1), (2)

and (3)). This approach though is not without

challenges: Firstly, one must determine appropriate

CONTEXT DIMENSIONALITY REDUCTION FOR MOBILE PERSONAL INFORMATION ACCESS

495

weights for each context type. Subsequently, it is

easy to realize that this would be a futile attempt, as

the weights of each context type are naturally

dynamic and can vary under different use contexts

(see scenario in figure 2). The example scenario

introduces the problem of context-derived specific

importance (vs. general importance) and shows that

a decent approximation to the calculation of this

term is very difficult, due to the infinite variability

of context itself and the complexity of its

dimensions. A possible solution however could

come from the field of Databases and IR, where

dimensionality reduction is a technique often used to

automatically extract important features from

complex data and reduce the complexity of searches

in multidimensional spaces.

Take the example of “John”, a contact that the user

calls every day and “Jane” another contact that is only

called once a year. “John” can be considered generally

important. However if the user is running late for a

meeting with Jane Doe that will take place in 10

minutes, then Jane becomes undeniably more

important than anyone else, and her importance should

rise and decay naturally as time flows around the

scheduled event.

Figure 2: The Importance Scenario.

5 DIMENSIONALITY

REDUCTION (DR) IN M-PIM

ACCESS

DR, as the name suggests, is an algorithmic

technique for reducing the dimensionality of data,

applied in several computer science fields such as

databases, information retrieval, data mining,

recommendation systems, signal processing etc.

Real-world data usually has a high dimensionality, a

fact that affects data processing performance (the so

called “curse of dimensionality” originally appearing

in (Bellman, 1957) that suggests exponential

dependence of an algorithm on the dimension of the

input). The idea is to transform data from a high-

dimensional space to a low-dimensional space,

preserving some critical relationships among

elements of the data set. In mathematical terms,

given a p-dimensional object x=(x

,…,x

)

, find a

lower dimensional representation of it, s=(s

,…,s

)

with k≤p, that captures the content in the original

data, according to some criterion (Fodor, 1992).

There are two categories of methods in order to

solve the problem: a) feature selection, where an

optimal subset of features (dimensions) is chosen

and b) feature extraction, where existing features are

combined and transformed to new ones.

Our research idea is to perform DR to context

augmented personal information items, such as

entries in a contact list, an idea that has not yet been

proposed and applied in scientific literature as far as

we know. Feature selection for context DR is not

practical, as it is neither possible to know the best-

describing features of the context vector nor their

weights in advance. Since, as already presented,

context augmented personal information items can

be represented as multidimensional vectors, we find

it highly appealing to try to extract a small number

of features that could accurately represent the

original items and their relationships, so as to enable

quick and accurate similarity searches for related

personal information items. Furthermore, after

reducing the dimensions of the items, it might be

desirable to map them to a 1-d, 2-d or 3-d space, as

often done in high-dimensional data projections,

since visualization tends to reveal existing groups of

objects.

FastMap algorithm:

1. Find two objects that are far away.

2. Project all points on the line the two objects

define, to get the first coordinate.

3. Project all objects on a hyperplane

perpendicular to the line the two objects

define.

4. Repeat k-1 times

Figure 3: The FastMap Algorithm.

There is a wide range of algorithms with diverse

characteristics that achieve dimensionality reduction,

following different approaches. An interesting

method that could be appropriate for the case of

mobile phones due to its simplicity and

computational efficiency is the FastMap algorithm

(Faloutsos and Lin, 1995). FastMap is a fast

algorithm that maps high-dimensional objects into

lower-dimensional spaces, while preserving well

distances between objects and the structure of the

data set, as a result preserving also dis-similarities

between objects. Experiments presented in

(Faloutsos and Lin, 1995) show that the algorithm

performs well for visualization and clustering. The

algorithm functions as shown in figure 3 and its

complexity is O(Nk

), where k are the dimensions of

the target space. A further advantage of this

approach is that context vector “atoms” need to be

defined by application developers just once – the

recursive and atom-agnostic nature of the algorithm

allow it to work for any context description vector.

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

496

6 CONCLUSIONS

In the previous sections, we underlined the

significant role that mobile social context can play in

retrieving and presenting information to the user

from repositories within the mobile device that may

contain large volumes of data. We have proposed the

use of a vector model for the representation of

context and since finding weights (that may change

dynamically) is a very complex task, we introduced

the idea to apply the technique of dimensionality

reduction. This technique is characterised by its

ability to automatically produce meaningful clusters

of related information and thus can make the

contextualised visualization of personal information

items in 2 or 3 dimensions feasible.

We believe that our technique as described

above can be very useful in order to build rich

singular, two or three-dimensional information

retrieval interfaces that will support data from

multiple information repositories and present them

in context, as demonstrated in the mock-ups

presented in this paper (figures 4 & 5).

Figure 4: Hybrid one-dimensional mapping of importance

(important items are highlighted in bold but ordered

alphabetically so as to maintain user familiarity with

existing UIs).

In figure 4 a one-dimensional hybrid interface is

presented, where contacts are sorted alphabetically,

but for each letter the important contacts are on top

of the list and the remaining follow in alphabetical

order. In figure 5 we show some concept renderings

of 2-d and 3-d retrieval interfaces. In the case of 2

dimensions, the dimension of time can be preserved

and as a result the user is able to retrieve the most

important contacts during each time period. Finally,

in the case of 3 dimensions an example of how this

technique extends beyond the domain of contact lists

is presented. The respective figure (5b) shows how

several personal information items (contacts, e-

mails, SMSs etc.) could be projected in a 3-

dimensional space (with possible dimensions

presented on the axes of time, importance and

distance from current location).

Figure 5a (left) and 5b (right): Retrieval UIs using 2D

(left) and 3D (right) projections of item (contacts, e-mails,

photos etc.) importance combined with retained,

unprojected dimensions (time, distance from current

location).

At this point in time our work focuses on a

context-enabled contact list and following trials will

extend to support a richer information space that will

include all types of media and information pertinent

to those contacts, enabling a new mode of context-

based search and retrieval for mobile devices.

ACKNOWLEDGEMENTS

This work has in part been supported by the National

Strategic Reference Framework (NSRF) (Regional

Operational Programme – Western Greece) under

the title “Advanced Systems and Services over

Wireless and Mobile Networks” (number 312179).

REFERENCES

Ankolekar, A., Szabo, G., Luon, Y., Huberman, B. A.,

Wilkinson, D., Wu, F., 2009. Friendlee: a mobile

application for your social life. In Proceedings of

MobileHCI '09, Bonn, Germany.

Beach, A., Gartrell, M., Xing, X., Han, R., Lv, Q., Mishra,

S., Seada, K., 2010. Fusing Mobile, Sensor, and Social

Data To Fully Enable Context-Aware Computing. In

Proceedings of HotMobile 2010, Annapolis,

Maryland, USA.

Bellman, R., 1957. Dynamic Programming, Princeton

University Press. Princeton, NJ, USA.

Bentley, F., Kames, J., Ahmed, R., Zivin, R.,

Schwendimann, L., 2010. Contacts 3.0: bringing

CONTEXT DIMENSIONALITY REDUCTION FOR MOBILE PERSONAL INFORMATION ACCESS

497

together research and design teams to reinvent the

phonebook. In Proceedings of the 28th of the

international conference extended abstracts on Human

factors in computing systems, Atlanta, USA

Bergman O., Komninos A., Liarokapis D., Clarke J., 2011.

"You never call: Demoting unused contacts on mobile

phones using DMTR", Personal & Ubiquitous

Computing, Online First, DOI 10.1007/s00779-011-

0411-3.

Boardman, R., Sasse, M. A., 2004. "Stuff goes into the

computer and doesn't come out": a cross-tool study of

personal information management. In Proceedings of

CHI'04, Vienna, Austria.

Böcker, M., Suwita, A., 1999. Evaluating the Siemens

C10 mobile phone – Beyond “Quick and Dirty”

Usability Testing. In Proceedings of HFT’99,

Copenhagen, Denmark.

Du, W., Wang, L., 2008. Context-aware application

programming for mobile devices. In Proceedings of

C3S2E Conference, Montreal, Canada.

Faloutsos, C., Lin, K. I., 1995. Fastmap: A Fast Algorithm

for Indexing, Data-Mining and Visualization of

Traditional and Multimedia Datasets. In Proceedings

of ACM SIGMOD '95, ACM, New York, USA.

Fodor, I. K., 1992. A survey of dimension reduction

techniques. LLNL Technical Report, UCRL-ID-

148494.

Hsu, S.-H., Cubauld, P., Jumpertz, S. 2009. Phorigami: A

Photo Browser Based on Meta-categorization and

Origami Visualization. Human-Computer Interaction:

Novel Interaction Methods and Techniques Lecture

Notes in Computer Science, Volume 5611/2009, 801-

810, Springer.

Gaur, S., 2008. Mobile Phone Contact. In: Ylä-Jääski A,

Takkinen L (eds) Technical Reports in Computer

Science and Engineering.

Gilbert, P., Cuervo, E., Cox, L., 2009. Experimenting in

Mobile Social Contexts Using JellyNets. In

Proceedings of HotMobile 2009, Santa Cruz, USA.

Jung, Y., Anttila, A., Blom, J., 2008. Designing for the

evolution of mobile contacts application. In

Proceedings of MobileHCI'08, Amsterdam,

Netherlands.

Klockar, T., Carr, D., Hedman, A., Johansson, T.,

Bengtsson, F., 2003 Usability of Mobile Phones. In

the 19th International Symposium on Human Factors

in Telecommunications, Berlin, Germany.

Komninos, A., Liarokapis, D., 2009 The Use of Mobile

Contact List applications and a Context-Oriented

Framework to Support their Design. In Proceedings of

MobileHCI'09, Bonn, Germany.

Korpipaa, P., Hakkila, J., Kela, J., Ronkainen, S., Kansala,

I., 2004 Utilising context ontology in mobile device

application personalisation. In Proceedings of

MUM'04, ACM, New York, USA.

Oulasvirta, A., Raento, M., Tiitta, S., 2005.

ContextContacts: re-designing SmartPhone's contact

book to support mobile awareness and collaboration.

In Proceedings of MobileHCI'05, Salzburg, Austria.

Rhee, Y., Kim, J., Chung, A., 2006. Your phone

automatically caches your life. In ACM Interactions

13, 4.

Tolos, M., Tato, R., Kemp, T., 2005. Mood-based

navigation through large collections of musical data.

Consumer Communications and Networking

Conference, 2005.

Toninelli, A., Khushraj, D., Lassila, O., Montanari, R.,

2008. Towards Socially Aware Mobile Phones, In

Proceedings of SDoW 2008, Karlsruhe, Germany.

Yoon, Y., Ahn, Y., Lee, G., Hong, S., Kim, M., 2008.

Context-Aware Photo Selection for Promoting Photo

Consumption on a Mobile Phone. In Proceedings of

MobileHCI’08, Amsterdam, Netherlands.

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

498