Ew36913917

A. Chaudhary et al Int. Journal of Engineering Research and Applications
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.913-917

REVIEW ARTICLE

www.ijera.com

OPEN ACCESS

Machine Learning Techniques for Mobile Devices: A Review
A. Chaudhary1, S. Kolhe2, Rajkamal3
1,3
2

Devi Ahilya Vishwa Vidyalaya, Indore, Madhya Pradesh, India
Directorate of Soybean Research, Indore, India

Abstract
Machine learning in a very broad sense denotes the acquisition of knowledge, skill and understanding by
machine from guidance, experience or reflection. Machine learning is the study of computer algorithms,
applying these algorithms machines improve automatically with experience. In other words it is the ability of the
computer or mobile program to acquire or develop new knowledge or skills from examples for optimizing the
performance. It is an emerging technology that helps in the discovery of rules and patterns in datasets. This field
has evolved from the field of Artificial Intelligence, which aims to imitate intelligent abilities of humans by
machines. The advent of Mobile technology has fostered development of Machine learning. This paper presents
a review on machine learning and presents a brief study on different machine learning techniques along with
their applications on mobile devices. It also presents an overview of performance related parameters of machine
learning techniques useful for mobile devices.
Keywords Mobile phones; CART, Bluetooth, ANN, GSM, k-NN, Mobile computers, Mobile Intelligent
System, PDA, Pocket PC, Wireless networks, 3G+.

I.

INTRODUCTION

The field of Machine learning has developed
from the field of Artificial Intelligence that aims to
follow intelligent capabilities of humans by machines.
Mitchell [2] accentuates machine learning as a study
of algorithms in computer science that perform better
with experience. In other words it implies computer
programs use their past experience of problem solving
tasks to improve their performance. But description
does not include the concept of acquisition of
knowledge for the stated computer programs.
Alpaydin [3] defines Machine learning as “the
capability of the computer program to acquire or
develop new knowledge or skills from existing or nonexisting examples for the sake of optimizing
performance criterion”. The stream of Machine
Learning has shown potential growth since last 50
years. Alpaydin [3] states the two important factors for
growth in this field are elimination of tedious human
effort and reduction of cost. Machine Learning
techniques when applied to different fields as Medical
diagnosis, Bioinformatics, Speech and handwriting
recognition, Natural Language Processing, Game
playing, Efficient search engine design, Mobile SMS
classification, Mobile learning , Web document
classification, Classifying DNA sequences, , object
identification in Computer vision, Robot Locomotion,
Stock Market Analysis, detecting credit card fraud in
financial institutions have proved to work with large
amounts of data and generate results in a matter of
seconds. Thus, machine learning algorithms work
better in improving the efficiency and accuracy of
intelligent decisions making process by intelligent
computer programmes. Machine learning techniques
have also proved efficient for mobile devices such as
www.ijera.com

Smartcards, Smartphones, Sensors, handheld and
automotive computing systems [20]. Mobile
Technology has fostered development in recent years
with the help of increasing mobile terminals (e.g.
mobile phones, Pocket PC, PDA, mobile computers)
and mobile networks (wireless networks, Bluetooth,
GSM, 3G+ etc.). Machine learning techniques like
Naïve Bayesian, C4.5, Decision trees etc are helpful
for mobile devices. Some important machine learning
applications for mobile devices include Sensor based
activity recognition, Mobile text categorization,
Malware detection on mobile devices, Language
understanding etc. But traditional machine learning
approaches have some limitation:
a) A mobile device has limited computational and
memory capacity which restricts the application
of machine learning algorithms that work with
large data set.
b) A mobile learning environment exhibits concept
drift, for e.g. the concept being learned changes as
the user’s mode of operation changes.
c) A user’s intolerance of any hindrance that
machine learning techniques may introduce in the
normal mobile phone operation.
d) Therefore, a study of different machine learning
techniques on different parameters is of much
importance in order to pave a way for future
works in this direction.

II.

TYPES OF MACHINE LEARNING

Machine learning can be categorized into
major groups as supervised, unsupervised machine
learning and reinforcement learning as shown in Fig.
1. These groups represent how the learning method
works.
913 | P a g e


www.ijera.com

samples with their feature vectors to make prediction.

Figure 1. Machine learning types
Supervised learning
It consists of algorithms that reason or learn
from externally supplied instances to result in a
general hypothesis that makes prediction about future
instances [1].

Reinforcement learning
The machine communicates with its
surroundings by producing certain actions in
reinforcement learning. These actions result in
affecting the state of the environment, which
ultimately result in the machine receiving some scalar
rewards (or punishments).In other words the
agent/machine receives some evaluation of actions
such as a scalar reward or punishment from its
surroundings.

Figure 4. Reinforcement Learning
Figure 2. Supervised Machine Learning
This type of learning works with examples.
In other words machine is supplied with some training
examples to learn the given function. Hence machine
learns the given function and it is capable to work for
extrapolated examples as well. In this type of learning
machine develops the capability to forecast or learn
new concept or given hypothesis on the basis of
examples that are provided. There exists an outcome
or output variable to guide the process of learning.
There are many supervised learning algorithms as
decision trees, K-Nearest Neighbor (KNN), Support
Vector Machines (SVM) and Random Forests [4].
Unsupervised learning
In Supervised learning there is an outcome or
output variable to guide the learning process whereas
unsupervised learning works with models without
predefined classes or examples [5].

Figure 3. Unsupervised Machine Learning
In this category there is no supervisor
available and learning depends upon the guidance
obtained heuristically by the system testing different
data samples or environment. This type of learning
makes use of training samples as well as testing
www.ijera.com

In this category the goal of machine is to
learn to act in such a way that maximizes the future
rewards it receives (or minimizes the punishments)
over its lifetime. This learning is related to decision
theory (in statistics and management science) and
control theory (in engineering).

III.

MACHINE LEARNING
ALGORITHMS

There are various machine learning
algorithms depending upon the domain of their
application [1]. The following Machine learning
algorithms are considered to give readers a basic
understanding of machine learning algorithms.
A. Decision Tree
A decision tree is a classifier that works with
recursive partition of the instance space. It is used to
represent a supervised learning approach. It is a simple
graphical model where non-terminal nodes represent
tests on one or more attributes and terminal nodes give
decision outcomes. This tree consists of one root,
branches, internal nodes and leaves. Each node
corresponds with a certain feature or characteristic or
feature and the branches correspond with a range of
values or decision outcomes. A decision tree works
with local regions that are identified in a series of
recursive splits in a smaller number of steps that
implements divide and conquer paradigm. It is used in
classification and regression related tasks [3]. A
decision tree works with input data and uses decision
rules for future predictions. Decision trees work with
rules that are well understood by humans and used in
knowledge system such as database. The strengths of
using decision tree learning include
a) It works with understandable rules.
b) It performs classification without much of
tedious computation.
914 | P a g e

c) It works well with categorical variables.
d) It provides clear classification and traces
fields
important
for
prediction
or
classification.
The limitations of using decision tree learning are
a. This learning tool is not suitable for
prediction of continuous attribute.
b. It does not perform well with multiple classes
and small data.
c. Training decision tree is computationally
expensive.
B. k-Nearest Neighbor (k-NN)
This is also known as k-NN or Instance
Based learning which a type of supervised learning
algorithm. k-NN works by simply storing in memory
current training dataset, when a new query is fired, a
set of related instances that show resemblance or
neighbors are retrieved from memory. These instances
are further used for classification of new instance [2].
k-NN considers more than one neighbor at a time
while classifying hence it is known as K-Nearest
Neighbor. This nearest neighbor is determined in
terms of Euclidean distance which measures the
dissimilarities between examples represented as vector
inputs and some related measures.
C. Case Based Reasoning
Case based reasoning is a popular technique
by which we solve new problems by making use of
the solutions for the problems solved earlier. This
technique takes the instances of solutions from the
problems solved earlier and tries to solve step by step
new problems by using these cases. Each solution
available is termed as a case [8]. A new case is
defined by the initial description of the problem. A
case is picked up from a set of past cases and this case
is then combined with the new case through reuse
strategy resulting into the solved case. This solved
case becomes a sample solution to the defined
problem. When this solution is known, it is
implemented practically to the real world problems
and therefore it is tested. This testing process is
termed as revision of the problem. The process of
retaining starts after testing process. In retain process
important experience is retained for reusing it in future
and case base is enhanced by a newly learned case or
by updating of some existing cases.
D. Artificial Neural Network (ANN)
ANN is based on Biological nervous system.
Artificial Neural Networks models consists of
nonlinear computational components or processing
units that work in parallel and are arranged in patterns
resembling biological Neural Networks [7]. These
Computational components or nodes are connected via
links that have weights. A set of data (input-output
pairs) is used for training the Neural Network. After
the network has been trained it can be given any input
data (should be from the input space that is to be
www.ijera.com

www.ijera.com

approximated) and results in an output, which would
corresponds to the expected output from the
approximated mapping [8].Single-layer ANN or
Perceptrons work with algorithms like Gaussian
maximum-likelihood classifiers. A multi layer ANN is
used for better decision making in a variety of
problems like disease diagnosis, to test oracle, project
effort and cost estimation, text identification and hand
writing recognition problems etc.
E. Classification and Regression Trees (CART)
It is a well known technique because it
requires less input data for further analysis. This is the
greatest advantage of this technique as compared to
other learning techniques which require extensive
input, detailed examination of intermediate results and
explanation of results are required [8]. CART makes
use of classification and regression trees for estimating
categorical predictor variables (classification) and
continuous dependent variables (regression). This
technique is useful in solving either regression type
problems or classification type problems. CART
analysis is basically designs a tree, which is different
from other conventional data analysis methods. It is a
process of designing a tree and finally selecting an
optimal tree that fits the required information.
F. Support Vector Machine (SVM)
It is a machine learning technique proposed
by Vladimir Vapnik and colleagues at AT&T Bell
laboratories in 1992. It is the state of art in machine
learning techniques. Support Vector Machines work
on decision planes or hyper planes that define decision
boundaries. A decision plane is a plane that separates
a set of objects having different class memberships.
SVM uses a set of supervised learning methods.
SVMs perform wonderful in bioinformatics in DNA
sequencing and modeling protein structure. SVM
learning is used for data analysis and patterns
identification that are useful for classification and
regression analysis. SVMs make use of a (nonlinear)
mapping function (Φ) that transforms input data space
to data in feature space in such a way as to make a
problem linearly separable. SVMs are used
tremendously as they stand on a good theoretical base
and are also applicable in real world applications.
There are two categories of software that provide
SVM training algorithms. The first category is
specialized software whose prime objective is to have
an SVM solver. LIBSVM [6] and SVMlight [6] are two
popular software of SVM for this category. The other
category of software is machine-learning libraries that
provide a variety of classification methods and other
facilities such as methods for feature selection,
preprocessing etc.
G. Rule Induction
It is widely used machine learning technique
and easy as rules are simple to frame and interpret
than a regression paradigm or a trained Artificial
915 | P a g e

Neural Network. It works on condition-action rules,
decision trees or same kind of knowledge structures.
In this technique the performance portion orders the
instances underneath the branches of a decision tree or
finds the first rule whose conditions match the
instance, making use of an all-or-none match process
[8]. This technique makes use of greedy search
method. Propositional learning and Relational learning
[8] are two ways of rule induction-.
H. Genetic Algorithms and Genetic Programming
These are a type of evolutionary computing
models that are used for problem solving and work on
the principle of biological evolution like natural
selection. Genetic algorithms have concepts of genes,
chromosomes and population. These models are based
on the Darwin theory “the survival of the fittest”
relying on the fitness function where the best solutions
are selected from a set of individuals. The individuals
that are found fit have good chance of their selection.
Once selection process concludes, new individuals
have to be formed. The process of crossover or
mutation forms the new individuals. In crossover
process, a combination of the genetic makeup of the
two solution candidates results in creation of new
individuals. In mutation process, some randomly
chosen portions of genetic information are varied to
obtain a new individual. This process of generation of
individuals does not complete till one of the
conditions such as minimum criteria is fulfilled or the
desired level of fitness is achieved or a specific
number of generations are attained or any of the
combinations above is reached [9].
I . Random Forest
Random forests models are used for
classification purpose. Random forests are a
combination of tree forecasters such that every tree
depends upon the values of an independently sampled
random vector where same distribution is used for all
trees in the forest. This technique consists of
generating a set of trees that vote for the most
prevalent class. There are two important
characteristics of using Random forests. The first
characteristic is that the generalization error converges
with the increase in number of trees increases and
second characteristic is that this type of learning does
not suffer from over fitting [10].

IV.

MACHINE LEARNING
TECHNIQUES FOR MOBILE
DEVICES: A STUDY

Machine learning techniques are useful for
mobile devices. XPod Mobile MP3 player was
developed by Sandor et al. [11]. XPod automated song
selection process as per user interest or liking. This
system successfully played contextually music as per
user’s choice using machine learning techniques.
SVMs suited the best for different tasks of mobile
music player as to begin with SVMs very less training
www.ijera.com

www.ijera.com

data. This is an interesting feature from the point of
view of user because a mobile user spends very less
time in system setting and more time in enjoying the
benefits. But designing SVMs on a constrained device
is a tedious task. SVMs can be designed easily on an
unconstrained device such as a Personal Computer or
a laptop and then be transferred to a portable device.
In this way a constrained device can evaluate an SVM.
Whereas decision trees are classifiers that can be
properly converted into a rule set that can be further
evaluated. The performance of Decision trees can be
improved with boosting strategy. Unsupervised
Machine learning techniques are useful in data mining
domain [12]. Machine learning techniques are also
helpful in network management and surveillance
domains [13]. Machine learning techniques like
reinforcement learning and instance-based learning
perform better in mobile devices. These techniques are
also useful for mobile music platform for different
categories of music expressions [15]. Soar cognitive
architecture is efficient to assist for mobile music
interactions. Mobile learning is a novel and useful
form of learning as it has all the strengths of elearning and bridges the gap of time and space in
classroom learning [16]. To enhance English learning
[16] presented a Personalized Intelligent Mobile
Learning System (PIMS). PIMS advised English news
articles or readings to the readers as per the reading
habits and abilities [16] of reader. PIMS used fuzzy
item response theory. PIMS was implemented
successfully on Personal Digital Assistant (PDA) to
furnish personalized mobile learning for improving
the reading ability of readers. PIMS also enriched
English vocabulary of learners. The advancement of
mobile technologies and efficient access to
multimedia resources are important contributions that
led to development of mobile learning systems [17]. A
mobile learning system using existing web resources
was designed that resulted in learning material as per
the requirements of user in [17]. Mobile agents are
agents that traverse in a network from one host to
other for successful completion of different tasks [18].
A mobile agent is an intelligent decision maker
concerning its itinerary and updates the decision(s) as
per the available information as it traverses from one
host to another [18]. Machine learning classification
techniques like Naïve Bayesian and ANN are useful
for document classification and efficient retrieval [18]
for mobile agents. Mobile devices offer a wide range
of services above several wireless network access
technologies [19]. Mobile devices are prune to threats
and machine learning techniques like K-Nearest
Neighbor, Bayesian Networks, Random Forests are
useful in intrusion detection for mobile devices [19].
Machine learning techniques are efficient for mobile
devices in several ways. Mobile learning applications
include language translation, game playing, intelligent
agents etc. An architecture for designing an Intelligent
System on mobile device was discussed by Jawad et
al. [21]. This design makes use of sensor networks.
916 | P a g e

The success of any mobile intelligent system depends
upon the performance of learning techniques
implemented on mobile devices. The performance
criteria of any machine learning algorithm depend
upon its Prediction accuracy, Sensitivity and
Specificity [14]. The Prediction accuracy is a measure
of accuracy of prediction or classification. The
sensitivity of an algorithm is a measure of how an
algorithm classifies correctly the positive instances
[14]. The Specificity is a measure of performance of
an algorithm in classifying the negative instances.
Performance can be improved by good prediction
accuracy and good sensitivity. Hence, the choice of
most appropriate algorithm plays a vital role designing
mobile learning applications.

[10]
[11]

[12]

[13]

V.

CONCLUSION

This paper presents a review of machine
learning techniques. It emphasizes that each machine
learning technique has some strengths and
weaknesses. It gives brief study of recent areas of
applications of machine learning techniques for
mobile devices. It also presents some performance
measures for a machine learning algorithms. It is
somewhat not an easy task directly to state that one
learning technique performs better than the other one.
However the performance of a machine learning
technique greatly depends on the application domain
and the requirement. Recent advances in Mobile
technologies are the areas of application of machinelearning techniques.

[14]

[15]

[16]

REFERENCES
[1]

[2]
[3]
[4]

[5]
[6]
[7]

[8]

[9]

Zanifa Omary, Fredrick Mtenzi, “Machine
Learning approach to identifying the dataset
threshold for the performance estimators in
supervised learning”, International Journal for
Infonomics (IJI), 3: 3, September 2010.
T.Mitchell, Machine Learning, MIT Press,
1997.
Alpaydin Ethem, Introduction to Machine
Learning, MIT Press, Second Edition, 2004.
L.Rokach and O.Maimon. Part C, “Top down
induction of decision trees classifiers – survey”,
Applications and Reviews, IEEE Transactions
on Systems, Man and Cybernetics,vol.35, pp
476-487, 2005.
T. Caelli, W.F. Bischof, Machine Learning and
Image Interpretation, Plenum Press, 1997.
Asa Ben-Hur, Jason Weston, A user’s guide to
Support Vector Machines.
Richard P. Lippmann, “An introduction with
computing with Neural Nets”, ASSP IEEE 4:2,
pp. 4-22, 1987.
Yogesh Singh, Pradeep Kumar Bhatia,
Omprakash Sangwan,“A Review of studies on
Machine Learning Techniques”, International
Journal of Computer Science and Security 1:1,
2007.
Stephen M Winkler, Michael Aenzeller and
Stefan Wagner, “Advances in applying Genetic

www.ijera.com

[17]

[18]

[19]

[20]
[21]

www.ijera.com

Programming to Machine Learning, Focusing
on Classification Problems”, Parallel and
Distributed Processing Symposium, IEEE,
IPDPS, 2006.
L. Breiman, Random Forests, Kluwer
Academic Publishers, Vol. 45, 2001.
Sandor Dornbush, Jesse English, Tim Oates,
Zary Segall, Anupam Joshi, “XPod: A Human
Activity Aware Learning Mobile Music
Player”, Proceedings of conference on
Advances in Ambient Intelligence, pp. 107-122,
IOS Press Amsterdam, The Netherlands, 2007.
Areej Shhab,Gongde Guo, Daniel Neagu, “A
study on applications of Machine Learning
Techniques in Data Mining”, 22nd BNCOD
Workshop, 2005.
Nigel Williams,Sebastian Zander,Grenville
Armitage, “A preliminary Performance
comparison of Five Machine Learning
Algorithms for practical IP traffic flow
classification”, ACM SIGCOMM Computer
Communication Review, 36:5, 2006.
S. Kotsiantis, C. Pierrakeas, P. Pintelas,
“Predicting student’s performance in distance
learning using machine learning techniques”,
Artificial Intelligence, Taylor and Francis
Group, 2004.
Nate Derbinsky,Georg Essl, “Cognitive
Architecture in Mobile Music Interactions”,
NIME’11, Oslo, Norway May-June 2011.
Chih-Ming
Chen,
Shih-Hsun
Hsu,
“Personalized Intelligent Mobile Learning
System for Supporting Effective English
Learning”, Educational Technology and
Society, 11:3, pp. 153-180, 2008.
Mohammed
Alzaabi,
Jawad
Berri,
Mohammed Jamal Zemerly, “Web-based
Architecture
for
Mobile
learning”,
International Journal for Infonomics(IJI) 3:1,
March 2010.
Jihoon Yang, Prashant Pai, Vasant Honavar,
Les Miller, “Mobile Intelligent Agents for
Document Classification and Retrieval : A
Machine Learning Approach”, Citeseerx , 2011.
Dimitrios Damopoulos, Sofia A. Menesidou,
Georgios Kambourakis, Maria Papadaki,
Nathan Clarke, Stefanos Gritzalis, “Evaluation
of Anomaly-Based IDS for Mobile Devices
Using Machine Learning Classifiers”, Security
Comm. Networks, 2011.
Rajkamal, Mobile Computing, Oxford University
Press, 2007.
Jawad Oubaha, Ahmed Habbani and
Mohammed Elkoutbi, “Mobile Intellient
System (MIS) and a multi-criteria in MPLS
networks”, Internationla Journla of Next
Generation Networks (IJNGN), 2:4, pp. 61-71,
2010.

917 | P a g e

Ew36913917

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Ew36913917 (20)

Recently uploaded (20)

Ew36913917