Escolar Documentos
Profissional Documentos
Cultura Documentos
net/publication/313734208
CITATIONS READS
19 178
6 authors, including:
Some of the authors of this publication are also working on these related projects:
High Performance Computing and Big Data: Technologies and Applications View project
All content following this page was uploaded by Iyad Katib on 13 April 2017.
Rashid Mehmood, Senior Member, IEEE, Furqan Alam, Nasser N. Albogami, Iyad Katib, Member,
IEEE, Aiiad Albeshri and Saleh M. Altowaijri
Abstract—The education industry around the globe is undergoing major transformations. Organisations such as Coursera are
advancing new business models for education. A number of major industries have dropped degrees from the job requirements.
While the economics of higher education institutions are under threat in a continuing gloomy global economy, digital and lifelong
learners are increasingly demanding new teaching and learning paradigms from educational institutions. There is an urgent
need to transform teaching and learning landscape in order to drive global economic growth.
The use of distance eTeaching and eLearning (DTL) is on the rise among digital natives alongside our evolution towards smart
societies. However, the DTL systems today lack the necessary sophistication due to several challenges including data analysis
and management, learner-system interactivity, system cognition, resource planning, agility, and scalability. This paper proposes
a personalised Ubiquitous eTeaching & eLearning (UTiLearn) Framework that leverages IoT, big data, supercomputing, and
deep learning to provide enhanced development, management and delivery of teaching and learning in smart society settings. A
proof of concept UTiLearn system has been developed based on the framework. A detailed design, implementation and
evaluation of the UTiLearn system, including its five components, is provided using eleven widely used datasets.
Index Terms— Big data, Computational and artificial intelligence, Distance learning, High performance computing, Internet of
Things
—————————— ——————————
1 INTRODUCTION
course (MOOC ) organisations such as Coursera, edX and
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only.Published
xxxx-xxxx/0x/$xx.00 © 200x IEEE
Personal by
usethe
is also permitted, but republication/redistribution requires IEEE permission. See
IEEE Computer Society
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
Modern day learners demand increasingly sophisti- The notion of smart cities can be extended to smart so-
cated, dynamic, anytime anywhere learning environ- ciety; i.e. a digitally-enabled, knowledge-based society,
ments. While the use of eLearning and Distance learning is aware of and working towards social, environmental and
on the rise among the digital natives and millennials, the economic sustainability.
DTL1 systems and environments of contemporary times We note that knowledge and human and social capital
lack the necessary sophistication. Physical learning spaces are at the heart of smart city and smart society develop-
enable interactive environments with a number of im- ments. Therefore, we believe that the role of distance learn-
portant features that are difficult to mimic in virtual and ing would extend beyond the mainstream “education for
distance learning environments. The distance learning en- employment” scope. It will extend to the notion of social
vironments of today are unable to connect teacher to the and collaborative governance where the society will col-
learner at the same level and quality that physical spaces laborate to train each other in maintaining its knowledge,
could provide. For example, it is difficult to monitor and moral fiber, good practice, resilience, competitiveness, and
manage virtual and distance learning environments due to for bringing innovation, and becoming a knowledge-based
the absence of physical presence of the learner and teacher economy. The key to such efforts would be a converged
at the same place. It may be possible to create realistic in- ubiquitous infrastructure comprising many systems that
teractions between the teacher, learner and the learning en- share information among them and collaborate with each
vironments but these require a systematic approach for other as necessary.
collecting and harnessing data related to these learning en- Motivated by the disruptions and emerging require-
vironments and their users. Physical spaces and their in- ments of the education industry, and considering the chal-
teractive environments are organized and conducted by lenges discussed above in this area, we propose in this pa-
humans; it is difficult to mimic human intelligence in vir- per a personalised UTiLearn (Ubiquitous eTeaching &
tual and distance learning environments. There is a need eLearning) Framework. UTiLearn is DTL framework de-
to embed cognitive abilities in the DTL systems in order to signed for enhanced development, management and de-
serve the needs of the next-generation learners. Also, a livery of teaching and learning in smart city and smart so-
DTL system would typically provide learning content via ciety settings.
videos and virtual reality media, in addition to text-based The UTiLearn framework leverages on emerging tech-
and voice-based content. This would require IP and other nologies to provide an environment and ecosystem for so-
networks based communication. Therefore, planning of re- cieties to enrich their knowledge, anytime, anywhere, any
sources and scalability of a distance learning system is an- age, through improved processes for collection, develop-
other major challenge given the large-scale data, computa- ment, accessibility, pedagogy and cognition of knowledge.
tional and networking requirements discussed above. The Its users could be teachers, students, governments, busi-
agility of the infrastructure is also important. Moreover, nesses, third parties, and virtual organisations or groups
the system should allow accessibility from a broad range formed for specific purposes. A learner may register for a
of devices and networks. course or programme; or she may register her goals, learn-
Another development that is relevant to the topic of ing levels, and timelines with the system and the system
this paper is around the concept of smart cities because it suggests the learner the available courses or programmes,
relates to the environment of the learner in which they live, based on her personal preferences and monitors her
learn, work, and socialise. Smart cities provide the state of through to the completion of the learning task. In addition
the art approaches for urbanisation, having evolved from to the usual DTL services that MOOCs typically offer, a
the developments carried out under the umbrella of certain group or a government department can develop
knowledge-based economy, and subsequently under the specific material to reach out to a group of people to create
notion of digital economy and intelligent economy [4] . awareness of a topic for social intervention purposes, such
Smart cities exploit physical and digital infrastructure, as as to change people behavior towards better driving prac-
well as the intellectual and social capital, for urban and so- tices or healthy lifestyle.
cial development. A city can be defined as ‘‘smart’’ when The UTiLearn framework uses IoT as part of the digital
“investments in human and social capital and traditional infrastructure needed to connect with the user for person-
(transport) and modern (ICT) communication infrastruc- alisation and optimisation of teaching and learning. Mo-
ture fuel sustainable economic growth and a high quality bile computing technologies are used to connect to the us-
of life, with a wise management of natural resources, ers, sense their continuous activities, and provide them an-
through participatory governance” [5]. Komninos [6] de- ytime anywhere opportunity for DTL. Big data technolo-
fines smart cities as, “territories with high capacity for gies are used to deal with the challenges related to the vol-
learning and innovation, which is built-in the creativity of ume, velocity, and other characteristics of the DTL data.
their population, their institutions of knowledge creation, Computational Intelligence including deep learning is
and their digital infrastructure for communication and used for analysis, prediction, personalisation, optimisation
knowledge management”. Technically, smart cities are and decision making. HPC technologies are needed to deal
complex systems of systems that rely on converged and with the high computational requirements of AI and other
ubiquitous infrastructures. algorithms. Cloud technologies are used to deal with the
agility, scalability, and ubiquity of the framework. The
1 In this article, we adopt the term “Distance eTeaching and eLearning” Teaching & Learning (T&L) through ICT-based distance education.
with the abbreviation “DTL” to refer to the broader concepts incorporating
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 3
UTiLearn Framework could be implemented on top of a comprising 11,904 cores in 496 nodes. Aziz was ranked the
hybrid cloud IaaS (infrastructure as a service) and can offer world’s top 491 supercomputer in November 2015 Top500
ubiquitous teaching and learning as a Service (UTLaaS). competition (http://www.top500.org/).
Smart city technologies could be used to benefit from con- Section 2.2.1 gives a background on the technologies
verged social and national infrastructure in order to pro- used in this research and reviews the relevant literature.
vide integrated, enhanced and optimized pedagogies, con- Sections 3 to 8 introduce the UTiLearn framework and its
tent delivery, and better cognition for the users. proof of concept, including the function, the methodology,
Moreover, based on the proposed framework, we de- implementation and evaluation of the UTiLearn system.
velop a system, called the UTiLearn system. A prototype Section 9 concludes this paper and gives future directions.
implementation of the UTiLearn system comprising five
components is described and evaluated in this paper. The
UTiLearn framework and its algorithmic refinement com- 2 Background Material and Literature Review
prising five components is discussed in Section 3, and the We now give a brief background of the emerging technol-
subsequent five sections detail the function, the methodol- ogies used in UTiLearn followed by a detailed literature
ogy, implementation and evaluation of the five compo- review in Section 2.2.
nents. The purpose of the first component, the Sentiment
Analyser, discussed in Section 4, is to analyse social media 2.1 Technology Background
for forming the business case of the decisions taken by the Internet of Things (IoT): The Internet of Things (IoT) refers
system on strategic, tactical, operational and real time lev- to “a global infrastructure for the information society, ena-
els. The User Activity Recognition & User Identification bling advanced services by interconnecting (physical and
(UARI) Component, detailed in Section , analyses user ac- virtual) things based on existing and evolving interopera-
tivity data acquired from IoT devices to understand and ble information and communication technologies” [8]. IoT
predict spatio-temporal activities and other patterns of the would make our environments smarter by providing sens-
user. The Adaptive Content Delivery Mode Adviser (AC- ing and control capabilities at a fine-grain level. The data
DMA), in Section 6, leverages the user activity and identi- generated by IoT would grow exponentially as the number
fication related information produced by the UARI com- of “things” is expected to reach the 50 billion mark by 2020
ponent to optimize the DTL material delivery, and the sys- [9]. The IoT data would allow us to comprehend and man-
tem and network performance. The System and Network age complex environments, make better decisions, increase
Provisioning & Demand Prediction (SNPDP) Component automation, and enable higher productivity and efficien-
uses the information produced by ACDMA to compute, cies [10]. We will see in this paper how IoT could enhance
predict and manage the system and network resources re- DTL theory. and practice.
quired by the UTiLearn system (see Section 7). The Perfor- Big Data: Big data technologies refer to “the emerging
mance Analysis, Prediction and Intervention (P2AI) Com- technologies that are designed to extract value from data
ponent informs UTiLearn users about their performance having four Vs characteristics; volume, variety, velocity
and intervenes as necessary. All five components of the and veracity” [11]. Big data is revolutionizing most aspects
system are designed to optimise personalisation for indi- of our lives, be it work or social interactions. IoT would be
viduals as well as performance for the UTiLearn he system. a major contributor of big data. A major challenge is the
We have carried out a detailed literature review of dis- management of big data and timely analyses of big data in
tance and eLearning systems and reported the challenges order to extract reliable insights and other “value” from
that future DTL systems face today. We used this infor- big data. Big data would play an important role in next
mation to develop the UTiLearn framework and a proof of generation DTL technologies.
concept system based on the emerging technologies. To the AI and Deep Learning: The use of computational intel-
best of our knowledge, we have not found any DTL frame- ligence, including artificial intelligence (AI), has seen tre-
work or system that addresses the breadth of issues as ad- mendous growth in the last decade, and this would con-
dressed by UTiLearn. tinue to grow at a much faster pace in the coming years
Data, Algorithms, Software, and Hardware: We (see e.g. [12]). This is mainly due to the increasing capaci-
have a used a total of 11 publically available datasets in- ties, and decreasing costs, of computing infrastructure. We
cluding Twitter data, news media data, seven activity have mainly focused on deep learning, a subfield of AI, in
recognition datasets, telecommunication networks and ap- this paper. Deep learning technologies, also known as deep
plications traffic data, and data from edX [7] (a collabora- structured learning, have made remarkable progress re-
tive effort of Harvard and MIT). We have used the R sta- cently and its use has grown in diverse areas including
tistical platform for modelling, analysis and decision mak- computer vision, bioinformatics, speech recognition, and
ing of all the data and processes in our UTiLearn system. natural language processing. In certain application areas,
The data analysis, data quality management, prediction deep learning has produced significantly better results
and decision making methods used in the prototype im- than other AI methods including machine learning [10],
plementation of the UTiLearn framework include Naïve [13], [14].
Bayes classifier, Random Forest, and deep learning artifi- High Performance Computing (HPC): High Perfor-
cial neural networks (DLANNs). The experiments were mance Computing (HPC) “most generally refers to the
carried out on a supercomputer, called Aziz, a Fujitsu practice of aggregating computing power in a way that de-
made machine with peak performance of 230 teraflops, livers much higher performance than one could get out of
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
a typical desktop computer or workstation in order to play critical role in assuring the effectiveness of the eLearn-
solve large problems in science, engineering, or business” ing systems [18]. Cognitive factors describe learner deci-
[15]. The other terms similar to HPC are supercomputing sion making ability that regulate processes in controlled
and parallel computing. The term distributed computing learning environment [19]–[21]. Motivation-related factors
typically has a different focus but in future it is expected to include personal goal orientation, personal attribution of
converge with HPC. The use of HPC initially was restricted success, and personal attitude toward performing a learn-
to some specialized applications such as in engineering, oil ing task [22], [23]. While cognitive and motivation-related
and gas due to the high cost of HPC machines. However, factors have received increasing attention in recent re-
HPC is now being applied to many diverse areas including search, few studies have directly examined the environ-
social media mining. The use of HPC in the education in- mental context’s influence on learning effectiveness. Envi-
dustry is also expected to grow in the near future. ronmental factors include the structure and content of a
Mobile computing has already been widely used in course, as well as learning conditions that may moderate
DTL. It allows close interactions with the user, collection of the effectiveness of learner-controlled instruction [24], [25].
leaner related preferences and other data, personalization How to trap all these factors is challenging aspect of
and improved delivery of learning. Cloud computing al- eLearning applications and allot depends on these factors
lows system agility. The AI technologies and big data that how learner perceive the systems.
would come together to extract novel insights and make The eLearning market is showing great potential
better decisions from data collected through IoT and other around the world. It is growing at a slow but steady rate
means in smart environment settings, including the DTL [26], [27] around the world. The growing acceptance rate
environments. Big data would affect the future network of eLearning solution both by institutions and students is
traffic and hence the network architectures and this also also attracting more research and developments by indus-
needs to be taken into account while designing a DTL sys- try.
tem. The use of the technologies (discuss above) in the
UTiLearn system would become clearer as we explain our
framework and system.
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 5
DTL developments based on technology, pedagogy, the- such as individuals, businesses and governments. Now
ory, policy, accessibility, design, environment and man- businesses are using SM such as Twitter, Facebook, Insta-
agement domains. This work illustrated the recent and gram and blogging to advertise their product and services,
past DTL trends. to give special offers as per user preferences. Last year in
American elections we have already seen how SM is used
2.2.2 Distance eLearning Frameworks by people to form an opinion about particular candidate
Motivation plays an important in the learning process. [38]. A new dimension of SM is now shaping up to form a
Positive motivation enhances better understanding of top- completely new segment in its services i.e. distance
ics among the students. MENTOR is one of the such multi- eLearning. In the work [39], found out that younger gen-
media based adaptive online DTL framework [31]. It re- eration is too enthusiastic for using SM for distance
tains positive emotions in students by proving user eLearning. The factors of this increasing enthusiasm is
friendly, personalised environment for learning which based on SM eLearning perception, convenience, academic
based student’s emotion, sentiment and personality. An- reasons, easiness and collaboration by social networking
other DTL model is proposed in [32] where authors inte- which are same as analysed in [40]. The work [39] also
grate both constructivist and objectivist approaches to in- highlights the interests of students and teacher towards us-
structional design. When creating or redesigning courses ing SM for eLearning as they believe it will ease and en-
for Internet distribution, content and media developers, hance communication and collaboration between different
faculty and researchers benefit from the identification of a eLearning entities. The results of [39], [40] see similar
developmental model taking into account both learning trends as shown in earlier work [41] which highlight the
and design principles. popularity of social networking and its possibilities as an
Bong and Zhang proposed read, reflect, display, and education tool in Romania. A more recent work [42] which
do (R2D2) DTL framework particularly for online learning focuses on SM aspect in actively encouraging cognitive
[33]. Online learners have one universal characteristic that and meta-cognitive learner development while using a
they have diverse learning preferences. The first compo- heutagogical teaching and learning approach. This work
nent of R2D2 have methods to assist the learners to gain also indicated that familiarity of students with SM is an im-
knowledge through internet based media. For example portant factor that encourages use of SM in eLearning.
readings, virtual explorations, and online audio lesson and From course design perspective the use of SM is increased
podcasts. The second component of R2D2 model uses the in preparing new eLearning courses and packages [43].
reflective activities of the learner such as online blogging These literatures demonstrate that SM can play an im-
and self-assessment quizzes. The third component is based portant role in distance eLearning thus must be explored
on representations of the learning content such as anima- further with new analyses and approaches.
tions to visually present the content. The fourth and the Sentiment analysis (SA) refers to analyzing the text to
last component accentuates what learners can accomplish find out key insights about the opinion of the particular
with the material in hands-on activities such as real-time entity with respect to specific topic. This is achieved by
cases and scenarios. Further R2D2 model is explored in computational linguistics algorithms that extract particu-
[34]. As the finding suggests that R2D2 model is not effi- lar specific subject related information using social media
cient in terms of data collection methods. However it (SM) data like tweets, Facebook message, Tag lines, blogs
shows positive learning effects with vocabulary pronunci- and online articles. The main goal of SA is to determine the
ation, acquisition, note taking and presentations by the stu- opinion, attitude and sentiment of a writer regarding spe-
dents. In China online learning faced problems initially. cific topic. Current issues, trends and future prospects of
First teachers find hard to adapt and secondly, students SA are discussed comprehensively in surveys [44], [45].
particularly adult students feel over-burdened with allot of Papers also highlighted the importance of public opinion
learning content. To address these issues, in paper [35] a mining in modern day world and its social and commercial
framework is proposed which assist students in interacting aspects. For example businesses are using SA to perceive
with an automated question answering system. This helps customer’s product buying choices by their online SM be-
the teacher to analyze different learning patterns of the stu- haviors. This is achieved by analyzing tweets, Facebook
dents. Further teachers are able to organize online learning posts, blogs, tag-lines etc. In one [46] of such works, SA is
contents. Data mining is used to help both teachers and done for political events like elections. Psychometric in-
students to personalize the services. Another DTL frame- struments are used to extract six different types of senti-
work proposed in [36] which exploits data mining to help ments such as anger, depression, confusion, fatigue, ten-
teachers to manage the learning contents on the web using sion, vigor. Whereas in [47], nodes based SA using graphs
available infrastructure, needs and experience. Later learn- has been done that highlights social connection and their
ing content can be reorganized according to student's per- correlation. Multi-language SA is always an issue because
formance. most of the SA models are hard to extend for other lan-
guages. Addressing this problem in [48], which uses sim-
2.2.3 Social Media and Sentiment Analysis
ple heuristics and minimal linguistic computation is per-
Social media (SM) has gained a great deal of popularity formed to make it easily extendible to other languages.
during the last decade (see e.g. [37]). Now it is not merely SVM classifier is further used to demonstrate high perfor-
a platform that is used to do social communication but now mance of this model. Sentiment analysis is applied to vari-
it is more like an online world which is used by everybody ous fields to grasp the insights for mass opinions such as
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
finance [49], entertainment industry [50], [51], politics [47], time web and one of them in IoT, geographically distrib-
[52], sales [53] and business operations [54]. The literature uted collaboration, and scientific workflows. Each of these
review given above clearly shows that SA and Social media trends is examined and discussed, and a suggested re-
sentiment analysis (SMSA) are playing critical roles in search and development path is offered. Similarly in [59],
planning and policy making for businesses. author describes the process to translate a “traditional” re-
mote Web lab into a set of smart devices which are able to
2.2.4 IoT and Remote Laboratories work together in the cloud and IoT. Further in [60], authors
With the emergence of Internet of Things (IoT), DTL can have highlighted IoT technologies including actuators,
take the leverage from IoT’s infrastructure and networking web cameras, sensors, RFID, and Arduino which are used
technologies to provide any-time, any-where learning so- to provide WBRL facilities particularly in South African
lutions to individuals. Several different perspective of IoT environment.
enabled learning activities are discussed by Adorni et al, in
their work [55]. Undoubtedly in recent years, IoT is ma- 2.2.5 Activity Recognition & Behaviour Analysis
tured and practical applications of IoT are successfully de- Understanding of leaners’ or students’ remote environ-
ployed around the world. This is due to the two main rea- ment can provide key insights about when and how to de-
sons: (1) IoT is getting wider acceptance from the users as liver learning material, lessons and instructions. Further
well as the industry and (2) hardware manufacturing cost DTL systems develop learner dependent cognitive. These
is coming down with mass productions of sensors, proces- cognitive can later be utilized for planning student specific
sors, actuator and other enabling devices. Now affordable lessons. One of the such technology is of activity recogni-
mobile devices are in market and introduction of Internet tion facilitated by using IoT’s infrastructure [61]. Activity
Protocol version 6 (IPv6) provides far more bigger address- recognition assists in developing better cognitive about
ing space so that billions of IoT enabled "Things" can use learner’s behavior and daily routines. Once the under-
it. Both internet availability and use have increased several standing is achieved, DTL system can deliver lessons in an
folds and it is reaching the remote areas around the world. adaptive manner rather than fixed schedule to the stu-
The integration of IoT with DTL can be far more beneficial dents. For example with the help of activity recognition,
than what we think today [56], [57]. We will see later in this we identified student is free and not busy in any sort of
paper the use of IoT in our UTiLearn system for personali- work and not resting even. This prompt the DTL system to
sation, system optimisation, and higher quality of experi- deliver some important short lesson on his mobile device,
ence. this is only done at the time when student is free. This sort
An important area where IoT applications and infra- of learning will be very useful for the people who are
structure can be used to enhance DL, is by developing re- working as well as registered in DTL courses. Other exam-
mote laboratories. It is often seen that schools and univer- ple is by activity recognition, DTL application can record
sities in urban areas have well established and well main- student body language and nervous monuments. Whereas
tained laboratories. However in rural or under developed in the other scenario, where the student solving a online
areas laboratories are not available or available with lim- exams smart watches can be used to analyze his heart beat
ited and poorly maintained resources. Also, the idea of set- which is connected to learning application [62]. Further
ting laboratories in remote areas does not sound economi- heart beat records corresponding to answering time can be
cally feasible due to the small number of students per geo- studied to see on which type of question student is nervous
graphic area. In the past, students who are enrolled in DTL and stressed while answering.
courses had to go to traditional and physical laboratories
to gain practical knowledge and skills. Currently, thanks 2.2.6 Artificial Intelligence in Distance eLearning
to the advances in networking and communication tech- AI gained by machine learning (ML) and Deep learn-
nologies, programmers programming languages, new ing (DL) methods, nowadays applied in almost every field
eLearning tools are available to make the best use of learn- of our environment where we are living today. The use of
ing methodologies such as DL. Web-based remote labora- ML methods for data mining (DM) is quite old in DTL
tory (WBRL) is one of them. WBRL is an application which field. However, their initial implementations addressed
allows students to work with real laboratory equipment’s simple applications in DTL [63]–[65], like web access pat-
to perform their experiments. To facilitates this, learner tern mining to customize DTL systems as per individual
only need a computer with an Internet connection. How- requirements and On-Line Analytical Processing (OLAP)
ever, emerging technologies and other hardware such as of student data. Further in [66], [67], ML classifiers are used
mobile devices, cloud computing and IoT, require the de- to identify factors which influence the performance of the
sign of different models to support innovative learning ex- students in DL. This assists to upgrade DTL systems for
periences. For this reason, remote labs are being reconsid- better learning outcomes. One of the major problem in DTL
ered to smart device paradigm. Physically, the considered courses is higher student dropout rates as compared to tra-
smart device is made of the junction of a controller (com- ditional classroom learning which is addresses using vari-
puter server) connected to physical equipment on the one ous classifiers such as Naive Bayesian (NB), k-Nearest
hand, and to the Internet on the other hand. In [58] authors Neighbor (KNN), Decision Tree (DT), Support Vector Ma-
examine four emerging trends in the Internet that have the chines (SVM) and Artificial Neural Network (ANN) [68]–
potential to influence the development of WBRL: the real- [70]. Literatures above are purely based on providing data
mining functionalities based on student historic records.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 7
However, modern day DTL systems requirements are far 3 UTILEARN: PROPOSED FRAMEWORK
more beyond data mining. One issue with in past DTL sys-
We introduce in this section our UTiLearn framework, give
tems is the absence of auto guide which can instruct teach-
its overview, and present its algorithmic refinement.
ers to point out problems and possible solutions. An artifi-
UTiLearn is a distance eTeaching and eLearning (DTL)
cial advisor is used in to solve this problem [71] which uti-
framework designed for enhanced development, manage-
lizes fuzzy logic.
ment and delivery of teaching and learning in smart city
It is known that around the world students face hard-
and smart society settings. The framework is depicted in
ship in learning mathematical concepts and find it difficult
Fig. 2 and is further elaborated in Fig. 3. Moreover, we
to use them for problem solving. One of such system which
have developed a system, called the UTiLearn system,
addresses this by using AI power is ARTIMAT [72]. It is
based on the proposed framework. A prototype imple-
based on the idea to enhance mathematical problem solv-
mentation of the UTiLearn system is described and evalu-
ing abilities of the students. Further in [73], ARTIMAT is
ated in the next five sections (Sections 4 to 8).
evaluated and finding suggest that, it ease mathematical
problem solving procedure for the students and under-
standability of the mathematical concepts among the stu-
dents. Present day DTL systems are far more capable, user
friendly and automate. We believe that integrating new
technologies like IoT and AI can be useful for DTL systems.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
The UTiLearn framework leverages on advanced and The UTiLearn framework uses IoT as part of the dig-
emerging technologies (big data, HPC, AI, IoT, mobile ital infrastructure needed to connect with the user for op-
computing, and cloud) to provide an environment and timisation of teaching and learning. The IoT infrastructure
ecosystem for societies to enrich their knowledge, anytime, could include smart phone sensors, smart watch sensors,
anywhere, any age, through improved processes for collec- virtual reality technologies to enable physical-like experi-
tion, development, accessibility, pedagogy and cognition ences in distance eLearning environments, virtual labs,
of knowledge. Its users could be teachers, students, gov- and other sensors. Mobile computing technologies are
ernments, businesses, third parties, and virtual organisa- used to connect to the users, sense their continuous activi-
tions or groups formed for specific purposes. The roles are ties, and provide them anytime anywhere opportunity for
not exclusive; for example, an individual or organization DTL. Big data technologies are used to deal with the chal-
could be both teacher and learner based on her expertise, lenges related to the Vs of DTL data and benefit from the
contributions, and requirements. A learner might need to opportunities these technologies offer. Computational In-
learn a specific skill so she registers its interests, goals, telligence and AI technologies are used for analysis, pre-
learning levels, and timelines with the system. The system diction, optimisation and decision making. HPC technolo-
then suggest the learner the available courses or programs gies are needed to deal with high computational require-
and monitors her through to the completion of the learning ments of AI and other algorithms. A continuing trend is to
task. The system would maintain a profile for each user move businesses in the cloud environments using infra-
over time and use the profile to understand the user pref- structure as a service (IaaS) or other models. The UTiLearn
erences and other cognitive abilities. The individuals could system could be implemented on top of hybrid IaaS and
also contribute teaching material in the areas of their ex- can be offered distance teaching and learning as a Service
pertise. A rating system could be maintained for individu- (UTLaaS). Smart city technologies could be used to benefit
als and their contributed material based on user satisfac- from converged infrastructure for integrated, enhanced
tion. Certain groups or a government department can de- and optimized pedagogies, content delivery, and cognition
velop specific material to reach out to a group of people or of the users. The UTiLearn system hence would contribute
the whole population to create awareness of a topic for so- to the development of smart society; a digitally-enabled,
cial intervention purposes. An example is the change of knowledge-based society, aware of and working towards
people behavior towards better driving practices, cycling, social, environmental and economic sustainability.
or healthy lifestyles. Third parties could also provide their The framework (and the system) comprises five main
services in bridging the service gaps that the framework is components:
not able to provide directly. 1. Sentiment Analyser
2. User Activity Recognition & User Identification
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 9
(UARI) Component text) of content delivery to suit the user preferences and
3. Adaptive Content Delivery Mode Adviser (ACDMA) cognitive abilities can be advised by ACDMA. The AC-
4. System and Network Provisioning & Demand Predic- DMA Component could restrict and enable access to sys-
tion (SNPDP) Component tem resources based on the user identification and activi-
5. Performance Analysis, Prediction and Intervention ties. Moreover, ACDMA computes and predicts the num-
(P2AI) Component ber of applications and clients that are needed at different
times and geographical locations.
Algorithm
Input:
UTiLearn main
D ⃪ Tweets and online news articles
The System and Network Provisioning & Demand Pre-
𝐴 ⃪ 𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝐷𝐵 diction (SNPDP) Component uses the information pro-
Output: 𝐷𝐶𝐶 𝑎𝑛𝑑 𝑃𝑂𝑅 duced by ACDMA to compute and predict the system (e.g.
1. Go to Sentiment Analyser
2. Get D core-hours) and network (e.g. bandwidth) resources re-
3. Clean content of D quired by the UTiLearn system. The Performance Analy-
4. Perform SA on D
5. If Sentiments == positive sis, Prediction and Intervention (P2AI) Component stores
6.
7.
Then
Motivation exists.
and analyses information related to the interaction of the
8. Develop/Update the DTL system user with the UTiLearn system continuously in the back-
9. Move to line 12
10. Else ground to inform UTiLearn users about their performance.
11.
12.
Terminate / continue to monitor
Recognize activities and identify users
Moreover, it suggests interventions to the users to help
13. Go to UARI them address specific shortcomings and improve their out-
14. Apply DLANN to predict activities and
identify users comes.
15.
16.
Update 𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝐷𝐵
Prepare course content corresponding to
All five components of the system are designed to op-
identified activity timise personalisation for individuals as well as perfor-
17. Go to ACDMA
18. Look in to Activity-to-Content table to identify mance for the UTiLearn he system.
which course content matches predicted activity
19.
20.
Network Provisioning & Demand Prediction
Go to SNPDP
3.1 UTiLearn: An Algorithmic Refinement
21. Analyzing students temporal activity patterns The framework presented in Fig. 3 has been refined into
from historic activity data from 𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝐷𝐵
22. Apply DLANN for Temporal activity pattern the algorithm presented in Fig. 4. We explain the algorithm
23. Calculate and predict the short-term and
long-term network requirements
as follows. The Main UTiLearn Algorithm takes Tweets,
(bandwidth, delay, etc.) online news articles and activity data as input and feeds to
24. Feed into the system for real-time, planning
and provisioning purposes the Sentiment Analyser component which performs senti-
25.
26.
If course content match == available bandwidth
Then
ment analysis in order to transform the social media data
27. DCC to student into opinions of users, businesses and governments. The
28. Record student and system interaction
29. Else UARI Component then uses DLANNs for classifying user
30. Secure additional bandwidth if possible activities and identifying users. The ACDMA Component
31. If successful
32. Go to line 21 suggests the course content corresponding to user activity
Else
33.
34. Look for available content quality which type and computes the number of clients and applications
suites available bandwidth in content quality table to be served by the system. The SNPDP Component com-
35. Record student and system interaction
36. Predict Students performance with respect to his putes and predicts the UTiLearn system and network re-
37.
interaction with system
Go to P2AI
sources. The P2AI component periodically, and as needed,
38. Apply DLANN to predict performance predicts user performance based on their interaction and
39. Relate performance of the student with the mode of
delivery, individually and for all the students. the interaction of other users with the UTiLearn system. It
40. Send performance report and predicted final
course outcome. Intervene and give suggestion for improvements
warns and intervenes users as necessary. The reader may
also find Fig. 2 and Fig. 3 useful in understanding the main
Fig. 4. The UTiLearn main algorithm. algorithm. Moreover, the main algorithm will become
clearer as we indulge into explaining each component of
Fig. 2 lists these five components and summarises their the UTiLearn system, one by one, in the upcoming sec-
purpose within the UTiLearn framework. The framework tions.
is further elaborated in Fig. 3 with additional details in a
workflow style. The purpose of the Sentiment Analyser
Component is to analyse social media for forming the busi- 4 THE SENTIMENT ANALYSER
ness case of the decisions taken by the system. The senti- Considering the increasing importance and benefits of so-
ments for business case can be explored on strategic, tacti- cial media sentiment analysis in all spheres of our life, as
cal, operational and real-time levels. The User Activity highlighted in the literature review (see Section 2.2.1 ), we
Recognition & User Identification (UARI) Component have added a sentiment analysis function in our UTiLearn
analyses user activity data acquired from smart phones, system, called the Sentiment Analyser Component. The
smart watches and other IoT units to understand and pre- task of this component is to analyse social media for form-
dict spatio-temporal activities and other patterns of the ing the business case for the decisions taken by the system.
user. The Adaptive Content Delivery Mode Adviser (AC- The sentiments for business case can be explored on mul-
DMA) uses the user activity and identification related in- tiple levels, at strategic, tactical, operational and real-time
formation produced by the UARI component to optimize levels. For example, monitoring and analyzing sentiments
the DTL material delivery, and the system and network to understand the need for UTiLearn system presence in a
performance. For example, a better mode (video, voice,
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
Number (%)
60
Input: D ⃪ Tweets and news articles
X ⃪ {Positive, Negative, Neutral} 40
Output S ⃪ Sentiments 20
1. Prepare D 0
Positive Negative Neutral
2. Filter 𝐷 for special character, white spaces and symbols
Opinion
3. Apply Naïve Bayes (NB) classier Є P(X|Y) = 𝑃(𝑌|𝑋)𝑃(𝑋)/𝑃(𝑌)
4. Where 𝑃(𝑋|𝑌) is P(𝑋) class | Y ⃪ 𝐷(1,2,3,4….𝑛) Tweets Online News
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 11
providing new DTL systems with advanced learning fea- 2. To identify students based on the spatio-temporal ac-
tures. The approach presented in this section can be used tivity data.
to collect and analyse real time data regarding specific 3. To use the obtained temporal activity patterns for
course or programs, certain pedagogies, or modes of deliv- computing and predicting the number and type of cli-
ery to make informed decisions about the users and sys- ents and applications.
tems. The data can be analysed on spatio-temporal basis to 4. To compute and predict the bandwidth, latency and
understand the emerging behavior of people across times, other characteristics of the networks required by the
cities, countries and regions. UTiLearn system by using the data obtained in Point
(3) above.
Table 1 5. To compute and predict computational and other sys-
Benchmark datasets for activity recognition
tem characteristics required by the UTiLearn system
by using the data obtained in Point (3) above.
No Dataset Dataset Details Activities Year
Name
1 Ugulino [76] The dataset is recorded using 5 activity classes: standing-up, 2012 Table 2
accelerometers for 8 hours by 4 standing, walking, sitting-down,
participants. and sitting.
HAR datasets and their respective input parameters for DLANN
2 Casale [77] The dataset is recorded using 7 activity classes: working at 2012
accelerometers by 15 partici- computer, standing, standing up,
No Dataset Epochs Two Layers
pants. walking going up-down stairs,
walking and going up-down 1 Ugulino [76] 800 400 Neurons each
stairs, talking while standing,
walking and talking with some- 2 Casale [77] 800 400 Neurons each
one.
3 Reiss [78] 60 200 Neurons each
3 Reiss [78] The dataset is recorded using 18 activity classes: cycling, run- 2012
inertial measurement units and ning, walking, Nordic walking, 4 Stisen [79] 120 60 Neurons each
a heart rate monitor by 9 partic- standing, sitting, lying, watching
ipants. TV, working on computer, driv-
Smart Watch accelerometer
ing car, climbing stairs, coming 5 Stisen [79] 80 60 Neurons each
down the stairs, vacuum clean-
ing, house cleaning, ironing, fold- Smart Watch gyroscope
ing laundry, rope jumping, play-
ing soccer, rope jumping and 6 Stisen [79] 60 180 Neurons each
other (transient activities).
Smart Phone accelerometer
4 Stisen [79] The dataset is recorded using 6 activity classes: standing, sit- 2015
smart watch sensors (accel- ting, walking, biking, stair-up
7 Stisen [79] 100 50 Neurons each
erometers and gyroscope) and and Stair-down. Smart Phone gyroscope
smart phone sensors (accel-
erometers and gyroscope) by 9
participants.
5 Anguita [80] The dataset is recorded using 6 activity classes: walking Up- 2013 It is possible to compute the application characteristics
smartphones by 30 partici-
pants.
stairs, walking downstairs, walk-
ing, standing, sitting and laying.
that would be needed to support the various teaching ma-
6 Roggen [81] The dataset is recorded using 4 activity classes: walk, stand, sit 2010
terial via different modes of content delivery over time.
ambient and on-body sensors
by 4 participants.
and lie. This in turn would allow planning the networks that are
needed to support the system generated traffic. The net-
7 Banos [82] The dataset is recorded using 33 sports related activities 2014
accelerometers and gyroscope. work planning, both long-term and short-term, could be
8 Reyes-Ortiz The dataset is recorded using 6 activity classes: walking, walk- 2016 done by analyzing the historical and real-time activities
[83] smartphone sensors such as ac-
celerometers and gyroscope by
ing upstairs, walking downstairs,
standing, sitting, lying and other data.
30 participants.
5.1 Methodology
A number for human activity recognition datasets are
5 ACTIVITY RECOGNITION & USER IDENTIFICATION available publicly [76]–[83]. Table 1 lists the most widely
used eight human activity recognition datasets. The da-
(UARI) COMPONENT tasets have not been given a specific name by their origina-
Integration of Human Activity Recognition (HAR) and IoT tors except the PAM (Reiss [78]) and Opportunity (Roggen
has gained significant importance in recent times, particu- [81]) datasets. For presentation and brevity reasons, we
larly in context-aware systems. Accurate and reliable HAR have named each dataset according to its first author’s sur-
has been used in several applications such as improving name (see Column 2). The third column gives brief details
quality-of-care for infants, elderly and people with poor of each dataset and Column 4 lists the name of the activity
health conditions [61]. It has been found helpful in opti- classes. Column 5 lists the year of publication of the respec-
mizing energy consumption of IoT based smart homes tive datasets. The accuracy of the location reported in the
[84]. It can also help to predict various human behaviors datasets can be found based on the devices used in the ex-
such as rash driving styles [85] and emergency situations periments.
in advance like aggressive actions which can lead to vio- We have used Random Forest (RF) generated feature
lence [86] using smartphone based sensors. In the current importance (relevance) to select a subset of attributes from
prototype implementation of the UTiLearn system, we the datasets based on their relevance [87]. We then fed da-
have used human activity recognition for the following taset to H2O’s deep learning artificial neural networks
goals: (DLANNs) algorithm [88], which is based on a multi-layer
1. To provide adaptive educational content delivery to feed forward ANNs that are trained using back-propaga-
the students who are registered for a course. tion (BP) with stochastic-gradient descent. We divided da-
taset into two parts, 60% for the training phase and 40% for
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
the testing phase. We have used a different number of iter- readable. The y-axis in the figure gives the list of activities
ations (for different datasets) with two hidden layers each that have been identified from the datasets using our algo-
containing a certain number of neurons. The number of it- rithm. The x-axis gives the time dimensions of the human
erations (epochs) and the number of neurons used can be activity (the time is relative, not absolute). The space di-
found in Columns 3 and 4 of Table 2, respectively. mension is represented by box sizes (z space dimension),
The Algorithm for the Activity Recognition & User box shape outline colours (x space dimension), and shape
Identification Component of the UTiLearn system as de- fill colours (y space dimension). The largest box represents
scribed above is given in Fig. 7. Lines 1 and 2 in the algo- the highest Z value of space dimension; red colour in the
rithm set response variables to the activity and user class. box outline represents the largest X value of the space di-
The algorithm is deliberately written descriptively and is mension; and the light blue colour represents the highest Y
self-explanatory. value of the space dimension. The figure shows, for exam-
ple, that the sitting activity was only carried out between 5
Algo- Activity Recognition & User Identification (ARUI) minutes to 12.5 minutes, and the biking activity was rec-
rithm orded between 22 to 33 minutes. The time for other activi-
Input: 𝑀 ⃪ 𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝐷𝐵
Fig. 7. The Algorithm for the Activity Recognition & User Identifica-
tion (ARUI) Component of the UTiLearn system.
5.2 Results
We used the datasets given in Table 1 and the algorithm
given in Fig. 7 to calculate spatio-temporal activities of the
users. These are depicted in Fig. 8 for the Stisen smart
phone accelerometer dataset [79]. The dataset contains ap-
proximately one million records. However, in the figure,
we have used 5% sample data from the dataset to make it
Fig.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. 9. Visualization
Personal of thebuthuman
use is also permitted, activity recognition
republication/redistribution results.
requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 13
6 ADAPTIVE CONTENT DELIVERY MODE ADVISER mode, availability of learning material and mode of deliv-
(ACDMA) ery, availability of network bandwidth and latency, etc.
For example, due to the safety reasons, it may not be safe
The Activity Recognition & User Identification Compo- to deliver any course content to a user who is biking, driv-
nent described in the previous section is able to recognize ing or climbing stairs; it may be ok, arguably, to deliver
human activities and identify them based on their activi- voice-based course material to this user. The user, how-
ties. The purpose of the Adaptive Content Delivery Mode ever, can override anytime the system suggested mode of
Adviser (ACDMA) that we will discuss in this section is to delivery and teaching material subject to legal and proce-
use the user activity and automatic identification infor- dural constraint. The algorithm for ACDMA is given in Fig-
mation to optimize the DTL material delivery, and the sys- ure 10 .
tem and network performance. We have already men- We now look at a potential scenario that we have de-
tioned the potential uses of the Activity Recognition & veloped as a prototype implementation of the Adaptive
User Identification Component as a list at the beginning of Content Delivery Mode Adviser Component. The scenario
Section on . The T&L material delivery can optimized, for is motivated by our earlier work on a healthcare system
example, for suggesting a better mode (video, voice, text) [11]. Fig. 11 depicts a potential scenario where UTiLearn
of content delivery to suit the preferences, learning charac- subsystems in four countries are connected with each other
teristics, pedagogical requirements, and cognitive abilities through IP communications (this is for demonstration pur-
of the users. The system performance can also be opti- poses and we would expect the UTiLearn system to have
mized by using the user activity and identification data for global presence). Fig. 12 gives a more detailed networked
system planning or system performance prediction; for system architecture by narrowing down the system be-
motivation see [89], where we have studied system perfor- tween two cities, Munich in Germany and Jeddah in Saudi
mance for healthcare systems in computational Grid envi- Arabia. The infrastructure does not need to be owned by
ronments. The user activity and identification data can also the UTiLearn system, it could be acquired as an infrastruc-
be used for network planning, location based services, pre- ture or platform as a service.
serving user privacy and security, including user authen- Table 3 gives the number of client applications distrib-
tication and authorization [90]. uted across three types of distance eLearning places (see
Fig. 12: remote classrooms, homes, and mobile devices) be-
Algorithm Adaptive Content Delivery Mode Adviser (AC-DMA)
tween Munich and Jeddah. We have kept the number of
Input: user_activity, user_preferences, other_factors devices and eLearning places symmetric across the cities in
Output: DCC and receive acknowledgement accordance with the original article [11] in order to train
1. Relate course content delivery media to our model with the networking related data and validate
2. Activity Type the networking results. Further details of networking
3. Application type training and validation are provided in Section 7.
4. Network type
5. Device type
6. DCC appropriately select content for student
7 SYSTEM AND NETWORK PROVISIONING &
7. Received delivery acknowledgement from student
DEMAND PREDICTION (SNPDP) COMPONENT
8. Update course content database By now, we have described and evaluated three of the five
9. Compute the appropriate mix of applications and devices components of the UTiLearn system. The first component
of the UTiLearn prototype system, the Sentiment Analyser,
Figure 10. Algorithm for Adaptive Content Delivery Mode Adviser provided us the results about the opinions of individuals,
(ACDMA). governments and businesses on distance eLearning. This
component could advise on making decisions regarding
Our earlier work has looked into creating geographic specific courses, pedagogies, or modes of content delivery.
firewalls based on user location data allowing us to restrict The second component, Activity Recognition & User Iden-
and enhance services for individual users within a specific tification, provided us the information on the users, their
geographic area such as a lecture or examination room locations and activities. This in turn enabled us, using the
[91]. The control of distant physical or virtual environ- third UTiLearn Component, Adaptive Content Delivery
ments, and the users within them, is typically a difficult Mode Adviser (see the previous section) to compute and
task given the remote nature of distance eLearning envi- predict the number of applications, clients that are needed
ronments. This work has been extended to policy based across time and space (see Fig. 13 and Table 3). In this sec-
control of pervasive mobile learning environments using tion, we describe the System and Network Provisioning &
detection and resolution of feature interaction [92]. The ap- Demand Prediction Component and present its results.
proaches and works mentioned above could alleviate these The SNPDP Component uses the (applications and clients)
problems. results obtained from the ACDMA Component to compute
To explain the function of ACDMA further, consider and predict the bandwidth, latency and other characteris-
Fig. 8 which plots human activities across time and space. tics of the networks required by the UTiLearn system.
The course material is provided to the user based on its Moreover, it is also able to use the results from the AC-
preferences and other contextual information such as the DMA Component to compute and predict computational
user current and future activity, light or heavy study
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
and other system characteristics (e.g. how much computa- network characteristics (Fig. 12) from the ACDMA Com-
tional power and memory would be needed) required by ponent to compute and predict the future network and sys-
the UTiLearn system. It could even be possible to compute tem requirements of the UTiLearn system. This could be
and predict energy requirement of the UTiLearn system done on long-term, short-term and near real time levels.
based on the computational system and network load. We have used DLANN in this paper. However, we wish to
compare and augment it with other computational intelli-
gence methods such as Markov chains. Some of our earlier
7.1 Methodology work on modelling (including large Markov models) of
Our methodology to design and implement the SNPDP communication networks and other applications could be
Component is to use the application traffic (see Table 3) and found in [93]–[98].
Fig. 11. UTiLearn System with networked communications between four countries.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 15
Algorithm System and Network Provisioning & Demand Prediction (SNPDP) vided the data into 60/40: 60% of the data is used for train-
Input: 𝑁𝑐 ⃪ {𝑑, 𝑁𝑏 , 𝑁𝑡𝑦𝑝𝑒 } ing and 40% for testing.
𝐴𝑝𝑝𝑐 ⃪ {𝐴𝑝𝑝𝑛 , 𝐷𝑞 } The data used to train the networks in this prototype
Output 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑡𝑟𝑎𝑓𝑓𝑖𝑐 ⃪ {𝑇𝑟 , 𝑇𝑠, 𝑇𝑙 } implementation of the SNPDP component is taken from
1. Measure 𝑁𝑐 [11] (see also Table 3 and Fig. 13). A detailed network mod-
2. Aggregate 𝐴𝑝𝑝𝑐 elling related work has been carried out in [11] to explore
3. Feed to DLANNs ⃪ {𝑁𝑐 , 𝐴𝑝𝑝𝑐 } QoS performance specifically for intracity, intercity, and
4. Set DLANNs P
global environments for healthcare systems. They have
5. Train and Test DLANNs
used a variety of applications with different network and
6. Tune P
system sizes. Future work will focus on gathering more
data, both in terms of quantity and diversity, for network
7. If (P== best results)
training purposes. The SNPDP algorithm provides the
8. Then
Network planning functionality; the function to calculate
9. M ⃪ 𝐷𝐿𝐴𝑁𝑁𝑠
and predict computational characteristics of the UTiLearn
10. Else
system could be implemented in a similar manner except
11. Tune P Further
that the attributes used in the function would be system
12. Go to Line 4
related such as computational time, memory etc (see [89]
13. 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑡𝑟𝑎𝑓𝑓𝑖𝑐 ⃪ 𝑀(𝑁𝑐 , 𝐴𝑝𝑝𝑐 ) for motivation). The load balancing of UTiLearn servers
14. Secure bandwidth for 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑡𝑟𝑎𝑓𝑓𝑖𝑐 would also be required to create high availability, we plan
to rely on our earlier work in [99] for this purpose.
Fig. 13. System and Network Provisioning & Demand Prediction
(SNPDP).
Video Performance Comparison: IntraCity
Table 3
4
Client applications and distance eLearning places
3.5
3
Munich, Germany Jeddah, Saudi Arabia
Delay (ms)
2.5
2
1.5
Place Client Applications Place Client Applications 1
0.5
FTP DB Video FTP DB Video 0
M96
L96
H96
VH192
VH384
VH768
M1248
H1248
L1248
UH192
UH384
UH768
VH96
M192
M384
M768
L192
H192
L384
H384
L768
H768
UH96
VH1248
UH1248
Clients Clients Clients Clients Clients Clients
78 78 52 78 78 52
Remote Remote
Classroom Classroom
Actual Predicted
78 78 52 78 78 52
Home Home
Fig. 14. IntraCity Video Delay for Jeddah and Munich (symmetric):
Actual Data vs Predicted by DLANN.
78 78 52 78 78 52
Anywhere Anywhere
106
Total 1248 Clients
104
102
100
98
The SNPDP Algorithm of the UTiLearn system for ap- 96
94
plications bandwidth prediction and network provision- 92
ing is given in Fig. 13. The algorithm uses 𝑁𝑐 for the net-
M96
L96
H96
VH192
VH384
VH768
M1248
H1248
UH192
UH384
UH768
L1248
VH96
M192
M384
M768
L192
H192
L384
H384
L768
H768
UH96
VH1248
UH1248
brevity reasons we will present the video results alone. The about her worsening performance due to a pattern in her
predicted results using the DLANN models are compared course related activities and inactivity. A general perfor-
with actual response times taken from [11]. Fig. 14 plots the mance pattern can be identified for the user population
video delay values encountered by the clients for 25 differ- based on their activities and inactivity and this can be used
ent traffic profiles in intracity environment. The term in- to anticipate, predicted and inform individuals and groups
tracity means that all the clients and servers are within the about their performance. Moreover, the P2AI Component
same city; Jeddah or Munich. Both the cities will have same can suggest interventions (strategies, actions, courses,
performance because the architecture and scenarios are modes of delivery, pedagogy related suggestions) to the
same. The traffic profile L96 in the figure means low net- user to help her address specific shortcomings and im-
work traffic for 96 nodes and UH1248 profile means ultra prove her outcomes. Such awareness and capabilities of
high traffic with 1248 nodes. The details of all these traffic the UTiLearn system about its users would allow users to
profiles, the bitrates etc. can be found in [11]. Similarly, Fig. get real-time, short-term and long-term detailed feedback
15 plots the video delay values for intercity traffic between and take necessary actions to improve their performance.
Jeddah and Munich. Note that the difference between the The system will report performance to the users by differ-
actual and predicted values is mostly minimal. Fig. 16 de- ent means including email, SMS, multimedia message and
picts the root mean square error (RMSE) between the pre- browser pop-ups. These functionalities of the P2AI Com-
dicted and actual values for video. This RMSE value is ponent are captured in the P2AI process cycle, depicted in
computed for all the network traffic including video, FTP Fig. 17.
and database queries. Note that the prediction error is very
small.
8.1 Methodology
Fig. 18 gives the performance prediction and intervention
algorithm for a prototype implementation of the P2AI
Component. The core of the P2AI component relies on the
H2O DLANNs algorithm [88], which is based on a multi-
layer feed forward ANNs that are trained using back-prop-
agation (BP) with stochastic-gradient descent. We divided
Fig. 16. RSME shows the differences between predicted and actual dataset into two parts, 60% for training and 40% for testing.
values obtained by DLANNs model.
We performed 400 iterations with two hidden layers each
containing 800 neurons.
8 THE PERFORMANCE ANALYSIS, PREDICTION For testing the P2AI Component, we have used a real
dataset [7] available online by edX published in 2014. The
AND INTERVENTION (P2AI) COMPONENT
major feature of this data [7] is that, it describes degree of
The Performance Analysis, Prediction and Intervention interaction between the student and edX. The edX online
(P2AI) Component is the fifth and last component of the education entity is a massive open online course (MOOC)
UTiLearn system. Its need and design is motivated by the provider, founded in year 2012 by the collaborative efforts
fact that the users (students, teachers and others) would of Harvard and MIT. It offers high quality online courses
like to know whether the objectives they had set for their that are delivered and designed by some of the best minds
teaching and learning are being met, the extent of their suc- around the world.
cess or failure, and the detailed reasons behind their The edX data that we used contained a large number
achievements and shortcomings. The P2AI Component of missing values. We therefore discarded rows which had
stores and analyses the information continuously in the more than four missing columns; i.e. the columns which
background to inform UTiLearn users about their perfor- were related to our analysis, containing students’ event-
mance. The historical and current information about a based records describing student and system interaction
user’s activities is analysed continuously to inform the user (Data Cleansing: Fig. 17). Further we imputed missing val-
about their current performance. The information related ues using MICE [100]. Subsequently, we selected a subset
to the individual users, as well as the collective infor- of attributes which are most relevant by using RF gener-
mation, is also used to predict course outcomes and other ated feature importance [87]. A subset of these selected fea-
user outputs. For example, a student can be informed
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 17
tures, only those that describe student and system interac- population interacted with the edX. The student
tion, are listed in Table 4. Using the selected features data, population is separated into two groups: the students who
we performed predictions based on the “certified” attrib- gained upto 50% marks (depicted by the yellow color bars)
ute (not mentioned in the table) with class label of 1 for and those who gained between 51% to 100% marks
certified (successful) students and 0 for unsuccessful stu- (depicted by the blue color bars). The average number of
dents. times the student population interacted with edX are
Table 4 plotted in four different categories: Average videos played,
Features that describes student and system interaction
average chapters read, average activities, and average
events (see Fig. 19. A clear message from the figure is that
No Feature Description
the students who received higher marks (blue bars)
1 explored Student who accessed minimum the half chapter of
interacted with the system significantly higher number of
a course
times (3239 vs 1051) than those who received lower marks
2 last_event_DI Last date of interaction between student and DTL
(yellow bars). We could evidently imply that a student’s
3 nevents Total interactions of the student with DTL system
marks are directly proportional to the number of times the
for the particular course
student interacts (uses the course content) with the system.
4 ndays_act Total days student interacted with the DTL system
Fig. 20. The classification error plot of the P2AI algorithm.Fig. 20
5 nplay_video Total videos played by student for the particular
depicts the accuracy of the DLANN prediction algorithm
course
(see Fig. 18) in terms of the classification error. It shows a
6 nchapters Total chapters from which student interacted
very small error (the y-axis scale is very small). The algo-
rithm achieved a 99.83% classification accuracy, a 0.0018
Algorithm P2AI: Performance Prediction and Intervention
mean square error (MSE), and a 0.0427 root mean square
Input: 𝑀 ⃪ 𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝐷𝐵 (historic data of student who finished
error (RMSE). These are relatively high accuracies
the course) achieved by the use of DLANNs algorithm though, as
𝐶 ⃪ 𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝐷𝐵 (on going student) mentioned earlier, it incurs high computational costs (see
𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑟𝑒𝑝𝑜𝑟𝑡 ⃪ 𝑆𝑡𝑢𝑑𝑒𝑛𝑡 ′ 𝑠 𝑖𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛 𝑤𝑖𝑡ℎ 𝐷𝐿𝑆 [10]).
Output: Predicted 𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝑜𝑢𝑡𝑐𝑜𝑚𝑒
1. Set response variable → Certified class name
2. Impute missing values using Multiple Imputation by Chained Equations
(MICE)
3. If Data matrix 𝑀 contain NAs
4. Then
𝑥1 𝑥2 𝑥3
Let 𝑀 = 𝑥4 𝑁𝐴 𝑥6
5. 𝑥7 𝑥8 𝑁𝐴
Imputation function = 𝐼𝑚𝑝𝑢𝑡𝑒𝑀𝐼𝐶𝐸 (𝑀)
𝑥1 𝑥2 𝑥3
𝑀𝑖𝑚𝑝𝑢𝑡𝑒𝑑 = 𝑥4 𝑥5 𝑥6
6. 𝑥7 𝑥8 𝑥9
7. Go to line 10
8. Else
9. Go to line 10
10. 𝑓1 , 𝑓2 , 𝑓3 , 𝑓4 … . 𝑓𝑛 ⃪ feature space of 𝑀 or 𝑀𝑖𝑚𝑝𝑢𝑡𝑒𝑑
11. Compute weight 𝑊(𝑥1,𝑥2 ,𝑥3….𝑥𝑛) using Random Forest feature selection Fig. 19. Students' scores against different type and number of inter-
12. Select subset of 𝑊𝑚𝑎𝑥 , | new data matrix 𝑀𝑠𝑢𝑏𝑠𝑒𝑡 = 𝑊𝑚𝑎𝑥 (M) actions with UTiLearn.
13. Set parameters of DLANN
14. Divide 𝑀 or 𝑀𝑖𝑚𝑝𝑢𝑡𝑒𝑑 into 𝑀𝑡𝑟𝑎𝑖𝑛 ⃪ 60% 𝑜𝑓 𝑛𝑟𝑜𝑤(𝑀) and 𝑀𝑡𝑒𝑠𝑡 ⃪ 𝑀 −
𝑀𝑡𝑟𝑎𝑖𝑛
15. Training and testing using DLANN
16. Parameter tuning if required, go to line 13
17. Feed 𝐶 to DLANN
18. Get 𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝑜𝑢𝑡𝑐𝑜𝑚𝑒
19. Send 𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝑜𝑢𝑡𝑐𝑜𝑚𝑒 report to students from time to time
20. Warn and intervene student by Email, Multimedia, text messages
and web browser pop-ups
21. Suggest possible way to improve 𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝑜𝑢𝑡𝑐𝑜𝑚𝑒
8.2 Results
The results related to the student performance obtained
through the P2AI Component prototype are depicted in
Fig. 19. The figure plots the number of times the student
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 19
These are required to enhance the precision of the training, Networked Healthcare Systems: A Review and Case Study,”
computation and prediction of networking characteristics, Big Data Concepts, Methodol. Tools, Appl., pp. 2429–2457, 2016.
application traffic models, and system models. In this re- [12] A. Chianese, F. Marulli, F. Piccialli, P. Benedusi, and J. E.
spect, our earlier work on city transportation modelling, Jung, “An associative engines based approach supporting
evacuation, and telematics, such as in [106], [107], [108], collaborative analytics in the Internet of cultural things,”
[109], would be useful both from activity recognition and Futur. Gener. Comput. Syst., vol. 66, pp. 187–198, 2017.
learning point of view. That is, information, learning and [13] I. Arel, D. C. Rose, and T. P. Karnowski, “Deep Machine
entertainment would be integrated, and system optimiza- Learning - A New Frontier in Artificial Intelligence Research
tion would benefit from all kinds of (big) data. [Research Frontier],” IEEE Comput. Intell. Mag., vol. 5, no. 4,
Finally, the use of computational and artificial
pp. 13–18, 2010.
intelligence and human-machine interaction demands a
[14] X. Du, Y. Cai, and S. Wang, “Overview of deep learning,”
careful look at the design of UTiLearn and similar systems.
Youth Acad. Annu. Conf. Chinese Assoc. Autom., 2016.
There is need to consider the theory and practice of
[15] “What is high performance computing?,” Inside HPC.
sociotechnical systems [110]–[113].
[Online]. Available: http://insidehpc.com/hpc-basic-
training/what-is-hpc/.
ACKNOWLEDGEMENT [16] Ryann K. Ellis, “A Field Guide to Learning management
The authors would like to acknowledge the technical and systems,” 2009.
financial support from the Deanship of Scientific Research [17] B. Khatri, P. Chouskey, and M. Singh, “Comparative analysis
(DSR) at the King Abdulaziz University (KAU), Jeddah, study of e-learning and traditional learning in technical
Saudi Arabia, under the grant number 1437/611/715. The institution,” Proc. - 2013 Int. Conf. Commun. Syst. Netw.
experiments reported in this paper were performed on the Technol. CSNT 2013, pp. 770–773, 2013.
Aziz supercomputer at KAU. We are also thankful to UCI [18] C. Sorgenfrei and S. Smolnik, “The Effectiveness of E-
data repository for making datasets available. Learning Systems : A Review of the Empirical Literature on
The corresponding author for this paper, Rashid Learner Control,” Decis. Sci. J. Innov. Educ., vol. 14, no. 2,
Mehmood, could be contacted via email at 2016.
RMehmood@kau.edu.sa. [19] P. H. Winne, “A metacogntive view of individual differences
in self-regulated learning,” Learn. Individ. Differ., vol. 8, no. 4,
REFERENCES pp. 327–353, 1996.
[20] B. J. Zimmerman, “Investigating Self-Regulation and
[1] “OpenCourseWare,” Wikipedia. [Online]. Available:
Motivation: Historical Background, Methodological
https://en.wikipedia.org/wiki/OpenCourseWare.
Developments, and Future Prospects,” Am. Educ. Res. J., vol.
[2] L. Burke, “University degrees irrelevant to big employers,”
45, no. 1, pp. 166–183, 2008.
new.com.au, 2016.
[21] E. Danili and N. Reid, “Cognitive factors that can potentially
[3] L. Sherriff, “Ernst & Young Removes Degree Classification
affect pupils’ test performance,” Chem. Educ. Res. Pract., vol.
From Entry Criteria As There’s ‘No Evidence’ University
7, no. 2, p. 64, 2006.
Equals Success,” The Huffington Post UK, 18-Jan-2016.
[22] M. S. Rosli, N. S. Saleh, B. Aris, M. H. Ahmad, A. A. Sejzi,
[4] L. A. Tawalbeh, A. Basalamah, R. Mehmood, and H.
and N. A. Shamsudin, “E-Learning and Social Media
Tawalbeh, “Greener and Smarter Phones for Future Cities:
Motivation Factor Model,” Int. Educ. Stud., vol. 9, no. 1, p. 20,
Characterizing the Impact of GPS Signal Strength on Power
2015.
Consumption,” IEEE Access, vol. 4, no. c, pp. 858–868, 2016.
[23] S. Essam and J. Al-Ammary, “The Impact of Motivation and
[5] A. Caragliu, C. Del Bo, and P. Nijkamp, “Smart Cities in
Social Interaction on the E-Learning at Arab Open
Europe,” 3rd Cent. Eur. Conf. Reg. Sci., pp. 45–59, 2009.
University, Kingdom of Bahrain,” Creat. Educ., vol. 4, no. 10,
[6] N. Komninos, “Intelligent cities: Variable geometries of
pp. 21–28, 2013.
spatial intelligence,” Intell. Build. Int., vol. 3, no. 3, pp. 172–
[24] B. Cheng, M. Wang, J. Moormann, B. A. Olaniran, and N. S.
188, 2011.
Chen, “The effects of organizational learning environment
[7] A. D. Ho, J. Reich, S. O. Nesterko, D. T. Seaton, T. Mullaney,
factors on e-learning acceptance,” Comput. Educ., vol. 58, no.
J. Waldo, and I. Chuang, “HarvardX and MITx: The First
3, pp. 885–899, 2012.
Year of Open Online Courses, Fall 2012-Summer 2013,”
[25] Y. Park and K. Lim, “Effects of Environmental and Human
SSRN Electron. J., no. 1, pp. 1–33, 2014.
Constructs on e – learning Effectiveness in Online University
[8] ITU, “Recommendation ITU-T Y.2060 (06/2012),” 2012.
Settinings,” Indian J. Sci. Technol., vol. 8, no. January, pp. 103–
[9] D. Evans, “The Internet of Things - How the Next Evolution
109, 2015.
of the Internet is Changing Everything,” CISCO white Pap.,
[26] “E-Learning market trends and forecast: 2014 - 2016 Report
no. April, pp. 1–11, 2011.
by Docebo,” 2014.
[10] F. Alam, R. Mehmood, I. Katib, and A. Albeshri, “Analysis
[27] I. E. Allen and J. Seaman, “Tracking Online Education in the
of Eight Data Mining Algorithms for Smarter Internet of
United States,” 2015.
Things ( IoT ),” Int. Work. Data Min. IoT Syst. (DaMIS 2016),
[28] K. Leach and S. Walker, “Internet-Based Distance
vol. 98, no. DaMIS, pp. 437–442, 2016.
Education : Barriers , Models , and New Research,” Proc.
[11] R. Mehmood, M. A. Faisal, and S. Altowaijri, “Future
WebNet World Conf. WWW Internet, 2000.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
[29] S. Mayes, T. and de Freitas, “JISC e-Learning Models Desk Network and Sentiment Analysis on Twitter : Towards a
Study Stage 2 : Review of e-learning theories , frameworks Combined Approach,” 2011.
and models,” Learning, vol. 202, no. September, p. 43, 2004. [48] G. Paltoglou, “Sentiment Analysis in Social Media,” Online
[30] M. Aydemir, E. E. Özkeskin, and A. A. Akkurt, “A Collect. Action, no. June, pp. 3–17, 2014.
Theoretical Framework on Open and Distance Learning,” [49] P. D. Azar, “Sentiment Analysis in Financial News,” vol.
Procedia - Soc. Behav. Sci., vol. 174, no. 2003, pp. 1750–1757, Bachelor o, p. 68, 2009.
2015. [50] S. Asur and B. A. Huberman, “Predicting the future with
[31] M. Leontidis, C. Halatsis, and M. Grigoriadou, “Using an social media,” Proc. Int. Conf. Web Intell. Intell. Agent Technol.,
affective multimedia learning framework for distance pp. 492–499, 2010.
learning to motivate the learner effectively,” Int. J. Learn. [51] V. K. Singh, R. Piryani, A. Uddin, and P. Waila, “Sentiment
Technol., vol. 6, no. 3, pp. 223–250, 2011. analysis of movie reviews: A new feature-based heuristic for
[32] K. Passerini and M. J. Granger, “A developmental model for aspect-level sentiment classification,” 2013 Int. Mutli-
distance learning using the Internet,” Comput. Educ., vol. 34, Conference Autom. Comput. Commun. Control Compress. Sens.,
no. 1, pp. 1–15, 2000. 2013.
[33] C. J. Bonk, K. Zhang, C. J. Bonk, and K. Zhang, “Introducing [52] A. Ceron, L. Curini, and S. M. Iacus, “Using Sentiment
the R2D2 Model: Online learning for the diverse learners of Analysis to Monitor Electoral Campaigns: Method Matters--
this world Introducing the R2D2 Model,” Distance Educ., vol. Evidence From the United States and Italy,” Soc. Sci. Comput.
27, no. 2, pp. 249–264, 2006. Rev., vol. 33, no. 1, pp. 3–20, 2014.
[34] H. Cartner and J. Hallas, “Exploring the R2D2 model for [53] L. C. Wood, T. Reiners, and H. S. Srivastava, “Expanding
online learning activities to teach academic language skills,” sales and operations planning using sentiment analysis:
Proc. ASCILITE Auckl., 2009. demand and sales clarity from social media,” ANZAM 2013,
[35] R. Shen, P. Han, F. Yang, Q. Yang, and J. Z. Huang, “An open Stream 15 Technol. Innov. Supply Chain Manag., pp. 1–17, 2013.
framework for smart and personalized distance learning,” [54] M. Zitnik, “Using sentiment analysis to improve business
Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. operations,” XRDS Crossroads, ACM Mag. Students, vol. 18,
Intell. Lect. Notes Bioinformatics), vol. 2436, pp. 19–30, 2002. no. 4, p. 42, 2012.
[36] S. A. Khoja, F. Sana, A. Karim, and A. A. Rehman, [55] G. Adorni, M. Coccoli, and I. Torre, “Semantic web and
“Implementing constructivist pedagogical model in internet of things supporting enhanced learning,” J. E-
dynamic distance learning framework,” Commun. Comput. Learning Knowl. Soc., vol. 8, no. 2, pp. 23–32, 2012.
Inf. Sci., vol. 20 CCIS, pp. 191–201, 2008. [56] I. Asseo, M. Johnson, B. Nilsson, C. Neti, and T. Costello,
[37] F. Piccialli and J. E. Jung, “Understanding Customer “The Internet of Things: Riding the Wave in Higher
Experience Diffusion on Social Networking Services by Big Education,” EDUCAUSE Review, 2016. [Online]. Available:
Data Analytics,” Mob. Networks Appl., pp. 1–8, Dec. 2016. http://er.educause.edu/articles/2016/6/the-internet-of-
[38] F. Manjoo, “Social Media’s Globe-Shaking Power,” The New things-riding-the-wave-in-higher-education.
York Times, 16-Nov-2016. [57] D. Kurzweil and S. Baker, “The Internet of Things for
[39] V. Balakrishnan and H.-S. Loo, “Social Media in E-Learning: Educators and Learners,” EDUCAUSE Review, 2016.
An Empirical Analysis among Students and Academicians,” [Online]. Available:
GSTF J. Comput., vol. 2, no. 4, 2013. https://er.educause.edu/articles/2016/8/the-internet-of-
[40] X. Li, K. Ganeshan, and G. Xu, “The role of Social things-for-educators-and-learners.
Networking Sites in e-learning,” 2012 Front. Educ. Conf. Proc., [58] M. F. Schulz and P. Long, “Possible Futures for Remote
pp. 1–6, 2012. Laboratories,” Internet Access. Remote Lab. Scalable E-Learning
[41] A. Stanciu, F. Mihai, and O. Aleca, “Social Networking As an Tools Eng. Sci. Discip., pp. 493–510, 2011.
Alternative Environment for Education,” Account. Manag. [59] M. L. García, G. C. Fernandez, E. S. Ruiz, A. P. Martín, and
Inf. Syst., vol. 11, no. 1, pp. 56–75, 2012. M. C. Gil, “Rethinking remote laboratories: Widgets and
[42] L. M. Blaschke, “Using social media to engage and develop smart devices,” Proc. - Front. Educ. Conf. FIE, pp. 782–788,
the online learner in self-determined learning,” Res. Learn. 2013.
Technol., vol. 22, no. 0, Nov. 2014. [60] N. Dlodlo, “The Internet-of-things in remote-controlled
[43] C. S. Fishpaw and M. Ketel, “Online social networking laboratories,” Proc. 13th Annu. Conf. World Wide Web Appl.,
practices and the implications for e-learning solutions,” Conf. no. September, pp. 14–16, 2011.
Proc. - IEEE SOUTHEASTCON, 2014. [61] A. Wilde and E. Zaluska, “Revisiting activity theory within
[44] B. Liu and L. Zhang, “A Survey of Opinion Mining and the Internet of Things. ,” 1st Int. Work. Internet Sci. Web Sci.
Sentiment Analysis,” Min. Text Data, pp. 415–463, 2013. Synerg., pp. 3–6, 2013.
[45] M. Tsytsarau and T. Palpanas, “Survey on mining subjective [62] Ericsson, “Wearable technology AND Internet.” [Online].
data on the web,” Data Min. Knowl. Discov., vol. 24, no. 3, pp. Available:
478–514, 2012. https://www.ericsson.com/res/docs/2016/consumerlab/
[46] J. Bollen and A. Pepe, “Modeling Public Mood and Emotion : wearable-technology-and-the-internet-of-things-ericsson-
Twitter Sentiment and Socio-Economic Phenomena,” Fifth consumerlab-2016.pdf.
Int. AAAI Conf. Weblogs Soc. Media, pp. 450–453, 2011. [63] S. H. Ha, S. M. Bae, and S. C. Park, “Web mining for distance
[47] P. Fornacciari, M. Mordonini, and M. Tomauiolo, “Social education,” Proc. 2000 IEEE Int. Conf. Manag. Innov. Technol.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access
MEHMOOD ET AL.: UTILEARN: A PERSONALISED UBIQUITOUS TEACHING AND LEARNING SYSTEM FOR SMART SOCIETIES 21
, vol. 2, pp. 715–719, 2000. [80] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz,
[64] W. Hämäläinen, J. Suhonen, S. Member, A. Hershkovitz, R. “A Public Domain Dataset for Human Activity Recognition
Nachmias, E. Sutinen, and H. Toivonen, “Data mining in Using Smartphones,” ESANN 2013 proceedings, Eur. Symp.
personalizing distance education courses,” Proc. 1st Int. Conf. Artif. Neural Networks, Comput. Intell. Mach. Learn., 2013.
Educ. Data Min., p. 16, 2004. [81] D. Roggen, A. Calatroni, M. Rossi, T. Holleczek, K. Forster,
[65] L. Wang, “Data mining in individuality long-distance G. Troster, P. Lukowicz, D. Bannach, G. Pirkl, A. Ferscha, J.
education system applied research,” 2nd Int. Symp. Inf. Sci. Doppler, C. Holzmann, M. Kurz, G. Holl, R. Chavarriaga, H.
Eng. ISISE 2009, pp. 99–103, 2010. Sagha, H. Bayati, M. Creatura, and J. R. Millan, “Collecting
[66] N. K. M. Tech and P. Engineering, “Study And Analyze The complex activity datasets in highly rich networked sensor
Factors Related To Distance Education Using Data Mining,” environments,” 2010 Seventh Int. Conf. Networked Sens. Syst.,
Int. J. Eng. Res. Technol., vol. 2, no. 2, pp. 1–5, 2013. 2010.
[67] José Gerardo Moreno Salinas and C. R. Stephens, “Applying [82] O. Banos, M. A. ttila Toth, M. Damas, H. Pomares, and I.
Data Mining Techniques to Identify Success Factors in Rojas, “Dealing with the effects of sensor displacement in
Students Enrolled in Distance Learning: A Case Study José,” wearable activity recognition,” Sensors (Basel)., vol. 14, no. 6,
Adv. Artif. Intell. Its Appl., vol. 9414, pp. 208–219, 2015. pp. 9995–10023, 2014.
[68] S. Pal, “Mining Educational Data Using Classification to [83] J.-L. Reyes-Ortiz, L. Oneto, A. Sama, X. Parra, and D. A. I.-O.
Decrease Dropout Rate of Students,” Int. J. Multidiscip. Sci. http://orcid. org/000.-0002-4943-3021 Anguita, “Transition-
Eng., vol. 3, pp. 35–39, 2012. Aware Human Activity Recognition using smartphones.,”
[69] E. Yukselturk, S. Ozekes, and Y. K. Türel, “Predicting Neurocomputing An Int. J., vol. 171, pp. 754–767, 2016.
Dropout student: An Application of Data Mining Methods [84] W. T. Cho, Y. X. Lai, C. F. Lai, and Y. M. Huang, “Appliance-
in an Online Education Program,” Eur. J. Open, Distance e- aware activity recognition mechanism for iot energy
Learning, vol. 17, no. 1, 2014. management system,” Comput. J., vol. 56, no. 8, pp. 1020–
[70] M. A. Santana, E. B. Costa, and J. B. A. Rego, “A predictive 1033, 2013.
model for identifying students with dropout profiles in [85] D. A. Johnson and M. M. Trivedi, “Driving style recognition
online courses,” CEUR Workshop Proc., 2015. using a smartphone as a sensor platform,” IEEE Conf. Intell.
[71] K. R. Roll, I., Baker, R. S., Aleven, V., McLaren, B. M., & Transp. Syst. Proceedings, ITSC, pp. 1609–1615, 2011.
Koedinger, “Generating Computer-Based Advice in Web- [86] N. Banovic, T. Buzali, F. Chevalier, J. Mankoff, and A. K.
Based,” User Model. Springer Berlin Heidelb., pp. 367–376, Dey, “Modeling and Understanding Human Routine
2005. Behavior,” Proc. 2016 CHI Conf. Hum. Factors Comput. Syst. -
[72] V. Nabiyev, H. Karal, S. Arslan, L. A. K. Erumit, and A. Cebi, CHI ’16, pp. 248–260, 2016.
“An artificial intelligence-based distance education system: [87] P. Romanski, L. Kotthoff, and M. L. Kotthoff, “Package
Artimat,” Turkish Online J. Distance Educ., vol. 14, no. 2, pp. FSelector: Selecting Attributes,” CRAN, p. 18, 2016.
81–98, 2013. [88] A. Candel and V. Parmar, “Deep Learning with H2O Deep
[73] H. Karal, V. Nabiyev, A. K. Erümit, S. Arslan, and A. Çebi, Learning with H2O,” 3rd ed., no. September, H2O.ai, Inc,
“Students’ Opinions on Artificial Intelligence based Distance 2015.
Education System (Artimat),” Procedia - Soc. Behav. Sci., vol. [89] S. Altowaijri, R. Mehmood, and J. Williams, “A Quantitative
136, pp. 549–553, 2014. Model of Grid Systems Performance in Healthcare
[74] J. Gentry, “R Based Twitter Client,” CRAN, 2015. Organisations,” 2010 Int. Conf. Intell. Syst. Model. Simul.,
[75] D. T. Lang, “General Network (HTTP/FTP/...) Client 2010.
Interface for R,” CRAN, 2016. [90] G. Ayres and R. Mehmood, “LocPriS: A Security and Privacy
[76] W. Ugulino, D. Cardador, K. Vega, E. Velloso, R. Milidiú, Preserving Location Based Services Development
and Hugo Fuks, “Advances in artificial intelligence - SBIA Framework,” Knowledge-Based Intell. Inf. Eng. Syst. KES 2010.
2012: 21th Brazilian symposium on artificial intelligence Lect. Notes Comput. Sci., vol. 6279, pp. 566–575, 2010.
Curitiba, Brazil, october 20-25, 2012 proceedings,” Lect. Notes [91] G. Ayres, R. Mehmood, K. Mitchell, and N. J. P. Race,
Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. “Localization to Enhance Security and Services in Wi-Fi
Notes Bioinformatics), vol. 7589, pp. 52–61, 2012. Networks under Privacy Constraints,” Lect. Notes Inst.
[77] P. Casale, O. Pujol, and P. Radeva, “Personalization and user Comput. Sci. Soc. Informatics Telecommun. Eng., vol. 16, pp.
verification in wearable systems using biometric walking 175–188, 2009.
patterns,” Pers. Ubiquitous Comput., pp. 563–580, 2012. [92] A. Alvi, Z. Nabi, D. Greaves, and R. Mehmood, Controlling
[78] A. Reiss and D. Stricker, “Introducing a new benchmarked real world pervasive environments with knowledge bases, vol.
dataset for activity monitoring,” Proc. - Int. Symp. Wearable 6279 LNAI, no. PART 4. 2010.
Comput. ISWC, pp. 108–109, 2012. [93] R. Mehmood, R. Alturki, and S. Zeadally, “Multimedia
[79] A. Stisen, H. Blunck, S. Bhattacharya, T. S. Prentow, M. B. applications over metropolitan area networks (MANs),” J.
Kjærgaard, A. Dey, T. Sonne, and M. M. Jensen, “Smart Netw. Comput. Appl., vol. 34, no. 5, pp. 1518–1529, 2011.
Devices are Different: Assessing and MitigatingMobile [94] R. Alturki, K. Nwizege, R. Mehmood, and M. Faisal, “End to
Sensing Heterogeneities for Activity Recognition,” Proc. 13th end wireless multimedia service modelling over a
ACM Conf. Embed. Networked Sens. Syst. - SenSys ’15, pp. 127– metropolitan area network,” in 11th International Conference
140, 2015. on Computer Modelling and Simulation, UKSim 2009, 2009, pp.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2668840, IEEE Access