MOBILE MIND: A FULLY MOBILE PLATFORM BASED MACHINE LEARNING APPLICATION

Abstract

In recent years, mobile devices have developed significantly in terms of technical capabilities, computing power, storage capacity and ability of sensing different activities via intelligent built-in sensors. In this perspective, capabilities of ultimate mobile phone technology has begun to be a candidate novel platform for machine learning and data mining activities by the help of its computing power. In this study, a fully mobile platform based machine learning application named Mobile Mind is designed and implemented. While, all other current mobile platform based machine learning and data mining applications are using central data mining servers to perform analysis, Mobile Mind does all tasks on cell phone's processor and memory. On the other hand, Mobile Mind currently supports support vector regression and kernel recursive least squares regression algorithms with polynomial and radial basis kernels to allow users performing predictive data mining operations on flat CSV (comma separated values) files. By this study, it is shown that mobile platforms are becoming native and ubiquitous platforms for machine learning purposes from now on. Therefore, the need of central data mining servers and web service usage for data transferring will started to be less and less in the future. Furthermore, a native fully mobile machine learning tool presents unlimited opportunities to the mobile application programmers especially dealing with sensor data driven applications has much potential in this point of view.

Full text

MOBILE MIND: A FULLY MOBILE PLATFORM BASED
MACHINE LEARNING APPLICATION
Ahmet Selman BOZKIR
Hacettepe University Computer Engineering Department, Ankara, Turkey
selman@cs.hacettepe.edu.tr
Ebru AKCAPINAR SEZER
Hacettepe University Computer Engineering Department, Ankara, Turkey
ebru@hacettepe.edu.tr
ABSTRACT
In recent years, mobile devices have developed significantly in terms of technical capabilities, computing power, storage
capacity and ability of sensing different activities via intelligent built-in sensors. In this perspective, capabilities of
ultimate mobile phone technology has begun to be a candidate novel platform for machine learning and data mining
activities by the help of its computing power. In this study, a fully mobile platform based machine learning application
named Mobile Mind is designed and implemented. While, all other current mobile platform based machine learning and
data mining applications are using central data mining servers to perform analysis, Mobile Mind does all tasks on cell
phone’s processor and memory. On the other hand, Mobile Mind currently supports support vector regression and kernel
recursive least squares regression algorithms with polynomial and radial basis kernels to allow users performing
predictive data mining operations on flat CSV (comma separated values) files. By this study, it is shown that mobile
platforms are becoming native and ubiquitous platforms for machine learning purposes from now on. Therefore, the need
of central data mining servers and web service usage for data transferring will started to be less and less in the future.
Furthermore, a native fully mobile machine learning tool presents unlimited opportunities to the mobile application
programmers especially dealing with sensor data driven applications has much potential in this point of view.
KEYWORDS
Mobile data mining, Machine learning, Support vector regression, Kernel recursive least squares
1. INTRODUCTION
Mobile data mining has started to be one the important fields of pervasive and ubiquitous mobile computing.
In their study, Talia and Trunfio described the three possible usage scenarios of mobile devices in data
mining oriented applications such as (1) terminal mode which users access to a central data mining server,
select data source and invoke operations on data along with a returning result; (2) data generator mode that
users or mobile devices generate data and send it to a remote server for further processing and (3) miner
mode that all data mining related operations are done on device itself (Talia & Trunfio, 2010). However, it’s
reported that third mode is unrealistic because of limited capabilities of mobile devices such as computing
power and storage area (Talia & Trunfio, 2010). In essence, if literature is reviewed, it can be seen that, many
of the mobile data mining studies rely on utilizing first two modes. In one study, authors designed and
implemented a distributed mobile data mining environment conforming to first usage scenario depicted
above to smartly monitor and visualize stock market data on PDAs by utilizing Fourier transform of decision
trees and employing Java technologies (Kargupta et al., 2002). In another study, a mobile data mining
application based on communication with a remote data mining server via web service techniques is
introduced (Talia & Trunfio, 2010). In that study, users allowed to select a data table on server and perform
data mining methods by invoking remote built-in functions of Weka Toolkit (Weka, 2010). On the other
hand, due to shortcomings of first and second modes in mobile data mining scenarios, an efficient mobile
data mining model is developed which helps to preprocess of local mobile data and gain descriptive statistics

before sending it to remote mining servers (Goh & Taniar, 2005). In another study, an intelligent system
named VEDAS which tracks and monitors of vehicles by using on-board PDAs connected through wireless
networks and performs data mining analysis over recorded paths is introduced (Kargupta et al., 2003).
As can be seen, mobile data mining centric applications have generally avoided using the mobile device
itself for performing whole data mining task. Instead, they have used central data mining servers for analyzes
and utilize web service or similar technologies for data transferring and gathering responses. This approach
has the advantage of high speed and scalable data analysis but suffers from well-known low bandwidth
problem of communication lines. However, as a result of innovations in mobile device industry, current
high-end mobile devices have been equipped with adequate size of multi-touch screens, built-in magnetic
field, proximity and accelerometer sensors, high resolution cameras, powerful processors up to 1 GHz, large
volumes of storage capacities up to 32 GB which eliminate the limitations over scenario three and allows
mobile devices to be a powerful candidate platform for data mining & machine learning activities.
Furthermore, due to rapid development of mobile devices and built-in sensors, the way of mobile application
development has evolved and mobile devices have became self data generators in this case. Therefore, local
analysis of sensor generated data has much potential and valuable. One important instance of this case is
motion based recognition. Liu et al., developed a single passes gesture recognition algorithm for three-axis
accelerometers which allows capturing and revealing patterns of various gestures (Liu et al., 2009). As
today’s mobile devices have these kinds of accelerometers, it is possible to recognize motions of users and
develop authentication mechanisms for mobile devices. However, to achieve this and similar goals, existence
of mobile classifier or regressor is very essential.
In this paper, we introduce a completely mobile based machine learning tool named Mobile Mind. To
allow users performing predictions in test data through employing pre-installed algorithms over train data
constitutes the main purpose of the present study. Mobile Mind currently supports support vector regression
and kernel recursive least squares (an online algorithm) algorithms with polynomial and radial basis kernel
choices. The algorithms used in Mobile Mind are migrated from Dlib Machine Learning Toolkit (King,
2009).
2. MOBILE MIND WITH ALL ASPECTS
As described in previous section, many data mining and machine learning related applications on mobile
platforms are designed and implemented in client-server architecture considering the limited capabilities of
mobile devices, such as processor power, low memory and battery life. However, as the aim of this study is
to run mining algorithms on device itself, current high-end cell phones and platforms attributes are
investigated at startup. Google’s Android (Android, 2011), Apple’s IOS 4 (IOS 4, 2011) and Samsung’s
Bada (Bada, 2011) were the choices stand out with their pervasiveness, robustness and easy programmability
features. However, as our core mining library Dlib Machine Learning Toolkit was written in standard C++
language and the platform which has C++ support was only the Bada platform, Samsung’s Bada platform has
been chosen. The reason behind the selection of Dlib library rather than others is being an open-source
project, having many kernel based methods including support vector machine, relevance vector machine and
kernel recursive least squares, being written in standard C++ language for easy migration between different
platforms and regular updates for new features and bug fixing. On the other hand, Dlib library presents many
other utility libraries as well as image processing, networking and optimization that can be used in this study
in the future for further aims.
Support vector regression (SVR) and kernel recursive least squares (KRLS) algorithms are selected and
adopted in Mobile Mind due to some reasons. First of all, SVR is a state-of-algorithm which proved its
reliability and pervasiveness by being used in many study fields such as medicine and finance (Farquad et al.,
2010) On the other hand, KRLS is an on-line learning algorithm developed by Engel, Mannor and Meir at
2003 which has advantage of avoiding re-training of whole training cases from startup for new upcoming
cases. On the other hand, KRLS is highly suitable algorithm for mining of streaming data and time series
analysis (Engel, Mannor & Meir, 2003). Therefore it is selected to be a part of Mobile Mind for streaming
data analysis and sensor activity mining purposes that are high probable in cell phone and sensor based future
applications. Readers can found deeper information about Bada platform, algorithms used in this study and
details of Mobile Mind in the sections below.

2.1 Bada smartphone platform
Bada platform is an operating system named BadaOS designed and implemented for cell phones developed
by Samsung. BadaOS is currently used various cell phones of Samsung including Wave 8500, Wave II 8530
and Samsung S7233. Samsung provides software development kits (SDK) regularly and version of ultimate
SDK is 1.2.1 which was released as of December 17 of 2010. SDK presents GUI, I/O, security, networking,
media, location based services and social networking libraries & namespaces for developers together with
easy-to-use integrated development environment (IDE) which is shipped with SDK. Although, Bada is a very
new platform, due to its robust structure and C++ basis, it has found worldwide usage throughout the mobile
application programmers. However, it has some limitations due to its immaturity such as lack of full support
for standard C++ libraries and multitasking issues that avoids an application to start another application
without blocking running status of its own. To overcome these problems Samsung is currently developing
BadaOS 2.0 platform which will have support for complete standard C++ and fully multitasking along with
new unique improvements such as built-in text to speech (TTS) and speech recognition.
2.2 Algorithms used in Mobile Mind
Mobile Mind is a developing tool and currently supports only SVR and KRLS algorithms which are chosen
for different purposes. In next versions new other algorithms which have already been implemented in Dlib
library will be adopted in Mobile Mind.
SVR which was developed by Vapnik at 1995 is based on statistical learning theory and aims to create a
hyperplane which lies near to or close to as many instances as possible (Farquad et al., 2010; Ancona, 1999).
On the other hand, as SVR is a kernel-based method, it does not have some shortcomings which some other
methods have such as decision trees in terms of over-fitting and competition of input variables on node
splitting. However, it needs more time than the other non-kernel based algorithms for computing support
vectors in case of increment in number of training examples. By considering the pros and cons of SVR, it is
decided to be included in Mobile Mind. SVR implementation in Dlib ML toolkit is epsilon insensitive
version of SVR that requires three parameters such as kernel type, C (regularization parameter) and epsilon-
insensitive parameter that sets the accuracy level of regression.
KRLS on the other hand, is another kernel based method which is an improved version of recursive least
squares algorithm, that is popular and practical in many fields covering signal processing, communications
and control kernel (Engel, Mannor & Meir, 2003). KRLS performs linear regression in the feature space
induced by Mercer kernel. For this reason, KRLS algorithm can be used to build a minimum mean squared
error regressor recursively. On the other hand, KRLS is an online algorithm which means that algorithm does
not need to be re-trained from scratch when new train cases are available. According Engel et al., “on-line
algorithms are useful in learning scenarios where input samples are observed sequentially, on at a time (e.g.
data mining, time series prediction, reinforcement learning)” (Engel, Mannor & Meir, 2003). Moreover, on-
line algorithms are useful and practical in real-time decision giving operations. For these reasons, KRLS is
included in Mobile Mind algorithms library. KRLS implementation in Dlib ML toolkit requires three
parameters such as kernel type, accuracy level and dictionary size determining how many of the dictionary
vectors will be memorized during training phase.
As stated before, Mobile Mind uses kernel based methods. Therefore currently polynomial and radial
basis (Gaussian) kernels are presented to users. As a result of this, different numbers of kernel parameters for
each of these kernel types are presented in GUI such as gamma value (for both of them), coefficient and
degree values (for only polynomial kernel)
2.3 Design, implementation and usage of Mobile Mind
Mobile Mind is written in C++ on Bada SDK 1.2.1 and includes Dlib machine learning toolkit library as a
background algorithm engine. Due to be compact and shrink the size of program, only machine learning and
essential components of Dlib are integrated at compiling stage. As a visual design requirement, mostly used
functions are located on main form as button elements instead of menu items. Thus, a perspicuous and easy
to use environment is tried to be obtained. As depicted in Fig. 1, usage of Mobile Miner consists of four
essential and one optional stage. First of all, users must pick training and testing data files which must be

conformed to CSV file standards. Predictions can be performed with existence of these two files. On the
other hand, to measure the generalization capability of prediction, R2 (determination of coefficient) measure
is selected. To obtain R2 value, user must provide also one another file named as “observed values” file
which contains only the observed values as a column matrix. After picking data files, users either perform
training by pushing on selected algorithm button or can continue to configure operation’s kernel type or
parameters as depicted in Fig. 2.b. After training stage, algorithms predict the cases given in test files and
outputs R2 result with output file (if filename is defined) as depicted in Fig. 2.d.
Figure 1. System chart of Mobile Mind
As an obligation, the output variable must be the in last order in training file and the test file should not
contain observed values (output variable) as reside in training file. Instead, observed values must be stored in
another file named “observed values file”. During the development stage, model saving and re-loading
capabilities are tried to be implemented with the co-operation of Dlib. However, as current Bada platform
does not have full support of standard C++ and <iostream> classes, model saving and re-loading methods
could not be implemented. Likewise, these features require of serialization and storing of created models and
need the <iostream> and some other standard C++ functions.
To measure the performance of the methods and check whether or not methods are suitable for mobile
machine learning purposes, Mobile Mind is benchmarked with three datasets by employing each algorithm
with two types of kernels. All these datasets have only continuous variables as well as output variable and
datasets are divided into train and test groups with 80% - 20% partitions. First dataset includes data gathered
via a tunnel boring machine and the aim of the data is to predict “field rate penetration” variable. This data is
also subjected in a study (Akcapinar Sezer et al., 2010). The remaining datasets are downloaded from UCI
Machine Learning Repository located at http://archive.ics.uci.edu/ml/datasets/. Subject of the second dataset
is predicting concrete compressive strength by using age and other ingredient variables. As a last and larger
one, third dataset contains “parkinsons telemonitoring” data in terms of voice recording measurements of 42
patients, aiming to predict “total UPDRS” variable by using other input variables. The properties of datasets
are given in Table 1 below.

Figure 2. Screenshot of Mobile Mind: [a] Main form, [b] Parameter settings form, [c] Main menu and [d] Result form
In benchmark stage, gamma, accuracy, dictionary size (specific to KRLS method), C (regularization
parameter specific to support vector machine methodology), coefficient and degree parameters (used by only
polynomial kernel) are kept as 0.0005, 0.0001, 10000, 10, 2 and 2 respectively. All tests are executed at
simulator. As the goal of the benchmark is to measure the capabilities of algorithms on mobile device and
determine the concordance of algorithms for mobile machine learning purposes, count of seconds elapsed
during operation and maximum memory usage records are determined. In time measurements, only training
and testing phases are considered while R2 calculation and file outputting are excluded .On the other hand,
600 seconds (10 minutes) is accepted as a maximum threshold value of operation. Thus, operations
exceeding 600 seconds are cut.
Table 1. Different datasets with features and corresponding computational cost in time
Algorithm - Kernel Dataset # of Input
Variables
# of Training
Cases
# of Test
Cases
# of Seconds
To Finish Job
Maximum Memory
Usage (MB)
SVR – Polynomial TBM 4 119 32 8 79
SVR – Radial Basis TBM 4 119 32 1,5 81
KRLS – Polynomial TBM 4 119 32 3 82
KRLS – Radial Basis
TBM 4 119 32 3.5 82
SVR – Polynomial Concrete 8 822 208 >600 87
SVR – Radial Basis Concrete 8 822 208 9 86
KRLS – Polynomial Concrete 8 822 208 19 79
KRLS – Radial Basis
Concrete 8 822 208 103 98
SVR – Polynomial Parkinsons
18 4703 1172 >600 275
SVR – Radial Basis Parkinsons
18 4703 1172 235 260
KRLS – Polynomial
KRLS – Radial Basis
Parkinsons
Parkinsons
18
18
4703
4703
1172
1172
98
131
81
79

3. DISCUSSION
If the results given on Table 1 are investigated, it can be seen that KRLS algorithm performs fast and scalable
analyses rather than SVR. As it uses a dictionary vector and it only remembers last important vectors (the
memorization level can be set by dictionary size) required memory size remains constant and therefore it is
not affected by the numbers of cases in training data. On the other hand, considering time needs, KRLS
performs faster mining with polynomial kernel than radial basis kernel. From this point of view, analysis with
polynomial kernel stands more plausible in KRLS usage. However, as KRLS uses a dictionary size, it
increments required memory space linearly; KRLS is suitable for both kernel types. Likewise it is
indispensible to use radial basis kernel for non-linear analysis.
In SVR, memory usage increments exponentially when number of cases and especially the number of
variables increases. Performing SVR with radial basis kernel requires acceptable amount time to perform.
However, when polynomial kernel is to be discussed, due to the nature and structure of algorithm, SVR
requires excessive amounts of time which is unacceptable. Therefore, it can be stated that SVR with
polynomial kernel is useless for current mobile cell-phone features. The main reason of this finding is SVR
generates exponential growing number of support vectors by the increment of training data size. Furthermore,
working with polynomial kernel on SVR requires degree value to bet as an integer and gamma value is small
enough so that the output of the kernel is not a huge number. However, SVR & polynomial kernel couple
runs slower than KRLS.
In the tests, the biggest training file contains 4703 cases. If the fact that ultimate mobile devices which
BadaOS runs on, have at most 256 MB usable memory, it can be concluded that KRLS is a very suitable
algorithm for mobile machine learning as it is not affected from the number of training cases. In contrast,
SVR is a reasonable algorithm unless polynomial kernel is used. However, as it is not designed as an online
learner, it can be clearly stated that SVR has limitations in mobile device platforms due to memory issues.
4. CONCLUSION AND FUTURE WORK
In this study, a fully mobile platform based machine learning tool named Mobile Mind was developed. It is
designed and implemented to let the users to perform whole machine learning processes on mobile device
instead of using remote servers. As can be seen in the study, ultimate mobile device technology has become
capable of doing such kind of tasks. Likewise, currently used processors in mobile devices can reach up
1GHz. If benchmark tests are considered together with this improvement, processor capability is not barrier
for many cases from now on. However, as the machine learning algorithms require memory, available
memory area could still be a bottleneck in mobile machine learning. Due to fact that, mobile devices do not
have virtual memory features as existed in PCs, algorithms are being limited with the amount of memory at
device. On the other hand, with this study, it is shown that mobile devices have enough power to be a mobile
classifier or regressor for artificial intelligence, machine learning and data mining centric applications.
As a further research, adapting new classifier algorithms which are independent from the count of training
cases is planned. Additionally, model load and save utilities are planning to be implemented in Mobile Mind
whenever standard C++ support is available in Bada.
REFERENCES
Akcapinar Sezer, E., Bozkir, A.S., Yagiz, S, Gokceoglu, S.,2010, Karar Ağacı Derinliğinin CART Algoritmasında
Kestirim Kapasitesine Etkisi: Bir Tünel Açma Makinesinin İlerleme Hızı Üzerinde Uygulama, Akıllı Sistemlerde
Yenilikler ve Uygulamaları Sempozyumu, Kayseri, Turkey.
Ancona, N., 1999, Classification properties of support vector machines for regression. Technical report, R.I.-IESCI/CNR-
Nr., 02/99
Android, 2011, Available: http://www.android.com/
Bada, 2011, Available: http://www.bada.com/
Goh, J. and Taniar, D., 2005, An Efficient Mobile Data Mining Model, Springer-Verlag LNCS, pp 54-58.
IOS 4, 2011, Available : http://www.apple.com/iphone/ios4/

Engel, Y., Manor, S., Meir, R., 2003. The Kernel Recursive Least Squares Algorithm, IEEE Transactions on Signal
Processing, Vol. 52, pp 2275-2285.
Farkuad, M.A.H., Ravi, V., Bapi Raju, S., 2010. Support vector regression based hybrid rule extraction methods for
forecasting, Expert Systems with Applications, Vol. 37, No. 8, pp 5577-5589.
Kargupta, H. et al, 2002. Mobimine: monitoring the stock market from a PDA, ACM SIGKDD Explorations. Vol. 3, No.
2, pp 37-46.
Kargupta, H. et al, 2003. VEDAS: A Mobile and Distributed Data Stream Mining System for Real-Time Vehicle
Monitoring. Proc. SIAM Data Mining Conference.
King, D., 2009, Dlib-ml: A Machine Learning Toolkit. Journal of Machine Learning Research, Vol. 10, pp 1755-1758.
Liu, J. Et al, 2009, uWave: Accelerometer-based personalized gesture recognition and its applications , Pervasive and
Mobile Computing, Vol. 5, No. 6, pp 657-675.
Talia, D. and Trunfio, P., 2010. Mobile Data Mining on Small Devices Through Web Services, Wiley, USA.
Weka, 2010, Available: http://www.cs.waikato.ac.nz/ml/weka/