Resolving Uncertainty in
Context Integration and Abstraction
∗
[Context Integration and Abstraction]
Juan Ye, Susan McKeever, Lorcan Coyle, Steve Neely and Simon Dobson
System Research Group, UCD
Dublin, Ireland
juan.ye@ucd.ie, susan.mckeever@ucd.ie, lorcan.coyle@ucd.ie, steve.neely@ucd.ie,
simon.dobson@ucd.ie
ABSTRACT
Pervasive computing is typically highly sensor-driven, but
sensors provide only evidence of fact rather than facts them-
selves. The uncertainty of sensor data will affect each com-
ponent in a pervasive computing system, which may de-
crease the quality of its provided services. We provide a
general model to represent semantics of uncertainty in dif-
ferent levels (e.g., sensor, lower-level context and higher-
level context). Within our model, fine-grained approaches
are applied to evaluate and propagate uncertainties. They
will help to resolve the uncertainty in each process of con-
text management so that the effect of uncertainty on system
services will be minimised.
Categories and Subject Descriptors
D2.2 [SOFTWARE ENGINEERING]: Design Tools and
Techniques; I2.4 [ARTIFICIAL INTELLIGENCE]: Knowl-
edge Representation Formalisms and Methods
General Terms
Management
Keywords
Context-aware computing, Context integration, Context ab-
straction, Uncertainty, Bayes’ Theorem
∗
This work is partially supported by Science Foundation
Ireland under grant numbers 05/RFP/CMS0062 “Towards
a semantics of pervasive computing”, 04/RPI/1544 “Secure
and predictable pervasive computing”, and Enterprise Ire-
land under grant number CFTD 2005 INF 217a, “Platform
for User-Centred Design and Evaluation of Context-Aware
Services”.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
ICPS’08, July 6–10, 2008, Sorrento, Italy.
Copyright 2008 ACM 978-1-60558-135-4/08/07 ...$5.00.
1. INTRODUCTION
Pervasive computing aims to provide services that respond
directly to their user and environment with minimal intru-
siveness and inherent pro-activity. This is achieved by as-
suming a number of invisible sensing and computational de-
vices in an environment, which collect information about
users and the environment. With the help of these devices, a
pervasive computing system can deliver customised services
to users in a contextual manner. Data in pervasive com-
puting environments may be generated by untrustworthy or
inaccurate sources and so should be taken “with a grain of
salt”. Because components of a pervasive computing envi-
ronment deal with the real world, they come with certain
caveats: sensors in the field are inherently inaccurate, since
they could break down; or they could report inaccurately
because they come up against a phenomenon for which they
have not been designed [26]. Uncertainty may have an in-
fluential effect on the quality of services that a pervasive
computing system provides. Inaccurate sensor data may re-
sult in misunderstanding of a user’s or an environmental
state, which leads to incorrect behaviour. Therefore, the
issue of resolving uncertainty must be taken into account
when dealing with pervasive computing systems.
Context-awareness is an enabling technology for perva-
sive computing. A context-aware computing system exhibits
appropriate and customised behaviours that adapt to the
change of users’ context. Context can be any information
that is used to characterise the situation of service con-
sumers, which includes information about consumers, their
environment, or their tasks [8]. Context is acquired from
various kinds of sensors that are distributed in a pervasive
computing environment. It can be sensed from physical de-
vices, profiled from users, or derived from application- or
meta-information existing in systems [29]. The uncertainty
in sensor data will be transferred to context uncertainty
and propagated through all the processes of context man-
agement. The question is how to model the semantics of
context uncertainty and how to resolve (or minimise) uncer-
tainty by distilling the most accurate context from a large
number of trivial and noisy contexts.
Our work aims to propose a fundamental model of context
uncertainty that represents semantics of context uncertainty
and exhibits fine-grained approaches to evaluate and resolve
uncertainty when processing and using context. To accom-
plish this, we analyse the essential characteristics of context
types and constituents of context values, based on which
131

different types of context uncertainty will be discussed: out-
of-date, incomplete, imprecise,andinaccurate [13]. We will
discuss how context uncertainty is acquired from sensor un-
certainty and explore the uncertainty propagation issue in
two processes of context management: context integration
and context abstraction. Context integration is about ex-
tracting the most accurate context from a number of noisy
and conflicting contexts. Context abstraction is about de-
riving a higher-level application-interesting contexts (for ex-
ample, a user state being in a “meeting” or “working”) from
a number of lower-level contexts (such as a user’s location
or a temperature in a room).
The remainder of this paper is organised as follows. Sec-
tion 2 investigates various sensor uncertainties in different
types of sensors. Section 3 provides a general definition of
context from the perspective of representing a context in
a real system. Within this definition, we explore different
types of uncertainty. Section 4 and Section 5 discuss fine-
grained approaches in resolving the uncertainty during con-
text management. Section 6 demonstrates the feasibility of
our model and provides a preliminary evaluation result. Sec-
tion 7 compares our work with the recent research works in
dealing with context uncertainty. Finally, in Section 8 we
summarise our work and outline the future direction of this
research.
2. SENSORS
Pervasive computing systems operate in large, open and
ever-changing environments, where a huge number of sen-
sors of different types are involved. These sensors can be
categorised into types according to the type of information
they provide: environmental sensors are those which gener-
ate information from the real world, for example noise level,
temperature, humidity, etc.; positioning sensors that locate
or track the movement of objects; device sensors that report
the state of hardware and equipment, e.g., whether a printer
is busy, idle, or off; virtual sensors that extract information
from other software or applications (for example, a virtual
sensor could be used to mine schedule information from an
online calendar).
A typical approach to representing the characteristics and
possible imperfections of sensed data is to describe sensor fi-
delity as meta-information in a quality matrix. We assume
that there should exist different types of quality matrix for
each category of sensor and individual sensors. Further, dif-
ferent types of sensors should have different quality parame-
ters that can be applied to the data they output.
We propose a general quality matrix that can be used
to describe any type of sensor. The metadata consists of
frequency, with a list of accuracy and precision pairs. Fre-
quency is defined as the sample rate – how often the sensor
data is updated. The resolution and frequency are deter-
mined by the technical specification of sensors given by the
manufacturer. Precision defines the range and accuracy is
the percentage of how often the accuracy is achieved [10].
Different ranges of precision result in different accuracies.
For example with our in-house location system Ubisense, to
achieve 70% accuracy, the precision on x- and y-axis are 3.30
and 2.22 meters. Accuracy and precision can be acquired
through different approaches for example training from ex-
perimental set up or calculated from component diagnostics.
The quality matrix of a given sensor should be referenced
by all sensed data when it is produced. The general quality
matrix as we describe is not definitive – it should be ex-
tended with more quality parameters for a particular type
of sensor.
These sensors are the inputs that drive the production and
derivation of context. This implies that the imperfections of
sensed data are one of the causes of context uncertainty.
There are at least three factors that are responsible for im-
perfect sensor data:
• technical limitation of sensors: each sensor is produced
with inherent errors. This is due to the manufacturing
process and hardware limitations. When sensors are
installed, they may suffer from breakdown, disconnec-
tion from network, or signal delay;
• environment noise : the accuracy of some sensors may
be subject to radio interference, temperature, humid-
ity, sound noise or reflective materials which signals
bounce off;
• and users: the configuration of sensors by users may
affect the accuracy of sensors. In addition, especially
for physical devices (like tag-based positioning sen-
sors), the reliability of their data will be decreased if
users do not correctly use them.
Technical limitations of sensors are reasonably fixed. They
can be provided by the manufacturer or empirically cal-
culated after installation. We can combine these values
with environmental noise which is intermittently gathered
through sampling and machine learning algorithms. In con-
trast, the influence of users on the process of gathering ac-
curately sensed data is far more unpredictable.
When users are taken into account, the confidence of the
sensor data can be computed by a function that takes the
precision and the impact factor of the use. For example, in
Middlewhere [20], the confidence on data from a tag-based
location sensor is the product of its accuracy and the prob-
ability that a user wears a tag.
3. CONTEXT AND UNCERTAINTY
3.1 Context
Context can be categorised into different data types ac-
cording to their particular properties. Each context type
indicates a set of context values that represent reality enti-
ties or one property (that is, aspect) of reality entities. For
example, the Location context type contains a set of location
data that represent the location property of entities, such as
a coordinate or a place with a human-friendly name [27]; the
Person context type that contains a set of person entities, or
social communities; and the Environment context type that
contains a set of physical properties about an environment
entity like temperature, humidity, or noise level.
A context type has a set of ground values, labelled as V
g
,
that are the irreducible (the smallest perceivable grained)
elements. The tangible context value of a context type is
defined as a set of its ground values through a mapping re-
lationship: m : V → 2
V
g
. For example in the Location con-
text type, the ground values are a set of single coordinate
points, like [12.22,5.26,0.09], and a tangible context value
“Lecture Room 01” maps to a set of coordinates that are
in this room. In the Temperature context type, the ground
values are a set of individual degrees in a certain unit (e.g.,
132