Content uploaded by Theodoros Thodoris Anagnostopoulos
Author content
All content in this area was uploaded by Theodoros Thodoris Anagnostopoulos on Apr 10, 2015
Content may be subject to copyright.
Predicting the Location of Mobile Users: A Machine
Learning Approach
Theodoros
Anagnostopoulos
Pervasive Computing Research
Group, Communication Networks
Laboratory, Department of
Informatics and
Telecommunications, University of
Athens, Panepistimiopolis, Ilissia,
Athens 15784, Greece, tel:
+302107275409
thanag@di.uoa.gr
Christos Anagnostopoulos
Pervasive Computing Research
Group, Communication Networks
Laboratory, Department of
Informatics and
Telecommunications, University of
Athens, Panepistimiopolis, Ilissia,
Athens 15784, Greece, tel:
+302107275409
bleu@di.uoa.gr
Stathes Hadjiefthymiades
Pervasive Computing Research
Group, Communication Networks
Laboratory, Department of
Informatics and
Telecommunications, University of
Athens, Panepistimiopolis, Ilissia,
Athens 15784, Greece, tel:
+302107275409
shadj@di.uoa.gr
Miltos Kyriakakos
Pervasive Computing Research Group,
Communication Networks Laboratory, Department of
Informatics and Telecommunications, University of
Athens, Panepistimiopolis, Ilissia, Athens 15784,
Greece, tel: +302107275409
miltos@di.uoa.gr
Alexandros Kalousis
Artificial Intelligence Laboratory, Department of
Computer Science, University of Geneva,
Uni-Dufour, Geneva 1211, Switzerland, tel:
+41223797630
Alexandros.Kalousis@cui.unige.ch
ABSTRACT
Mobile context-aware applications experience a constantly
changing environment with increased dynamicity. In order
to work efficiently, the location of mobile users needs to be
predicted and properly exploited by mobile applications. We
propose a spatial context model, which deals with the
location prediction of mobile users. Such model is used for
the classification of the users' trajectories through Machine
Learning (ML) algorithms. Predicting spatial context is
treated through supervised learning. We evaluate our model
in terms of prediction accuracy w.r.t. specific prediction
parameters. The proposed model is also compared with other
ML algorithms for location prediction. Our findings are very
promising for the efficient operation of mobile context-
aware applications.
Categories and Subject Descriptors
I.5.2 [Pattern Recognition]: Design Methodology – Pattern
analysis.
General Terms
Algorithms, Performance.
Keywords
context-awareness, machine learning, location prediction,
spatial context representation.
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, or
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
ICPS’09, July 13–17, 2009, London, United Kingdom.
Copyright 2009 ACM 978-1-60558-644-1/09/07...$5.00.
1. INTRODUCTION
During the recent past, we have witnessed an
impressive growth in the domains of wireless-mobile
telecommunications and context-aware services. The
general framework that evolved from
distributed systems, and is currently described under
the term mobile computing has attracted the extensive
interest of academia and the industry. New protocols,
schemes, applications and services are developed and
applied in real-life situations. This escalating
development has also triggered discussions on other,
more enhanced paradigms like pervasive computing.
In order to render mobile applications intelligent
enough to support modern users everywhere / anytime
and materialize the so-called ambient intelligence,
information on the present context of the user has to be
captured and processed accordingly. Contextual
information may refer to the user’s position, time,
physical properties like temperature or other general
parameters (e.g., the specific devices that the user
carries). The efficient management of context requires
detailed and thorough data modeling along with
specific processing, classification, inference and
prediction capabilities.
A well-known definition of context is the following:
“context is any information that can be used to
characterize the situation of an entity. An entity is a
person, place or object that is considered relevant to
the integration between a user and an application,
including the user and the application themselves”
65
[26]. Context refers to the current values of specific
parameters that represent the activity of an entity.
Context-awareness allows an entity to adapt to its
environment, thus, offering a number of advantages
and possibilities for new applications. One of the more
intuitive capabilities of mobile context-aware
applications is their pro-activity. Predicting user
actions and contextual parameters enables the
development of new, advanced applications.
Moreover, predicting the location of a mobile user is
an inherently interesting and challenging problem.
Location prediction has received increased attention
driven by applications in location management, call
admission control, smooth handoffs, and resource
reservation for improved quality of service. Spatial
information prediction can be used to facilitate the
provision of advanced location-based services by
preparing and feeding them with the appropriate
contextual information (in advance). Pro-active
context-aware applications address context pre-
evaluation / pre-determination introducing innovative
proactive services, e.g., alerts related to traffic
conditions, content pre-fetching and triggering
actuation rules in advance (download company
documents while entering downtown).
Context pre-evaluation may match with information
classification and prediction, in the sense that, the
values of certain contextual parameters are estimated
/evaluated, clustered and classified in advance. The
concept of predicting context through ML algorithms
and techniques is quite novel. ML refers to the study of
algorithms that improve automatically through
experience. A wide spectrum of applications based on
ML algorithms relates to search engines, medical
diagnosis, object recognition in computer vision and
natural language processing. ML tasks can be
addressed through: supervised learning, (e.g.,
classification and regression), and unsupervised
learning, (e.g., clustering and rules extraction). In this
paper, we adopt classification in order to predict spatial
context - location. The (training) examples used for the
supervised learning task are vectors of attribute-value
pairs. Each example is assigned to a specific class from
a fixed set of classes (e.g., symbolic locations). The
goal of the classification task is to develop a model
from a set of examples capable of predicting the class
of a given unseen example.
A context model is proposed in order to support
location prediction for mobile users. Such model
predicts, with a certain accuracy level (portion of
successful predictions), the future position (cell) of a
mobile user in a cellular environment and can be used
for the pro-active management of network resources
(e.g., packets, proxy-cache content). The model can be
trained using a variety of ML classification algorithms.
In this paper we experiment with such algorithms, in
order to perform location prediction. We can divide the
assessed algorithms into two broad categories
according to the way they construct their classification
models: standard classification algorithms, (e.g.,
decision trees, k-nearest neighbors, support vector
machines), and ensemble learning algorithms that learn
to combine classification models constructed by base
learning algorithms in order to improve predictive
performance. Typical examples of this last category are
the boosting and voting schemes. Finally, the
performance of the ML models is compared with that
of the most cited location prediction models namely
the LeZi-Update [12] and Mobile Motion Prediction
(MMP) [13] algorithms.
This paper is organized as follows: Section 2 reports
certain ML schemes used for classification while in
Section 3 the context representation model taking into
account the user mobility pattern is proposed. In
Section 4, we evaluate the discussed model and we
compare it with the certain ML schemes. Section 5
reports prior work on that research area and, finally,
Section 6 concludes the paper.
2. CLASSIFICATION IN MACHINE
LEARNING
Classification is the task of learning to categorize
(predict) an unseen example to a discrete class value.
An example is a (m+1)-dimensional vector e of m
attributes ei, i = 1, …, m and a class attribute em+1, that
is e = [e1, …, em, em+1]. The ei attributes determine the
value of the class attribute em+1 of e. An ei attribute
assumes values from the corresponding domain
Dom(ei). The input of a classification algorithm is a set
E of s example vectors, E = {ei, i = 1, …, s}, and the
output is a classification model M(E). Such model is
capable of predicting / estimating the value of the class
attribute of an unseen and yet unlabeled (m+1)-
dimensional example vector q based only on the values
of its qi attributes with Dom(ei) = Dom(qi), ∀i.
The performance of a classification algorithm is
estimated on the basis of the quality of predictions
delivered by the trained models. In order to estimate
the classification performance, a test phase is required.
In the test phase we employ a test set V that was not
used in the training phase. A classification model,
M(E), is established through the training set, E, as a
result of the learning phase. The produced model M(E)
is then applied on the test set V and the correctness of
its predictions is assessed and quantified. Prediction
accuracy ε is a quantitative measure for the
classification performance. It refers to the proportion
66
of the correctly predicted examples C ⊂ V out of V,
i.e., the fraction ε = |C| / |V|, where |C| denotes the
cardinality of C. One method to estimate the prediction
accuracy of a classification algorithm is a re-sampling
method called cross-validation [1, 2]. In n-fold cross-
validation, the training set E is divided into n subsets of
equal size. A different model M is trained n times, each
time setting aside one of the subsets which will be used
as the test set on which the predictive accuracy will be
computed. The final accuracy estimation is the
averaged accuracy over the n different repetitions of
the training and testing phases.
A number of different learning paradigms have been
developed over the years for classification. Since there
is no single classification algorithm that is better than
all the others irrespective of the application domain,
each time face a new classification problem we have to
assess anew the suitability of the algorithms. We have
experimented with several classification algorithms
trying to cover a range, as broad as possible, of
different learning paradigms. We found a voting
ensemble-learner to be the best for our application
problem according to the estimated predictive
performances. In the next paragraphs we give a brief
description of the different types of learning paradigms
that we are going to consider in this paper.
2.1 Bayesian learning
Bayesian classification algorithms are statistical
learning algorithms based on the Bayes theorem. The
Naïve Bayes algorithm (the simplest Bayesian
classifier) [5] assumes that, the effect of the value of an
attribute on the class attribute is independent of the
values of the other attributes given the value of the
class attribute (conditional independence).
2.2 Decision Tree learning
A decision tree consists of decision nodes and leaves.
A leave is usually associated with a single class; the
majority class of the training examples that arrive to
that leave. Splits are introduced in the building of the
tree according to the outcome of a function f
(information gain ratio). When examples are classified
a function is used in each split to determine the
downstream path to be followed [6]. An indicative
decision tree-based algorithm is the C4.5 classifier [7].
2.3 Rule-Induction learning
Rule-induction performs a depth-first search in a graph
G(V,E) generating one path. V is a set of attributes and
E is the set of edges denoting dependencies among
attributes. Such path represented as a classification
rule [8] is a conjunction of conditions with discrete or
numeric attributes. A rule is said to cover an example if
the later fulfills all the conditions of that rule. A
representative rule-induction algorithm is the RIPPER
( Repeated Incremental Pruning to Produce Error
Reduction) [9].
2.4 Instance-based learning
An instance-based algorithm uses a distance function
||e - q|| in order to determine which example vector e of
the training set is closest to an un-classified example q
(or instance). Once the nearest example vector has
been determined, its class label is selected as the class
label for q. A representative instance-based algorithm
is the k-nearest neighbor classifier.
2.5 Ensemble-Learning Algorithms
Ensemble-learning algorithms combine a number of
base classification models, classifiers, produced by
different learning algorithms or by different training
sets, in order to achieve better classification
performance than their constituents. Base models can
be combined in different ways in order to generate
ensemble-learning algorithms. Popular ensemble-
learning algorithms are the following:
2.5.1 Voting
Each base classifier predicts, votes, a class. The final
class is that, which assumes the greatest number of
votes.
2.5.2 Bagging
Several training sub-sets Ei are formed from the initial
training set E by random re-sampling with
replacement. A base classifier is learned from each
training sub-set. The final class is determined by voting
of the base classifiers.
2.5.3 Boosting
In boosting also the diversity of the base classifiers is a
result of different training sets. The method works in
an iterative manner re-sampling the current training
dataset by giving higher resample weights to instances
that are hard to classify. The final class is determined
by a weighted voting of the base classifiers where the
weights are determined on the basis predictive
performance of the base classifiers. A typical example
of a boosting algorithm is AdaBoost M1 boosting [6].
3. CONTEXT REPRESENTATION
3.1 Spatial Context Model
The contextual information considered for
classification refers to the history of user movements.
Such history is represented by a (m+1)-dimensional
vector e of m time-ordered visited locations ei, that is,
ej < ei if the user visited location ej before ei, i, j = 1,
67
…, m. Such locations refer to the antecedent-part of a
classification rule while the consequent-part is also the
location em+1, which is the predicted location. In the
considered context model, the user roams through a
cellular network thus a network cell with a unique
identifier represents a location. All attributes of the e
vector assume values in the set Dom(e) of network
cells identifiers. Assume that a user u is currently
positioned at cell e, which becomes the point of
reference. We want to predict which cell the user is
going to move to in the next transition. Let eS(u) denote
the sequence of the last m cells from which user u went
through together with the class attribute, em+1, i.e., the
cell to which he/she is going to move in the next
transition. Note that the eS(u) vector does not model
time. Instead, it only models cell transitions. Then,
from the complete sequence of transitions of a user, we
extract a number of transition sequences, eS(u), applying
a sliding window of length m+1, where :
eS(u) = [e1, e2, …, em, em+1] (1)
In the above vector em is the cell in which the user is
currently positioned, em+1 is the cell to which he/she is
going to move in the next transition and which
constitutes the prediction target and [e1, e2, …, em-1] is
the sequence of cells from which the user passed
before reaching em. We explain the above model
through an example. Consider a set of cells Dom(e) =
{e1, e2, e3, e4, e5}. Assume a user that had the following
sequence of transitions [e1, e2, e3, e4, e5]. Applying a
sliding window of length 3 derives the following three
eS(u) [e1, e2, e3], [e2, e3, e4], and [e3 , e4, e5]. Obviously
the corresponding class labels are e3, e4, e5. Figure 1
depicts an area divided into cells in which several
sliding windows of a user movement are shown.
Movements:
t
1
t
2
t
3
Network
Cell
Figure 1. The three eS(u) vectors of a sliding window of length m = 4.
Note that the sliding windows are overlapping.
3.2 User Mobility Profile
We introduce the parameter degree of movement
randomness, δ ∈ [0, 1], in order to express the mobility
behavior of a user, i.e., the way a user transits between
cells and changes directions. Such degree is used for
assessing the performance of the proposed models
under various levels of uncertainty and unpredictability
w.r.t. mobility behavior. The δ degree denotes the
possible transition patterns of a user trajectory between
locations. A certain trajectory can derive either from a
deterministic movement (assuming a low value of δ) or
a random movement (assuming a high value of δ). The
adoption of δ provides an objective criterion for
assessing movement prediction algorithms. It allows a
correct interpretation of the performance evaluation
results (e.g., a high accuracy may not necessarily
indicate an efficient algorithm if the testing patterns
were quite deterministic).
The deterministic trajectories represent regular
movements (e.g., the route from home to work). On the
other hand, random trajectories represent purely
random movements between predefined locations (e.g.,
a quick detour for a coffee after leaving home and
before getting to work). Therefore, a value of δ ~ 1.0
does not mean an explicitly non-deterministic mobility
behavior. Instead, such a movement (i.e., δ ~ 1.0) is
constrained by obstacles in the examined space.
In our experiments, we adopted the mobility pattern
generator discussed in [10]. Through this generator we
obtained trajectories with specific δ values in the set
{0.0, 0.25, 0.5, 0.75, 1.0}. The five discrete values of δ
range from the regular pattern (δ = 0.0 with 500
example trajectories) to completely random pattern (δ
= 1.0 with 1000 example trajectories). It should be
noted that, the value of δ influences the size of the
training patterns (movement history) since the more
random the movement is, the more transitions are
generally required for a certain itinerary (i.e., moving
from a given origin to a given destination).
Table I. The value of the majority class ρ w.r.t., degree of randomness
δ.
Degree of Randomness
(δ)
Majority Class ratio (ρ),
(Default accuracy)
0% 9.06%
25% 6.53%
50% 5.48%
75% 3.81%
100% 3.69%
4. CONTEXT MODEL EVALUATION
We derive a number of context models that, in fact,
correspond to the different classification models build
from the different classification algorithms that we
consider, and examine their relative predictive
performance. All the considered classification
algorithms are provided with the Weka machine-
learning workbench [4]. We represent user trajectories
68
through a series of waypoints. Each waypoint is
defined by the location in terms of a cell and speed.
The number of cells is 100, hence, |Dom(em+1)| = 100.
Traces were obtained by the RMPG tool [10] which
allows a controllable degree of randomness. We have
used five discrete categories of randomness from the
regular pattern (δ = 0.0 with 500 training instances) to
completely disordered trajectories (δ = 1.0 with 1000
training instances).
We experimented with the following classifiers from
Weka: (i) the Naïve Bayes classifier, (ii) the J48
Decision Tree-based classifier (an implementation of
the C4.5 algorithm), (iii) the JRip Classification Rule
classifier (an implementation of the RIPPER
algorithm), and (iv) the IBk incremental algorithm (an
implementation of k-nearest neighbor algorithm). From
the category of the Ensemble learning algorithms we
experimented with: (i) Voting with the base classifiers
constructed by J48 and Ibk, (ii) Bagging were the base
classifiers were learned by IBk, and finally (iii) the
AdaBoost M1, where the base classifiers were also
learned by IBk. The prediction accuracy ε for each
classification algorithm is estimated by 10-fold cross-
validation.
We assessed the level of statistical significance for
the differences in the accuracies of the different
classification algorithms using the t-test [11]. We
extensively evaluated the performance of models
produced by the best classification algorithm under
different levels of movement randomness (δ) and
different sizes m of the sliding window (parameter m).
4.1 Classifier Selection
At first we have to compare the predictive performance
of each classification algorithm with the default
accuracy, i.e., the prediction accuracy obtained from a
naïve classifier that always predicts the majority class.
A classification algorithm is considered appropriate for
a certain classification problem if its classification
accuracy is significantly better than the default
accuracy. Table I depicts the values of default accuracy
w.r.t., the different levels of randomness δ.
Figure 2 depicts the prediction accuracy ε of the IBk,
J48, Naïve Bayes, JRip, Vote, Bagging and AdaBoost
M1 classifiers for the eS(u)representation. The
prediction accuracy of all classification algorithms is
significantly better than the default accuracy for any
given δ (see Table I). The classification algorithm that
achieves the top performance (for all levels of
randomness) is Vote.
Table II depicts in the t-test results of all the pairs of
classifiers for δ = 0.25. The (i,j) element of the matrix
denotes the statistical significance of the difference
between the classification algorithm of the ith row and
that of the jth column. Specifically, if the element at
(i,j) is ‘0’, then the difference |εi - εj| between the ith and
the jth classification algorithm is insignificant. If the
element at (i,j) is ‘*’ then the prediction accuracy of
the ith classification algorithm is better than that of the
jth algorithm and this is validated through statistical
significance test; if the element at (i,j) is ‘v’ then the jth
classifier is better. The threshold of the statistical
significance test was set to a = 0.05.
00.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
10
20
30
40
50
60
70
80
90
100
Degree of randomness (δ)
Prediction accuracy –ε-(%)
Naïve Bayes
IBk
J48
JRip
AdaBoost (IBk)
Bagging (IBk )
Vote (IBk, J48)
Figure 2. The behavior of the prediction accuracy ε of classifiers vs.
degree of randomness δ.
00.2 0.4 0.6 0.8 1
0
10
20
30
40
50
60
70
80
90
100
Degree of randomness (δ)
Prediction accuracy –ε-(%)
m=4
m=8
m=12
m=16
m=20
m=2
Figure 3. The behavior of the prediction accuracy ε of eS(u) -Vote vs. the
window length m.
4.2 Experimenting with the window length
An important factor of the proposed model is the
selection of the sliding window size m. The appropriate
value of m is estimated by experimentation. In Figure
3, we can observe the prediction accuracy of the eS(u)-
Vote for 2 ≤ m ≤ 20 with a training set size s = 40 days.
The best results are obtained for m = 4. The case m = 4
assumes better prediction results even when the user
exhibits a high degree of randomness. In case where
the size m is small, the considered vectors refer to
69
many different class attributes (almost equi-probable),
thus, the prediction accuracy decreases. As long as the
size m raises then the considered vectors of cell
identifiers is quite large to achieve a close match
(between sample and test data), especially, when
randomness increases. The size m = 4 reflects the
regularity of a common user movement as reported in
[27].
Table II. Significance difference between the classifiers.
classifier
N
a
ï
ve
Bayes
IBk
J48
JRip
Ad
a
B
oos
t
M1
Bagging
Vote
Naïve
Bayes - v v v v v v
IBk - - * * 0 * v
J48 - - - * v 0 v
JRip - - - - v v v
AdaBoost
M1 - - - - - * v
Bagging - - - - - - v
Vote - - - - - - -
4.3 Comparison with other Machine Learning
algorithms
We compare eS(u)-Vote with the LeZi-Update [12] and
MMP [13] schemes by means of prediction accuracy.
Specifically, in [14], the mobility tracking problem in a
cellular network has been considered in the
information theoretic framework. Comparison of user
mobility models has been based upon the concept of
entropy. A dictionary of user’s path updates is built
and maintained by the proposed scheme. Such
dictionary supports an adaptive online algorithm that
learns the profiles of users. This technique is based on
ideas and concepts coming from the area of lossless
compression and, specifically, the Lempel-Ziv
algorithm. This algorithm is also called “LeZi-update”
and is exploited to reduce the location update related
costs while its predictive power is used to reduce
paging cost. In [12], the LeZi-Update scheme is
applied in an intelligent home-environment in order to
track down an inhabitant, both inside and within
surroundings in order to satisfy connectivity
requirements. After processing the input trace of the
user through the LeZi-update algorithm, a blending
strategy called exclusion is used to predict the next
location of the user (see Figure 4).
Moreover, the algorithm discussed in [13] is based
on Mobile Motion Prediction (MMP) scheme for the
prediction of the future location of a roaming user
according to his / her movement history pattern. The
scheme consists of Regularity-Pattern Detection (RPD)
algorithms and Motion Prediction Algorithm (MPA).
Regularity Detection is used to detect specific patterns
of user movement from a properly structured database
(IPB: Itinerary Pattern Base). Two RPD algorithms are
proposed: (i) Movement Circle (MC) for the detection
of long way routes and (ii) Movement Track (MT) for
the detection of regular routes. Three classes of
matching schemes are used for correlation analysis of
the MCs or MTs namely the state matching, the
velocity or time matching and the frequency matching.
The Motion Prediction Algorithm (MPA) is invoked
for combining regularity information with stochastic
information (and constitutional constraints) and thus
reach a decision - prediction for the future location (or
locations) of the mobile user. Figure 5 shows an
overview of the suggested scheme.
Figure 6 depicts the prediction accuracy of the eS(u)-
Vote with that of the LeZi-update and MMP schemes
with m = 4 and s = 40 days for four same training sets.
Specifically, eS(u)-Vote achieves better performance
than the other two algorithms for each degree of user
randomness.
LeZi-update
INPUT
Exclusion
Prediction
Dictionary
DB
LeZi-update
INPUT
Exclusion
Prediction
Dictionary
DB
LeZi-update
INPUT
Exclusion
Prediction
Dictionary
DB
LeZi-update
INPUT
Exclusion
Prediction
Dictionary
DB
Figure 4. The LeZi-update scheme.
Regularity
Detection
Algorithm
Motion
Prediction
Algorithm
Stochastic
Processes,
Markov
Chain, Constitution
Itinerary Pattern
Base (IPB)
Input
Random Regularity
Prediction Output
Source: Liu - Maguire, 1996
Regularity
Detection
Algorithm
Motion
Prediction
Algorithm
Stochastic
Processes,
Markov
Chain, Constitution
Itinerary Pattern
Base (IPB)
Input
Random Regularity
Prediction Output
Source: Liu - Maguire, 1996
Figure 5. The Predictive Mobility Management Algorithms.
5. PRIOR WORK
There are a lot of prediction models based on
Machine Learning techniques. Specifically, the
70
probabilistic model in [15] is based on the user
movement history of handover behavior. This model
considers the history of all handovers that occurred in a
given cell using the Naïve Bayes classification. The
authors in [16] report a probability - based and a
learning - based model for trajectory prediction. The
algorithm in [17] predicts the next inter - cell
movement of a mobile user. The user mobility patterns
are mined from the history of the trajectories resulting
in mobility rules extraction. The location prediction is
based on such rules.
00.2 0.4 0.6 0.8 1
0
10
20
30
40
50
60
70
80
90
100
Degree of randomness
Prediction accuracy (%)
model - V ote
LeZi Update
MMP
(δ)
(ε)
Figure 6. The behavior of the prediction accuracy ε of eS(u) -Vote vs.
the LeZi-update and MMP algorithms.
The authors in [18] implement an algorithm based on
user mobility patterns discovery. Such patterns, which
derived from trajectory clustering, are used for location
prediction and dynamic resource allocation. Moreover,
an efficient online (incremental) algorithm that
classifies routes between important locations and
predicts the future location is reported in [19].
Specifically, clusters of cell - sequences are built to
represent physical routes. The prediction is based on
destination probabilities and temporal reasoning. A
data - mining algorithm is proposed in [20], which
efficiently discovers sequential mobile patterns. These
patterns exploit both spatial and application context
(e.g., service requests). Moreover, the mobility
tracking in a cellular network is based on information
theory by using the compression Lempel-Ziv algorithm
[21]. The algorithm [13] consists of the regularity-
pattern detection and the model prediction processes
used for predicting the user movements. The work
presented in [22] discusses pattern matching
techniques and extended self-learning Kalman filters in
order to estimate the future location. In addition, a
learning automaton that follows a linear reward -
penalty scheme is used in [23] in order to facilitate
location prediction. The authors in [24] apply
evidential reasoning based on the Dempster-Shafer
theory in mobility prediction when adequate
knowledge about the history of user’s travelling
patterns is not available. Finally, the authors in [25]
refer to a conceptual context prediction model based on
geometrical location prediction.
6. CONCLUSIONS
We propose a context model for location prediction
based on the user mobility behaviour. We present how
Machine Learning is applied to context-awareness for
predicting future locations in pervasive computing
environments. The proposed context model exploits the
user history and degree of movement randomness in
order to classify and predict future movements. The
model is evaluated with the 10-fold cross validation
method with sets of simulated user movements with
varying degrees of randomness. We experiment with
several ML classifiers and evaluate the model through
certain parameters derived from the ML field in order
to choose the appropriate classifier for location
prediction. Our findings show that, the Voting
classification scheme is appropriate for location
prediction since it exhibits satisfactory prediction
results for diverse user mobility behaviour. Moreover,
we compare the performance of the proposed eS(u)-Vote
model with that of the LeZi-Update and MMP
algorithms justifying the importance of the ML
classification in predicting spatial context.
Our model can be enhanced with more semantics and
contextual information, like temporal context (e.g.,
time period within a day), application context (e.g.,
service requests), proximity of people (e.g., social
context) and destination / velocity of the user
movement. Finally, the combination of several local
classification models has to be considered in order to
exploit local spatial and temporal contextual
information.
7. REFERENCES
[1] M. Weiss, and C.A. Kulikowski, Computer Systems That
Learn: Classification and Prediction Methods from Statistics,
Neural Networks, Machine Learning and Expert Systems,
Morgan Kaufmann, 1991.
[2] M. Plutowski, S. Sakata and H. White, Cross-validation
estimates integrated mean squared error, Advances in Neural
Information Processing Systems, vol. 6, 1994.
[3] R. Duda, P. Hart, and D. Stork, Pattern Classification, Wiley-
Interscience, 2001.
[4] I. H. Witten and E.Frank, Data Mining: Practical Machine
Learning Tool sand Techniques, Morgan Kaufmann Series in
Data Management Systems, 2005.
71
[5] J. Han and M. Kamber, Data Mining: Concepts and
Techniques, Morgan Kaufmann Series in Data Management
Systems, 2001.
[6] E. Alpaydin, Introduction to Machine Learning, The MIT
Press, 2004.
[7] J. R. Quinlan, C4.5: Programs for Machine Learning, Morgan
Kaufmann Series in Machine Learning, 1993.
[8] M. Mahmood, M. Zonoozi, and P. Dassanayake, User
Mobility Modeling and Characterization of Mobility Patterns,
IEEE Communications, vol. 15, no. 7, 1997.
[9] W. Cohen, A. Prieditis and S.J. Russel, Fast Effective Rule
Induction, Proc .Int. Conf. on Machine Learning, 1995.
[10] M.Kyriakakos, N.Frangiadakis, S.Hadjiefthymiades and
L.Merakos, RMPG: A Realistic Mobility Pattern Generator
for the Performance Assessment of Mobility Functions,
Simulation Modeling Practice And Theory, vol. 12, no. 1,
Elsevier, 2004.
[11] F. Pernkopf, J. Bilmes, Discriminative versus Generative
Parameter and Structure Learning of Bayesian Network
Classifiers, ICML, 2005.
[12] S.K.Das, D.J.Cook, A.Bhattacharya, E.O.Helerman and T-
Y.Lin, The Role of Prediction Algorithms in the MavHome
Smart Home Architecture, IEEE Wireless Communications,
vol. 9, no. 6, pp. 77-84, December 2002.
[13] L. George and G. Maguire, A class of mobile motion
prediction algorithms for wireless mobile computing and
communications, Mobile Networks and Applications 1, J.C.
Baltzer AG, Science Publishers, 1996.
[14] A.Bhattacharya, and S.K.Das, LeZi Update: An Information
Theoretic Approach to Track Mobile Users In PCS Networks,
ACM/IEEE Mobicom ’99, Seattle, USA, August 1999.
[15] S. Choi and K. G. Shin, Predictive and adaptive bandwidth
reservation for hand-offs in QoS-sensitive cellular networks,
Proc. ACM SIGCOMM ’98, 1998.
[16] L. Xiong and H. A. Karimi, Location Awareness through
Trajectory Prediction, Computers, Environment and Urban
Systems, Elsevier, 2006.
[17] G. Yavas, D. Katsaros, O. Ulusoy, and Y.Manolopoulos, A
data mining approach for Location Prediction in Mobile
Environments, Data & Knowledge Engineering vol. 54, 2005.
[18] D. Katsaros, A. Nanopoulos, Clustering Mobile Trajectories
for Resource Allocation in Mobile Environments, Intelligent
Data analysis, 2003.
[19] K. Laasonen, Clustering and Prediction of Mobile User
Routes from Cellular Data, Proc. PKDD, 2005.
[20] V. S. Tseng, and K. W. Lin, Efficient mining and prediction of
user behavior patterns in mobile web systems, Information and
Software Technology vol. 48, Elsevier, 2006.
[21] A. Bhattacharya and S.K. Das, LeZi update: An Information
Theoretic Approach to Track Mobile Users in PCS Networks,
Proc. ACM/IEEE Mobicom ’99, 1999.
[22] T. Liu, P. Bahl and I. Chlamtac, Mobility modeling, location
tracking, and trajectory prediction in wireless ATM networks,
IEEE JSAC vol. 16, no. 6, 1998.
[23] S.Hadjiefthymiades, L.Merakos, Proxies + Path Prediction:
Improving Web Service Provision in Wireless-Mobile
Communications, ACM/Kluwer Mobile Networks and
Applications, Special Issue on Mobile and Wireless Data
Management, vol.8, no. 4, 2003.
[24] A. Karmouch, N. Samaan, A Mobility Prediction Architecture
Based on Contextual Knowledge and Spatial Conceptual
Maps, IEEE Transactions on Mobile Computing, vol. 4 no. 6,
2005.
[25] C. Anagnostopoulos, P. Mpougiouris, S. Hadjiefthymiades,
Prediction Intelligence in Context-aware Applications, Proc.
ACM Mobile Data Management (MDM05), pp. 137-141,
2005.
[26] A. Dey, Understanding and using context, Personal and
Ubiquitous Computing, vol. 5, no. 1, pp 4-7, 2001.
[27] M.Kyriakakos, Mobility Simulation and Path Prediction in
Wireless Mobile Communications, PhD thesis, University of
Athens, Dept. of Informatics and Telecommunications, 2005.
72
