• 沒有找到結果。

A personalized learning content adaptation mechanism to meet diverse user needs in mobile learning environments

N/A
N/A
Protected

Academic year: 2021

Share "A personalized learning content adaptation mechanism to meet diverse user needs in mobile learning environments"

Copied!
45
0
0

加載中.... (立即查看全文)

全文

(1)

DOI 10.1007/s11257-010-9094-0

O R I G I NA L PA P E R

A personalized learning content adaptation mechanism

to meet diverse user needs in mobile learning

environments

Jun-Ming Su · Shian-Shyong Tseng · Huan-Yu Lin · Chun-Han Chen

Received: 15 April 2010 / Accepted in revised form: 18 December 2010 / Published online: 28 January 2011

© Springer Science+Business Media B.V. 2011

Abstract With the heterogeneous proliferation of mobile devices, the delivery of learning materials on such devices becomes subject to more and more requirements. Personalized learning content adaptation, therefore, becomes increasingly important to meet the diverse needs imposed by devices, users, usage contexts, and infrastructure. Historical server logs offer a wealth of information on hardware capabilities, learners’ preferences, and network conditions, which can be utilized to respond to a new user request with the personalized learning content created from a previous similar request. In this paper, we propose a Personalized Learning Content Adaptation Mechanism (PLCAM), which applies data mining techniques, including clustering and decision tree approaches, to efficiently manage a large number of historical learners’ requests. The proposed method will intelligently and directly deliver proper personalized learn-ing content with higher fidelity from the Sharable Content Object Reference Model (SCORM)-compliant Learning Object Repository (LOR) by means of the proposed adaptation decision and content synthesis processes. Furthermore, the experimental results indicate that it is efficient and is expected to prove beneficial to learners.

J.-M. Su (

B

)

Department of Information and Learning Technology, National University of Tainan, Tainan 700, Taiwan e-mail: junming.su@gmail.com

S.-S. Tseng

Department of Applied Informatics and Multimedia, Asia University, Taichung 413, Taiwan e-mail: sstseng@asia.edu.tw

H.-Y. Lin· C.-H. Chen

Department of Computer Science, National Chiao Tung University, Hsinchu 300, Taiwan e-mail: linhuanyu@gmail.com

C.-H. Chen

(2)

Keywords Personalized learning content· Content adaptation · Mobile learning environment· Data mining · Learning object repository

1 Introduction

The proliferation of e-learning has been sparked by the rapid development of the Inter-net. This educational approach has become more and more popular because it allows learners to conveniently study anytime, anywhere; no longer are students restricted to the confines of a classroom. Due to the emergence of heterogeneous mobile devices and wireless technologies, such as cell/smartphones, Personal Digital Assistants (PDAs), netbooks, and other similar mechanisms, the requirements for delivering and display-ing learndisplay-ing content on mobile devices are increasdisplay-ing rapidly (Tretiakov and Kinshuk

2004;Yang et al. 2007a;Chang et al. 2008;Queirós and Pinto 2009;Pettersson and

Gil 2010). While the types of mobile devices and the available e-learning subject

mat-ter proliferate, a large gap still exists between traditional desktop and mobile-based systems. That is, most existing learning content is created for desktop devices, such as Personal Computers (PCs), and not for the mobile devices currently used. Conse-quently, the insufficient hardware capabilities, e.g., the limited computation power, memory, screen size, and unstable bandwidth of a wireless network, have led to poor navigation experience and unfavorable presentation of learning content. In many cases, it is difficult to display the original content, which is customized to a desktop com-puter and a larger monitor, on the display of a handheld device in a usable and efficient manner. How learning content is adapted to meet the needs of various users via various mobile devices needs to be addressed and improved.

This leads to the issue of content adaptation, which concerns the act of transforming learning content to adapt to any mobile device capabilities. How to successfully trans-form existing learning content into a suitable version that can be efficiently delivered and displayed to meet the needs of learners and of mobile devices has been a growing area of study. In order to solve this problem, a significant amount of research has been conducted that proposes numerous content adaptation approaches. Two notable approaches include: (1) static adaptation, which preprocesses and stores multiple ver-sions of the content; and (2) dynamic adaptation, which adapts the content real-time during the user’s request (Fudzee and Abawajy 2008).

The static adaptation approach is capable of reducing the time it takes for learning content to download, but it requires a preprocessing task and larger storage allocation

(Mohan et al. 1999;Villard et al. 2000;Hinz et al. 2004). In contrast, the dynamic

adaptation approach can apply content structure analysis (Buyukkokten et al. 2001,

2002;González-Castaño et al. 2002;Chen et al. 2003;Yin and Lee 2004;Ramaswamy

et al. 2005;Laakko and Hiltunen 2005;Yang et al. 2007b;He et al. 2007;Kim et al.

2008;Hsiao et al. 2008) to dynamically transcode, rearrange the layout, and distillate

the content to improve the delivery latency and meet the devices’ capabilities. Alter-natively, it can apply context-based adaptation (Lum and Lau 2003;Mohomed et al.

2004,2006a,b,2007; Lee et al. 2006;Yang et al. 2008) to consider both the

envi-ronmental context and the preference of the requesting user to offer more precisely adapted content.

(3)

However, the content quality of mobile devices is limited to fixed page fragments, which are compatible with content adaptation processing approaches. Furthermore, the information of content, e.g., text and semantic relations, may be lost during the decomposition or summarization processes. Based on our review of the literature

(Beekhoven et al. 2003,Chen and Mizoguchi 1999;Dewhurst et al. 2000;Chen et al.

2000; Smith 2001; McIlroy et al. 2001; Riding and Cheema 1991; Wilson 2000;

Franzoni et al. 2008), different students have different learning approaches and

pref-erences for learning. That is, not all students are the same, nor are their prefpref-erences for learning. Students, therefore, favor various ways to receive information and acquire knowledge, such as: content type preference (Gilbert and Han 1999;Stern and Woolf 2000), presentation style, cognitive styles (Riding and Cheema 1991), and learning style (Kolb 1976,2004), among others.

In mobile learning environments, the response time of requested information can significantly impact and influence learning performance. With the proliferation of mobile device usage for educational purposes, a learner must be able to retrieve requested information with few hindrances. For example, although a specific request for learning content has been appropriately tailored for the learner, if this content is delivered too slowly, or if the overall quality is poor, it will explicitly affect the learner’s satisfaction (Muntean 2008;Ding et al. 2010). Accordingly, a given learner’s satisfaction will be affected not only by his or her perception and preference, but also by the network-related issues and mobile device factors (Muntean 2008). Most existing research and systems either separately or partially considered the features of mobile devices, network conditions, and learner preferences (Zhao et al. 2008;

Nimmagadda et al. 2010). Therefore, a pressing issue is how to consider not only the

mobile devices’ capabilities, but also the available network conditions and learners’ diverse preferences in order to efficiently provide properly personalized learning con-tent in mobile learning environments (Tong et al. 2006;Mylonas et al. 2007;Basaeed

et al. 2007;Zhao et al. 2008;Franzoni et al. 2008;Muntean 2008;Pettersson and Gil

2010;Nimmagadda et al. 2010).

The historical records of a given learner, including hardware capabilities, various learning preferences, and the current situation of wireless networks, can help solve the problems arising from variant mobile device capabilities, vacillating network con-ditions, and learners’ diverse preferences. The proposed resolution is a concept that focuses on historical learners’ requests to provide successful new requests, which share similar preference attributes. That is, if we provide a new request with the per-sonalized learning content created from a previous similar request, not only will the performance of content delivery improve, but the learner’s overall satisfaction will be greatly enhanced as well. This is complicated by the fact that there are vast amounts of educational data and information available for e-learners. In an e-learning system, teaching materials are usually stored in a Learning Object Repository (LOR), which can be formatted based on one of the most popular standards on the e-learning system, Sharable Content Object Reference Model (SCORM) (2010). In an LOR, a substantial amount of teaching materials, including related learning objects, will result in man-agement issues over wired/wireless environments (Ko and Choy 2002; Wong et al.

(4)

Therefore, in this paper, our main concern is how to efficiently manage a large num-ber of historical learners’ requests while delivering learning content that is precisely tailored to meet the individual needs of each user and to fall within the scope of various mobile device requirements. Successful management of similar content and delivery of that information to various mobile devices can be accomplished by providing learners with adapted content created from a similar, previously processed request. To achieve this end goal, the following must be realized:

(1) Model and define constructive and meaningful data in a format that is tailored to meet the individual needs and mobile requirements of each user.

(2) Group several historical learners’ requests, which share similar preference attri-butes, into the same cluster to construct a useful relationship among them. (3) Determine high-fidelity, personalized learning content quickly and efficiently for

a new request based on the analysis of historical learners’ requests.

While taking diverse learners’ needs, wireless network conditions, as well as mobile device capabilities and constraints into consideration, a Personalized Learning Con-tent Adaptation Mechanism (PLCAM) has been proposed. The proposed PLCAM can efficiently manage a large number of historical learners’ requests, and intelligently deliver suitably personalized learning content, with higher fidelity, from a Learning Object Repository (LOR) directly to the learner. Subsequently, an adapted, or mod-ified, version of the learning content would be prepared for the next similar request. Thus, the PLCAM is composed of two distinct phases: (1) the Adaptation Data For-mat Definition Phase and (2) the Personalized Learning Content Delivery Phase. The former phase defines an adaptation data format based on Composite Capabilities/Pref-erence Profiles (CC/PP) (2010) and User Agent Profile (UAProf) (2010). These profiles identify constructive information about the requesting learner, applicable hardware, network capabilities, and media during the learning content adaptation process. The latter phase applies distance-based clustering and decision tree approaches to create a working relationship among a cache of historical learners’ requests. At this stage, the proposed adaptation decision and content synthesis processes can resourcefully identify and organize a precise and suitable version of the learning content to meet the specific request.

Furthermore, the framework of the PLCAM can be extended to address many more diverse user needs and to enhance the effectiveness of the adapted content composition. In an effort to evaluate our proposed approach, a prototypical system of the PLCAM, based on a Sharable Content Object Reference Model (SCORM)-compliant LOR has been developed, and experiments have also been performed. The experimental results show that the PLCAM is efficient and can be expected to benefit learners.

2 Related work

Diverse content adaptation approaches have been proposed to render learning con-tent on the mobile devices.Fudzee and Abawajy(2008) grouped content adaptation approaches into two basic types: static adaptation and dynamic adaptation. The former usually generates multiple variants for each content component, attaching a layout

(5)

description for the presentation of component-based Web content (Mohan et al. 1999;

Villard et al. 2000;Hinz et al. 2004). These static adaptation approaches can reduce

download time, but they require preprocessing tasks and greater storage allocation. Another limitation is that they do not take into account the user’s preference and the situation of the wireless network.

Many dynamic adaptation approaches, including content structure analysis and context-based adaptations, have been proposed to resolve these issues (Fudzee and

Abawajy 2008). A Hierarchical Atomic Navigation Concept (HANd) was proposed

byGonzález-Castaño et al.(2002) to navigate on small-scale devices, using the

con-tent structure analysis approach. In the HANd approach, an automatically generated navigator page is used to indicate some or all elements embedded in a World Wide Web (WWW) page. To generate the navigator page, a Web page must be analyzed and fragmented into several separate “clipped” versions, which can be delivered to a small-scale device, according to an importance value levied for every page fragment. Based on a similar concept, many fragmentation and summarization processes have been proposed to organize a Web page into a thumbnail representation that indexes detailed information (Chen et al. 2003), breaks each Web page into several text units

(Buyukkokten et al. 2001,2002), and detects the important parts (Yin and Lee 2004)

or the interesting fragments in dynamic Web pages (Ramaswamy et al. 2005), thus reducing delivery latency.

However, not all Web pages are suitable for text summarization because summa-rized statements, as lossy information, may mislead users. To help improve understand-ing, the semantically coherent perceivable units of the Web content can be extracted and presented together on a mobile device according to their semantic relationships

(Laakko and Hiltunen 2005;Yang et al. 2007b;Kim et al. 2008). However, most of the

aforementioned content adaptation approaches do not consider flexible and extensible content adaptation. Recognizing this,He et al.(2007) applied the rule-based approach to propose a flexible adaptation system, Xadaptor, for managing the new content type by adding its corresponding content adaptation rule. Similarly, a Versatile Transcoding Proxy (VTP) (Hsiao et al. 2008) executes the transcoding preference script provided by the client or server to transform the corresponding data or protocol according to the user’s specification based on CC/PP (2010). Although most of the previously men-tioned content structure analysis approaches can improve delivery latency, the quality of the content shown on mobile devices is limited to fixed page fragments, which are compatible with content adaptation processing approaches. These approaches, how-ever, may experience loss of information, e.g., text and semantic relations, during the decomposition or summarization processes, and environmental context and user’s preferences cannot yet be taken into consideration.

To consider the user’s preference in the context-based adaptation approach,Lum

and Lau(2003) proposed a decision engine, which can determine automatically an

appropriate content adaptation version based on the Quality of Service-sensitive (QoS) approach. A typical score tree with several score nodes is created to evaluate the QoS of the content versions in various quality domains. Because they assume that user preferences and the adaptation process are independent of content, the score tree can be established during the preprocess phase. This predefined score tree with limited content versions may not meet the user’s needs and may constrain the flexibility and

(6)

extensibility of the system. Yang et al. (2008) took the learner’s environmental context into account, proposing a context-driven content adaptation planner. This planner can determine a proper version of content created by static or dynamic adaptation, based on the predefined context profile to meet a learner’s context status (e.g., brighten the background light in the outdoor environment). However, efficiently selecting and deciding upon an appropriate adapted version of content from previous requests is still an important and under-evaluated issue.

To take advantage of similar navigational behaviors,Mohomed et al.(2004) pro-posed an automatic content adaptation approach based on the community-driven con-cept, which artificially groups users into communities according to some similar char-acteristics and assumes that users in the same community have similar adaptation preferences. This system can learn the common adaptation preference of one com-munity based on users’ feedback. Later, a feedback-driven context selection approach

(Mohomed et al. 2006b) was proposed to leverage user interaction to automatically

split users into several groups according to feedback from their previous research sys-tem, URICA (Mohomed et al. 2006a). This system adapts content for mobile devices based on usage semantics. Using the same concept and detecting correlations in the adaptation requirements of past users, they initially applied the standard K-means clustering algorithm to partition users into multiple groups. They proposed the use of an online classification algorithm to make predictions that allowed rapid classification of each user into a single cluster, defined as suitable for offering a specific adapted content version to users (Mohomed et al. 2007). Their experimental results showed that bandwidth consumption and browsing time can be significantly reduced. How-ever, this approach clusters users according to the fidelity of accessed image objects and must refine the prediction of adapted content by means of the user’s interaction and feedback. According to existing research (Muntean 2008;Ding et al. 2010), the long waiting and response time can explicitly affect a user’s satisfaction, often leading to negative feedback. Therefore, if the desired adapted content is received by several interactions, a poor navigation experience may result due to the long waiting time.Lee

et al.(2006) also proposed an intelligent adaptation system to reduce response time

by classifying users into basic categories. This process can be adjusted according to the users’ feedback; content generated by the same group category can be reused.

However, in learning environments, the learners’ preferences vary greatly because each learner is an individual student, and thus, has his or her individualized learn-ing behaviors and preferences. The variety of behaviors and preferences can include content type preferences such as text, picture, audio, video, or a hybrid of those types

(Gilbert and Han 1999;Stern and Woolf 2000). Learners also differ in their inclinations

for how learning material is presented and how they intellectually receive and digest that material. That is, learners vary in their preferences of presentation styles, cogni-tive styles (field dependence/independence) (Riding and Cheema 1991), and learning styles (such as whether they are “doers,” “watchers,” “thinkers,” or “feelers”) (Kolb 1976,2004). Students will achieve higher learning performance if the learning content can be customized and offered according to their diverse learning needs (Beekhoven

et al. 2003;Chen and Mizoguchi 1999;Dewhurst et al. 2000;Chen et al. 2000;Smith

2001;McIlroy et al. 2001;Riding and Cheema 1991;Wilson 2000;Su et al. 2005).

(7)

predefined adaptation rules to adapt the learning content to meet learning prefer-ence and environmental conditions.Franzoni et al. (2008) proposed learning styles integration taxonomy to facilitate and personalize the learning process. Nevertheless, the predefined approach is time consuming and to the efficient use of the adaptation rule is still problematic. In addition, response time of the content adaptation system is an important factor that affects the user’s satisfaction because it measures waiting time (Ding et al. 2010).Muntean(2008) proposed a learner Quality of Experience (QoE) model in an educational area derived from delivery performance-based content personalization to investigate the influence of delivery latency on user experience and learning performance. The experimental results demonstrate the significant benefits in learning achievement, performance, and satisfaction when learners are offered supe-rior content quality and delivery performance. In other words, learners’ satisfaction and performance are affected not only by their perceptions and preferences, but also by network conditions and mobile device factors in mobile learning environments. Nevertheless, most existing research and systems either separately or partially con-sidered the features of mobile devices, network conditions, and learner preferences

(Zhao et al. 2008). Consequently, they may not be suitable for efficient management

of diverse user needs in mobile learning environments. 3 Personalized learning content adaptation mechanism

3.1 Framework of the personalized learning content adaptation mechanism

This paper proposes a PLCAM, to address the increasing number and diversity of user requests. The architecture of the PLCAM is shown in Fig.1. The PLCAM can manage efficiently a large number of historical learners’ requests and intelligently deliver proper personalized learning content with higher fidelity from a Learning Object Repository (LOR) directly to the learner. An adapted content version can then be prepared for the next similar request.

The PLCAM includes two phases, as described below:

1. Adaptation Data Format Definition Phase: First, we define the adaptation data format including Learner Preference (LP), Hardware Profile (HP), and Media

(8)

Parameter (MP) based upon CC/PP (2010) and UAProf (2010). Here, the LP describes the diverse needs of a learner, e.g., the desired maximum deliv-ery time, image format, presentation style, the ratio of audio to picture, etc. The HP describes the hardware capabilities of a learner’s mobile device, e.g., the device type, screen size, etc. The MP describes how to adapt and transcode the multimedia, e.g., image, audio, and video, within requested learning objects. 2. Personalized Learning Content Delivery Phase: To deliver the suitable learning content with associated learning resources efficiently to learners in accordance with their preferences, their device’s hardware capabilities, and variable wireless bandwidth, we propose the following three modules.

- Learning Content Adaptation Management Scheme (LCAMS): To manage the learners’ historical request data efficiently, we first apply the distance-based clustering approach to group the historical learners’ requests into several sets, according to Learner Preference (LP). After the initial clustering phase, every cluster created with a similar LP will be tagged with a cluster label. The Hard-ware Profiles (HP) within the historical requests with corresponding cluster labels are used as training data to create a decision tree, called Content Adap-tation Decision Tree (CADT). The CADT can be used to efficiently determine the appropriate adapted content.

- Adaptation Decision Process (ADP): To determine a precise and proper adapted content version based on CADT, we propose an ADP Algorithm, called ADPAlgo, which can determine a suitable version of the existing adapted con-tent.

- Learning Content Synthesizer (LCS): According to the results of ADPAlgo, the LCS will use the identified adaptation parameters to transcode the content if necessary.

The details of each phase are described in the following subsections. 3.2 Adaptation data format definition

To manage existing user requests efficiently, we must model and define a data for-mat including the LP, HP, network condition, and MP, which will be recorded in a database to represent every learner’s request based uponCC/PP(2010) andUAProf

(2010) (Kobsa 2007). Thus, a Content Adaptation Rule (CAR) is defined to represent

a processed learner request transaction in the PLCAM. Definition 1 Content Adaptation Rule (CAR)

CAR= (LOi, pj, (B, HP, LP), MPset), where:

- LOi: the ith learning object in the SCORM-compliant LOR.

- pj = {r1, r2, . . . , rm}: the jthpage of LOi, i.e., (LOi j) consists of several

associ-ated learning resources (r).

- B: the bandwidth of the network condition in a mobile learning environment during the learner request.

- HP = a1, a2, . . . , an : every attribute (a) denotes a specific capability of a mobile device, e.g., the machine type (PDA or smartphone), Central Processing Unit (CPU) speed, memory capacity, screen size, sound rate, etc.

(9)

- LP= b1, b2, . . . , bk: every attribute (b) denotes a learner’s specific requirement, e.g., maximum delivery time, preferred picture format ordering, preferred audio property, media switch, preferred content type, cognitive style, learning style, etc. Thus, in this paper, we can initially define the LP= Delivery Time (DT), Pre-ferred Picture Format Ordering (PPFO), Picture Switch (PS), Audio Switch (AS) described in Example 1. The preferred content type, cognitive style, and learn-ing style can also be extended into LP definition if necessary, and the details are described in Sect.6.3, The Extensibility of PLCAM.

- MPset={MP1, MP2, . . . , MPk}: denotes all associated Media Parameter (MP) used to adapt and transcode all physical media resources belonging to a page. Here, MP= Version (V), Type (T), Attribute (A), Size (S), TP, where TP is the Transcoding Parameters used by the existing transcoding tools or approaches, e.g.,

ImageMagicK(2010) andSoX(2010), to transform the original media object (V0)

into adapted media version (Vi) with type (T), e.g., image, audio, video,

asso-ciated with media attributes (A), e.g., image= width, height, color depth, type, audio= precision, rate, video = width, height, sound precision, sound rate, and file size (S). In the PLCAM, the media transcoding process will be triggered if there is no suitable adapted content version for a new learner’s request.

Example 1 A CAR of a given learner’s request transaction can be recorded as ((LOi, pj), (80, HP, LP), MPset), where HP = 1,400, 128, 480, 640, 16, 16, 44

denotes that a learner uses a PDA (1) with 400 Mhz, 128 MB, 480× 640 resolution, 16 bits color depth, 16 bits sound precision and 44 KHz sound rate (U denotes Unsup-ported) under 80 kbps bandwidth (B) to retrieve the page (pj) of LOi, and LP = 5,

JPGB, 1, 0 denotes that the maximum delivery time (DT) is less equal than 5 sec-onds (s), the order of preferred picture format (PPFO) is JPG (J) > PNG (P) > GIF (G) > BMP (B), the switch attribute of media, PS= 1, enables to show the picture, and the AS= 0 disables the audio play, respectively. Then, the PLCAM is able to use the MPs in MPset, which are selected according to the B, HP, and LP, to transcode the

physical resources in the pj. Table1shows the example with 15 CARs for the pj in LOi, i.e., LOi j. The attribute definitions of LP and HP can be extended to meet the

various requirements.

4 Learning content adaptation management scheme

In this section, we will describe how to use existing CARs to construct a Content Adaptation Decision Tree (CADT) in the Learning Content Adaptation Management Scheme (LCAMS). The CADT can be used to efficiently and quickly determine the suitable adapted content in an LOR for learners according to the mobile device features, the preferences of learners, and network bandwidth. As shown in Fig. 2, the LCAMS includes three processes to construct the CADT: (1) clustering process; (2) decision tree construction; and (3) CADT maintenance process.

4.1 Clustering process of the learning content adaptation management scheme As mentioned in Sect.3, each learner’s request, with his/her preference(s) logged as a transaction, can be represented by a Content Adaptation Rule (CAR). As shown in

(10)

Table 1 Example of CAR for the same page (pj) in an LO (LOi), (LOi j)

ID Bandwidth (B) Hardware profile (HP) Learner preference (LP) 1 213 2, 528, 384, 320, 480, 16, 32, 120 7, JBGP, 1, 1 2 175 2, 600, 384, 320, 480, 24, 16, 40 2, GPBJ, 0, 1 3 487 0, 1200, 4000, 1366, 768, 32, 16, 300 2, JGPB, 1, 0 4 223 2, 528, 288, 320, 480, 16, 8, 30 3, JBPG, 1, 0 5 281 2, 528, 384, 320, 480, 16, 32, 120 7, PJGB, 0, 1 6 69 0, 2000, 8000, 1366, 768, 32, 32, 500 1, GPBJ, 1, 0 7 232 2, 528, 288, 320, 480, 16, 8, 30 5, JPGB, 0, 1 8 290 1, 1000, 448, 480, 800, 24, 16, 140 1, GPJB, 1, 0 9 95 0, 1200, 4000, 1366, 768, 32, 16, 300 1, BJGP, 0, 0 10 167 0, 1200, 4000, 1366, 768, 32, 16, 300 5, GPJB, 1, 0 11 220 0, 1200, 4000, 1366, 768, 32, 16, 300 5, JPGB, 0, 1 12 326 2, 528, 288, 320, 480, 16, 8, 30 4, JPGB, 0, 0 13 339 2, 528, 288, 320, 480, 32, 8, 20 7, PBGJ, 0, 0 14 313 2, 528, 288, 320, 480, 32, 8, 20 4, GJPB, 1, 1 15 95 2, 528, 384, 320, 480, 16, 32, 120 4, PBJG, 0, 0

Fig. 2 Process of LCAMS

Fig.2, in the LCAMS, all new CARs requested by learners will be stored temporarily in a CAR Pool. We can apply the distance-based clustering algorithm to group these historical CARs into several clusters according to learners’ preferences, where every learner in the same cluster shares similar LPs (Tan et al. 2005;Romero and Ventura 2006,2010). However, it is difficult to determine the number of clusters while apply-ing the clusterapply-ing approach. To resolve this problem and equip the PLCAM with the automatic maintenance process, the renowned clustering algorithm, ISODATA (Hall

and Ball 1965), can be employed. This process can dynamically change the number

of clusters by lumping and splitting procedures and by iteratively changing the num-ber of clusters to produce better results. The ISODATA clustering approach has been used successfully in many applications, such as image processing (Rahimi et al. 2007;

Chen et al. 2009), data and document classification (Zhu et al. 2007), and so on. In

this study, we apply the ISODATA clustering approach to group CARs into different clusters automatically.

(11)

4.1.1 Similarity measure of the clustering process

To apply the ISODATA clustering approach, a similarity measure estimating the sim-ilarity value between two CARs based on the LP must be determined. Because the attribute of an LP might consist of a numerical attribute, e.g., maximum delivery time, and a symbolic attribute, e.g., preferred picture format ordering, as described in Example 1, the similarity measure of an LP can be formulized by means of the distance measure approach as follows:

Given two LPi=a1, a2, . . . , an and LPj=b1, b2, . . . , bn, the similarity measure of numerical attribute can be formulized as follows:

SimofNumk= 1 −

|ak− bk|

Maxk− Mink,

where 1<=k<=n, the Maxkand the Mink are the predefined maximum and minimum

values of kthattribute in an LP, respectively.

Regarding the symbolic attribute in an LP, the value, JPGB, is like a string. To calculate the similarity between two symbolic attributes, their string-based values can be encoded into a numerical value by the numerical order of predefined symbol pri-ority. For instance, the numerical value of string JPGB can be encoded as ‘1234’ based on the priority order definition {“J”= 1, “P” = 2, “G” = 3, “B” = 4}. Also, the maximum value will be the ‘4321’ of the string BGPJ. The symbolic attribute can thus be transformed into a numerical value and its similarity can be measured by the SimofNumkformula. Because attributes in an LP may have different degrees of

impor-tance, we define a Weight Vector (WV), which can be manually defined by a learner, to adjust for the degree importance of each attribute. Therefore, the similarity measure between two LPs can be formulated as:

Si milar i t yL P  LPi, LPj  = SimofNumk(ak, bk) × wk  , wherewk ∈ W V, wk = 1, and 1<=k<=n.

To evaluate when to split and merge the cluster, the DeviationL P, which is used to

calculate the standard deviation of the samples, must be defined as: DeviationL Pk = | ak− bk

Maxk− Mink|, ,

where1<=k<=n, the Maxkand the Mink are the predefined maximum and minimum

values of kthattribute in an LP, respectively.

Example 2 Given two LPs, LP1=3, JPGB, 1, 0 and LP2=2, JGBP, 1, 1, and a learner predefined related attribute WV= 0.5, 0.3, 0.1, 0.1. We can apply the above similarity measure to calculate the similarity between LP1and LP2. For example, the similarity of the numerical attribute, Delivery Time (DT), between LP1and LP2is:

SimofNum1= 1 − |a1− b1| Max− Min = 1 −  |3 − 2| 5− 1  = 1 −1 4 = 0.75.

(12)

The similarity of the symbolic attribute, Preferred Picture Format Ordering (PPFO), is: SimofNum2= 1 −  |1234 − 1423| 4321− 1234  = 1 − 189 3087 = 0.938.

Hence, by the same way, the similarity between LP1 and LP2is the Similarity

L P(LP1, LP2) =0.75 × 0.5 + 0.938 × 0.3 + (1 − (|1 − 1|/(1 − 0))) × 0.1 + (1 −

(|0 − 1|/(1 − 0))) × 0.1=0.7564. Besides, the DeviationL P

1 = 3−2 5−1   = 0.25. Algorithm 1 Algorithm 1: LP Clustering Algorithm (LPCALgo) Symbols Definition:

DT: the Delivery Time (DT) in a learner preference vectors (LP). LPset: the set of LP.

K: the initial number of clusters.

C: a cluster with several learner preference vectors (LP). CC: the Center of Cluster.

Cset:the set of clusters with the Center of Cluster (CC)

Ts:the split threshold (Standard Deviation) for splitting a cluster into two ones.

Tm:the merge threshold (Mean Distance) for merging two clusters into one.

Tn:the minimum number of the members in a Cluster for deleting a cluster.

Ti: the maximum iteration number for executing the clustering process

Tp:the minimum number of Cluster pair for merging clusters process.

Input: LPset, K, Ts, Tm, Tn.

Output: The set of Clusters, Cset.

Step 1: Initial Clusters Selection: Step 1.1: For i = 1 to K.

(1) random select LPi ∈ LPset to insert LPi into Ci with CCi= LPi and

then insert Ciinto Cset.

Step 2: ISODATA Clustering Process:

Step 2.1: Execute the following sub-Steps (2.2–2.6) repeatedly until there is no difference between two iterations or exceed the Ti.

Step 2.2: Insert each LPj ∈ LPsetinto appropriate cluster Ci ∈ Cset

according to the SimilarityL P(CCi, LPj).

Step 2.3: Delete the Ciif the number of LP is less than Tn.

Step 2.4: Split a Ci into two clusters according to the Ts and Tn.

Step 2.5: Merge two clusters into one according to the Tmand Tp.

Step 2.6: Re-compute the Cluster Center (CCi) for each Ci ∈ Cset.

Step 3: Output the Cset.

4.1.2 Clustering algorithm based on ISODAT

An LP Clustering Algorithm (LPCALgo) based on ISODATA is proposed to group these LPs into several clusters according to the aforementioned similarity and deviation measure, shown in Algorithm 1. After applying the LPCALgo, the CARs in Table1can be grouped into three clusters, as depicted in Table2.

(13)

Table 2 Result of applying LP

clustering algorithm (LPCALgo) with the cluster parameters (K= 5, Ts= 0.01, Tm= 1.0, Tn= 1,

Ti= 50, Tp= 1) based on data

in Table1

Cluster label ID of CAR

1 {3, 4, 6, 8, 10}

2 {2, 5, 7, 9, 11, 12, 13, 15}

3 {1, 14}

Table 3 Result of mapping the numerical value in HP

Numerical attribute Representative symbol CPU speed (CPU) L:low, M: medium, H: high

System memory (SM) L: low, LM: low-medium, MH: medium-high, H: high,

Screen horizontal size (SHS) T: tiny, S: small, M: medium, L: large

Screen vertical size (SVS) T: tiny, S: small, M: medium, L: large

4.2 Decision tree construction

After the clustering process, each cluster will be tagged with a label, as shown in Table2. Determining a suitable cluster for a new request is an issue which can be resolved by using the decision tree approach. Based on the Hardware Profiles (HPs) in these CARs, with cluster labels defined in Table2, we can apply a decision tree induction algorithm, ID3 (Quinlan 1986), to create a decision tree, called CADT. ID3 can process only the symbolic value of an attribute, so the numerical attribute values of the HP in Table1, e.g., CPU speed, system memory, etc., have to be discretized by the following approach.

In all HPs, and μ are the minimal and maximal values of an attribute, respectively. Let = ( − μ)/N,where N is the number of desired discrete ranges. Then, a numeric value of an attribute can be mapped into the symbolic value. For example, given N is 3, the corresponding symbolic values are L in [,  + ] , M in [ + ,  + 2] , and H in [ + 2,  + 3] .

Therefore, the numerical attribute of HP in Table1 can be mapped into several discrete ranges, as shown in Table3.

In ID3, the information gain measure is used to select the test attribute at each node in the decision tree. The attribute with the highest information gain is chosen as the test attribute for the given set (Quinlan 1986). The following example describes how to apply ID3 to create the CADT of the PLCAM based on the HPs in the CARs with cluster labels.

Example 3 Table 4 shows six HP data with an ID and cluster label, which have been classified into two subsets: {4, 7, 12} and {5, 15, 1} according to the attri-bute, “Sound Precision.” The expected information needed to classify six samples is given by I (the number of CARs in C1, the number of CARs in C2, the number of CARs in C3) = I (1, 4, 1) = −  1 1+4+1  log2  1 1+4+1  − 4 1+4+1  log2  4 1+4+1  −  1 1+4+1  log2  1 1+4+1  = 1.252.

(14)

Table 4 HP in CAR data with

ID and cluster label classified by attribute, “Sound Precision”

ID HP in CAR Cluster label

4 2, 528, 288, 320, 480, 16, 8, 30 1 7 2, 528, 288, 320, 480, 16, 8, 30 2 12 2, 528, 288, 320, 480, 16, 8, 30 2 5 2, 528, 384, 320, 480, 16, 32, 120 2 15 2, 528, 384, 320, 480, 16, 32, 120 2 1 2, 528, 384, 320, 480, 16, 32, 120 3

The Entropy (E), or expected information based on the partitioning into two subsets by the attribute, “Sound Precision,” is given by:

E(Sound Precision) = (1 + 4 + 1)(1 + 2) I(1, 2) + (2 + 1) (1 + 4 + 1)I(2, 1) = 3 6  −1 3log2 1 3 − 2 3log2 2 3  +3 6  −2 3log2 2 3 − 1 3log2 1 3  = 0.918.

Finally, the encoding information that would be gained by branching on attribute “Sound Precision” is Gain(Sound Precision)= I(1, 4, 1) − E(Sound Precision) = 1.252

− 0.918 = 0.334.

Consequently, by means of the above ID3 approach, the information gain of each attribute (of each HP) in Table4 will be computed. The attribute with the highest information gain will be chosen as the test attribute. A node is created and labeled with the attribute, branches are created for each value of the attribute, and the samples are partitioned accordingly. Figure3depicts the result of applying the ID3 algorithm data given in Tables1, 2, and 3.

4.3 Content adaptation decision tree maintenance process

As stated previously, after the clustering and decision tree construction processes are complete in the LCAMS, all CARs in the CAR Pool, a temporary buffer, can be grouped into several clusters and retrieved by the CADT structure. In the CADT maintenance process (see Fig.4), all new CARs are first temporarily stored in a CAR Pool. While the amount of CARs (N) in a CAR Pool is more than a threshold, which is estimated automatically by the CADT Rebuilding Equation (Y= α + βX) generated by the ordinary least squares approach and described in Sect. 6, the LCAMS will rebuild the CADT automatically offline by the clustering and decision tree processes. Then, these processed CARs in the CAR Pool will be shifted to the final storage and become the historical CARs indicated by the newly rebuilt CADT structure. Each CAR indicates the associated media objects consisting of original (V0) or adapted versions

(15)

Fig. 3 CADT based on the HP in Table1

Fig. 4 Flowchart of the CADT maintenance process

(Vi|i > 0), all of which are stored in the LOR. Moreover, in order to efficiently

manage the storage space of the LOR, the LCAMS will check the Utilization Rate (UR) of every adapted media object version, except for its original version. If the UR of any adapted version (Mi j) < Threshold, it will be deleted from the LOR.

5 Personalized learning content delivering process 5.1 Content adaptation process

To meet diverse learner needs, including varied mobile device capabilities, network conditions, and individual learner preferences, the Content Adaptation Process (CAP) has been proposed to automatically determine an appropriate MPsetfrom the Media

Parameter (MP) database to adapt and transcode all media resources in a desired page according to the requirement of the new Learner Request (LR). The process is described below.

(16)

5.1.1 Satisfaction measure on the quality of media

In the CAP, we would like to determine an adapted media which can meet the require-ment of an LR very well. Therefore, the satisfaction measure on the quality of media has been defined to estimate the satisfaction degree between the adapted media selected by CAP and the media requested by the learner.

Given HP= a1, a2,…, am and MPi=b1, b2,…, bn, the similarity measure of each numerical attribute between HP and MPi can be formulized as:

SimofNum(HP, MPi) = Max



1−|aj − bk| aj , 0



, where1<=j<=m, 1<=k<=n.

Regarding the symbolic attribute of the image type between the MP and the requested LP, e.g., Preferred Picture Format Ordering (PPFO), the particular similarity measure of image type is formulized as:

SimofImageT ype(PPFO ∈ LP, Type ∈ MPi) = 1 − (k − 1) × 0.25,

where k= the order of type in the string of PPFO.

We also define a Satisfaction Weight Vector (SWV) to adjust the degree of importance. The satisfaction measure on the quality of media between an LR and an MP in the media database can thus be formulated as:

Sati s f acti onQualityOfMedia(LR, MPi)

= ((SimofNum(aj ∈ HP, bk ∈ MP)|SimofImageT ype (PPFO ∈ LP, Type ∈ MP)) × wj),

wherewj ∈ SW V, wj = 1, and 1  j  m, 1  k  n.

Example 4 Given MPa= Version (V), Type(T), Attribute(A), Size(S), TP = v1,

image, 800, 400, 16, J, 160, TP with SMV = 0.5, 0.1, 0.3, 0.1, and a new LR= ((LOi, pj),(B, HP, LP)) = ((LO1, p1), (500, 1, 1000, 448, 480, 800, 24, 16, 140,1, GPJB, 1, 0)). Then, the satisfaction between LR and MPacan be estimated as follows:

Sati s f acti onQualityOfMedia(LR, MPa) = 0.5 × SimofNum(480 ∈ HP, 800) +0.1 × SimofNum(800 ∈ HP, 400) + 0.3 × SimofNum(24 ∈ HP, 16) + 0.1 ×SimofImageT ype(“GPJB” ∈ LP, ‘J’)=0.5 × Max (0.33, 0) +0.1 × Max (0.5, 0)

+0.3 × Max (0.66, 0) + 0.1 × (1 − (3 − 1) × 0.25) = 0.165 + 0.05 + 0.198 + 0.05 = 0.463.

5.1.2 Satisfaction score of the media parameter

By means of the SatisfactionQualityOfMedia(LR, MPi), we can understand which

(17)

mentioned in Sect. 1, the response time to LRs will explicitly affect learner satis-faction (Muntean 2008;Ding et al. 2010). Accordingly, we take the response time of the PLCAM into account and define the satisfaction score of the MP to estimate the satisfaction degree of applying the MP to adapt the Media Object (MO), whereby the most appropriate MPset={MP1, MP2,…, MPk} can be determined by the CAP.

The definition of the satisfaction score is as follows and the CAP algorithm is described in Algorithm 2:

Sati s f acti on Scor eM P(Satis f actionQualityOfMedia(LR, MPi) , Tex pect ed, Tused) =

Sati s f acti on

QualityOfMedia

1+(Tused−Texpected)/Texpected, if Tused > Texpected, Sati s f acti onQualityOfMedia, Otherwise

where Tex pect ed = the maximum available DT and Tused= the actual time spent

delivering this adapted media Version (V) transcoded by MPi.

Example 5 Given an LR= ((LO1, p1), (500, 1, 1000, 448, 480, 800, 24, 16, 140,

1, GPJB, 1, 0)). Assume there are two MOs, MO1= Image, MO2= audio, in requested page p1 of LO1, where MO1 has an MP1= (“MO1v1”, image, 800, 400, 16, J, 160) without a corresponding physical adapted media file in the LOR.

Therefore, the MO1will be added into Mediar eqonly due to the Audio Switch (AS)

is 0, i.e., Mediar eq = {MO1} (Step 1). Then, all MPs of MOs in Mediar eqwill be

inserted into MPcandifor calculating the satisfaction score. Thus, the MPcandi={MP1}

(Step 2). Afterwards, we can estimate the Tex pect ed= 1/1=1 to understand how much

time we can use to do the CAP for each requested MO (Step 3).

For each MPi ∈ MPcandi, we estimate how much time we need to spend

delivering the media size over the Bandwidth (B) of the wireless network, i.e., Tdeliver= 160/500 =0.32 s; and how high the SatisfactionQualityOfMedia is, i.e.,

Sat1= 0.463, as described in Example 4. Then, because MP1 has no correspond-ing physical media file in the LOR, the nearest physical media file, (“MO1v2”, image,

1000, 500, 16, J, 160), in the LOR will be selected to estimate its transcoding time in

advance if we deliver it to the user. Here, we can assume Ttr anscodi ng= 1s. Therefore,

the satisfaction score of MP1can thus be calculated by:

Sati s f acti on Scor eM P(0.463, 1, 0.32 + 1) =

0.463

1+ (1.32 − 1)/1 = 0.35. (Step4)

Finally, if the MP1has the maximum satisfaction score in terms of MO1, it will be selected to insert into MPset, which will be used to perform the learning content

(18)

Algorithm 2

Algorithm 2: Content Adaptation Process Algorithm (CAPAlgo) Symbol Definition:

LR: denotes a learner request, i.e., LR = (LO, (B, HP, LP)). MO: denotes a Media Object∈ LO

MPcandi:the candidate MP list

Mediar eq: the set of requested media object

MPset={MP1, MP2,…, MPk}: stores all appropriated MPs selected by CAPAlgo.

Sati: the SatisfactionQualityOfMediaof MPi = Vi, T, A, S, TP

TM DT: maximum available delivery time, default is DT∈ LP in LR

Tex pect ed:the average expected time of deivering each requested media object

Tdeliver: the estimated deliver time of the media version (V) in MPi

Ttr anscodi ng: the estimated transcoding time of the media version (V) in MPi

Input: a LR, TM DT

Output: MPset

Step 1: add all requested media objects (MOs)∈LO in LR into Mediar eq Step 2: for each MO∈ Mediar eq, add all MPi ∈MO into MPcandi

Step 3: calculateTex pect ed= TM DT/(the number of MO in Mediar eq)

Step 4: for each MPi ∈ MPcandi

(1) Calculate Sati= SatisfactionQualityOfMedia(LR, MPi)

(2) Calculate Tdeliver= si ze(S) ∈ MPi/Bandwidth(B) ∈ LR

(3) Calculate T tr anscodi ng=

0, i f physical f ile of MPii s i n Lear ni ng Obj ect Reposi t or y(L O R) Esti mat e tr anscodi ng ti me f r om near est physi cal f ile i n L O R (4) Calculate Satisfaction Score of MPi= SatisfactionScoreM P(Sati, Tex pect ed,

Tdeliver + Ttr anscodi ng)

Step 5: for each MO∈ Mediar eq,

(1) Select MPiwith maximum Satisfaction Score and store it into MPset

Step 6: Return the MPset.

5.2 Adaptation decision process

In the Learning Content Adaptation Management Scheme (LCAMS), the Content Adaptation Decision Tree (CADT) can be used to search, retrieve, and maintain his-torical CARs. The desired adapted contents can be delivered quickly to learners if there is a similar existing learner request held by CADT. Determining how to effi-ciently deliver an appropriate adapted content from the existing CARs or how to redo the aforementioned Content Adaptation Process (CAP) is a concern. We propose an Adaptation Decision Process Algorithm (ADPAlgo) to process the adapted content decision quickly. The ADPAlgo is shown in Algorithm 3 and illustrated in Fig.5.

In the Adaptation Decision Process (ADP), as shown in Fig. 5, we are given a new LR= ((LOi,pj), (BL R, HPL R, LPL R)). First, a suitable cluster will be selected

by traversing the CADT based on HPL R. Second, these CAR= ((LOk,pm), (BC A R,

HPC A R, LPC A R)) in the selected cluster will be merged with those in the CAR Pool,

(19)

e.g., (LOi,pj) = (LOk,pm). Fourth, if there is a remaining CAR with higher similarity

compared to the LR, the Learning Content Synthesizer (LCS) in Sect.5.3will com-pose the personalized learning content and transcode the associated MOs based on necessity. Then, the adapted learning content will be delivered to a learner directly without or with low transcoding latency. Otherwise, the CAP will be triggered to create a new CAR based on the LR.

Algorithm 3

Algorithm 3: Adaptation Decision Process Algorithm (ADPAlgo) Symbol Definition:

CARset: stores several historical CARs

LR: denotes a learner request, i.e., LR= (LO, (B, HP, LP)). CARnew: stores the new CAR created according to the LR.

α : denotes the acceptable percent threshold of bandwidth deviation.

β : denotes the acceptable weight Threshold of the amount (NH P) of

attributes in HP.

γ : denotes the acceptable threshold of Similarity value.

SMi n= β × NH P: denotes the minimum amount of the same attributes value

between HPC A Rand HPL R.

Input: a LR

Output: a suitable CAR

Step 1: If the CADT is not Empty,

Then use the HP in LR to traverse the CADT for finding the suitable cluster with similar HP.

Step 2: Insert CARs into CARsetfrom the selected Cluster in CADT

and CARs Pool.

Step 3: Delete these CARs from CARset, if (LOi j∈CARset) = (LOmn∈LR).

Step 4: Delete these CARs from CARset, if|(B ∈ CARset) − (B ∈ LR)|  α × B ∈ LR.

Step 5: Delete these CARs from CARset, if the number of HP attributes with

similar value in CAR compared with LR < SMi n.

Step 6: Delete these CARs from CARset, and if the similarity between CAR in

CARsetand LR according to the SimilarityL P(LPL R, LPC A R) < γ .

Step 7: If∃ a CAR∈ CARsetwhose attribute values in HP and LP is the same as LR, Then goto Step 9.

Step 8: do the Content Adaptation Process (CAP) according to the LP in LR and create the CARnewstored in CARs Pool.

Step 9: If CARsetis not empty,

Then Output the CAR with the highest similarity in CARset.

Else Output the CARnew.

Example 6 Based on the data in Table 1, given a new Learner Request (LR), LR= (LOi j, (B, HP, LP))= (LO, (90 KB, 2, 528, 384, 320, 480, 24, 32, 120,

5,JGBP,1,1 )) and a new CAR 16 = (LOi k, (150 KB,1, 133, 128, 480, 640, 16, 16, 44, 12, GJBP, 0, 1)) in the CAR Pool, according to the CADT in Fig.3 and Adaptation Decision Process Algorithm (ADPAlgo), we can find the rule: if Machine

(20)

Fig. 5 ADP process

Type (MT)= ‘2’ and Color Depth (CD) = ‘24,’ then ‘C2,’ so that we can use the CAR ID, {2, 5, 7, 9, 11, 12, 13, 15}, of C2in Table2and CAR 16 in CAR Pool to select a suitable CAR (Steps 1 and 2). Then, CAR 16 is deleted due to (LOi k = LOi j), and

CARs 2, 5, 7, 11, 12, and 13 are deleted due to their bandwidth deviation>=45(α × B) whileα is 0.5 and B is 90 KB (Step 3 through Step 4). Afterward, CAR 15 with 7 sim-ilar attributes while Smi n= 0.9(β) × 8(NH P) = 7.2 and the similarity value = 0.772

(>γ =0.6) compared with an LR is a suitable CAR for the user (Step 5 through Step 6).

However, because CAR 15 is not completely the same as the LR, a new CAR, CAR 17, will be created by the Content Adaptation Process (CAP) based on the LR and stored in the CAR Pool for the next similar learner request (Step 8). Thus, the CAR Pool will hold two new CARs, i.e., {16 and 17} and the adapted content version based on CAR 15 will be delivered to a learner directly. Because the content version of CAR 15 was adapted according to the previous similar learner request, the CAP process does not need to be executed again. Therefore, the adaptation and transcoding latency can be omitted and saved.

5.3 Learning content synthesizer

The Learning Content Synthesizer (LCS) aims to compose appropriate personalized learning content according to the adaptation decision of the PLCAM based on diverse learner preferences. As stated previously, when dealing with a new LR without any suitable existing adapted content to be delivered, the CAP will decide a corresponding MPset to transcode the associated media resources. Hence, given that a page has n

media resources and its corresponding MPset={MP1, MP2,…,MPm}, where 1 m 

n, it is implied that the (n–m) resources do not need to be transcoded and shown due to the satisfaction degree and Switch Attribute (SA) of media, e.g., PS and AS in the LP. To notify users, media resources that are not shown will be automatically replaced by some additional annotations from the SCORM metadata. To efficiently manipulate

(21)

Algorithm 4

Algorithm 4: Learning Content Synthesis Algorithm (LCSAlgo) Symbol Definition:

TC A P: the spending time of executing the Content Adaptation Process (CAP).

TA D P:the spending time of executing the Adaptation Decision Process (ADP).

Tdeliver: the estimated deliver time of the media version (V) in MPi.

Tscor e: the minimum threshold of Satisfaction for the content adaptation process.

Tused:The used time of MPs, which needn’t be re-adapted.

DT: the maximum available delivery time, DT∈ LP in LR. MO: denotes a Media Object∈ LO.

MPset={MP1, MP2,…, MPk}: stores all appropriated MPs selected by CAPAlgo.

Mediaadapt: store the media objects, which need to do content adaptation process

: denotes the original media resource in a page. Trγ: denotes the transcoded media resource. Input: a LR with corresponding MPsetand LOi j.

Output: an adapted and transcoded learning content version, XHTML. Step 1: if (CAR= ADPAlgo(LR)) = null,

Then

(1.1) Estimate the TC A Pand TA D P

(1.2) MPset= CAPAlgo(LR, DT−(TC A P+ TA D P)), where TM DT in

CAPAlgo= (DT−(TC A P+ TA D P)).

Else

(1.3) for each MPi ∈ MPsetin CAR

(a) Calculate Sati = SatisfactionQualityOfMedia(LR, MPi)

(b) Calculate Tdeliver = size (S)∈ MPi/ Bandwidth(B)∈LR

(c) CalculateT tr anscodi ng=

0, if physical file of MPiis in Learning Object Repository(LOR)

Estimate transcoding time from nearest physical file in LOR (d) Calculate Satisfaction Score of MPi = SatisfactionScoreM P(Sati,

DT/(the number of MPi∈ MPset), Tdeliver + Ttr anscodi ng)

(e) If Satisfaction Score of MPi< Tscor e, then add media object (MO) of

MPiinto Mediaadapt

(1.4) Estimate the TC A Pand TA D P

(1.5) for each MPi ∈ MPsetand/∈ Mediaadapt,

(a) Calculate Tused = Tdeliver+Ttr anscodi ng

(1.6) for each MOk ∈ Mediaadapt,

(a) the MPiof MOk= CAPAlgo(LR, DT−(TC A P+ TA D P + Tused)),

where Mediar eqin CAPAlgo= Mediaadaptand TM DT

in CAPAlgo= (DT− (TC A P+ TA D P + Tused)).

(b) replace the original MP of MOk∈ MPsetof CAR by MPi.

Step 2: for each media resource, rγ,in a page pi.

(1) apply MPγ ∈ MPset, to transcode the rγinto the trγ. Step 3: transform the original HTML into XHTML format Step 4: replace all rγ by trγ into the XHTML.

Step 5: replace all unshown media resources by the useful annotation from SCORM metadata.

(22)

the diverse versions of learning content, a page’s original HTML will be transformed into a well-formed XHTML and tree-like Document Object Model (DOM) structure (2010). The details of the LCS are described in Algorithm 4.

Example 7 Assume the CADT structure has already been built. Then, given LR= ((LO2, p1), (500 KB, 1, 1000, 448, 480, 800, 24, 16, 140, 2, GPJB, 1, 0)) and the minimum threshold of satisfaction, Tscor e= 0.7. After the ADPAlgo(LR) process,

there are two corresponding MOs in MPset, i.e., MPset={MP1, MP2}= {(“M1v0”, image,800, 400, 16, J, 160), (“M2v1”, image,480, 700, 24, P, 50)}. Thus, assume TA D P= 1 s, and Tdeliver of MP1= (size in MP1)/ (bandwidth in LR) = 160 KB/ 500 KB= 0.32 s and Tdeliver of MP2 = 50 KB/ 500 KB= 0.1 s (Step 1.3.b). The associated Media Object, MO2, of MP2does not need to be adapted again because MO2has the adapted media version (V1) based on MP2, i.e., M2v1, in the LOR. On the contrary, M1v0, which is the original version (V0) of MO1, must be adapted according to the definition of MP1before it can be delivered to the learner. Therefore, assume Ttr anscodi ng= 1 s to adapt M1v0into the nearest version, i.e., (“M1v2,” image,800, 400, 16, B, 500)(Step 1.3.c).

In addition, according to the result of Example 4, the Sat1of MP1is 0.463 by the equa-tion: SatisfactionQualityOfMedia(LR,MP1) (Step 1.3.a). Consequently, the satisfaction

score of MP1is calculated by:

SatisfactionScoreM P(0.463, 2/ 2, 0.32 + 1) =

0.463

1+ (1.32 − 1)/1 = 0.35. Besides, Sat2is calculated by:

Sati s f acti onQualityOfMedia(LR, MP2) = 0.5 × SimofNum(480 ∈ HP, 480)

+ 0.1 × SimofNum(800 ∈ HP, 700) + 0.3 × SimofNum(24 ∈ HP, 24) + 0.1 × SimofImageT ype(“GPJB” ∈ LP, P) = 0.9625 (Step1.3.a).

Consequently, the satisfaction scores of MP2is calculated by: SatisfactionScoreM P(0.9625, 2 / 2, 0.1 + 0) = 0.9625

1+ (0.1 − 1)/1= 0.9625, because the Ttranscoding= 0 s. for adapting M2v1(Step1.3.d).

Because the satisfaction score of MP1= 0.35, which is less than 0.7 (Tscor e), and the satisfaction score of MP2= 0.9625, which is larger than 0.7, the MO of MP1, M1v0, will be added to the Mediaadapt for further adaptation process, i.e., Mediaadapt = {“M1v0”} (Step1.3.e).

Assume TC A P= 0.2 s and TA D P= 1 s for the “M1v0” (Step 1.4), “M2v1” of MP2 doesn’t need to be adapted again, so the Tused for MP2= Tdeliver + Ttr anscodi ng=

(23)

240, 24, P, 20) for the “M1v0” by calling the CAPAlgo(LR, TM DT) =

CAP-Algo(LR, DT−(TC A P + TA D P+ Tsave)) = CAPAlgo(LR, 2−(0.2 + 1 + 0.1)),

where Mediar eq in CAPAlgo = Mediaadapt={MO1} (Step 1.6.a). Therefore,

MPset={MP3, MP2}= {(“M1v0,” image,480, 240, 24, P, 20), (“M2v1,” image,480, 700, 24, P, 50)} (Step 1.6.b). According to the MPset, the media version, M1v0, of MP3will be transcoded by the definition of MP3first, and then output the XHTML with associated transcoded media resources (Step 2 through Step 6).

6 Implementation and experiment

The prototypical PLCAM system was developed on an Apache Server, PHP, Perl and C Language, and a SCORM-compliant LOR. Implementation details and experimental results are described in this section.

6.1 Mobile learning scenario for the personalized learning content adaptation mechanism system

The mobile learning scenario for the prototypical PLCAM system deploys a SCORM-compliant Learning Object Repository (LOR) that allows teachers to upload teaching materials. Students, in turn, can search and download the desired learning content onto their mobile devices. Students can log in to the system through their mobile devices and manually configure their own Leaner Preference (LP) with Weight Vector (WV). They may then select the learning content to download, and proper personalized learning content will be adapted and delivered by the PLCAM according to the stu-dents’ Hardware Profile (HP) and LP, and the status of the wireless networks. As the number of Learner Requests (LRs) increases, the Learning Content Adaptation Management Scheme (LCAMS) will automatically rebuild the Content Adaptation Decision Tree (CADT) and manage the adapted version of the Media Object (MO) in the LOR.

The operational flow of the PLCAM is illustrated in Fig.6. First, a student uses an ID to log in to the prototypical PLCAM system [Step (a)]. After logging in, a student can use the menu on the index page of the LOR to manually configure the following: (1) the HP setting, (2) the LP setting, (3) browse the content of the LOR, and (4) read the system manual [Step (b)]. The PLCAM can automatically detect the HP of a mobile device and the current bandwidth of a wireless network, as shown in the “Dynamic Attributes” [Step (c)]. Next, a student can use the “User Preference Configuration” to manually define the data of an LP, such as the preferred maximum Delivery Time (DT), presentation ratio of audio and picture, and the preferred priority of picture, etc. [Step (d)]. After configuring the HP and LP, a student can click “(3) browse the content of LOR” in Step (b) to browse the contents by the “Category Page” of the LOR [Step (e)]. The system will list the learning content stored in the selected categories [Step (f)]. A student can also select interesting learning content to browse its “SCORM metadata” and “Table of Contents” [Step (g)]. Consequently, the PLCAM will adapt the chosen content according to the mobile device profile (HP),

(24)

Fig. 6 Operational flow for a user to retrieve the learning content by the PLCAM system

LP, and the current Bandwidth (B) together, and offer personalized learning content to the student’s mobile device [Step (h)].

Figure 7 illustrates screenshots of the PLCAM delivering proper personal-ized adapted learning object content to the similar LR without waiting for the

(25)

a b c

Fig. 7 a Adapted content version of LPa; b delivered the adapted content version of LPafor LPbdue to

the higher similarity; and c delivered the content version created by LPbin advance for LPc

adaptation time of the requested content. The performance and effectiveness of the PLCAM will be evaluated and described in Sect. 6.2. Assume there is an existing adapted content version created by LPa=5,JGBP,0,0 of learner A, as shown in

Fig.7a. This existing version will be selected in advance and delivered directly to the new request with LPb=5,JGBP,1,0 of learner B, due to the higher similarity

estimation, as shown in Fig.7b. Therefore, learner B does not need to wait for the adaptation to take place again. In the meantime, the PLCAM will prepare an accu-rate content version with a background picture to meet LPb=5,JGBP,1,0, which is

stored in the LOR for the next similar request. For example, this prepared version of LPb can be delivered directly to meet the new LPc=5,JBGP,1,0, as shown in

Fig.7c.

Figure8shows several experimental screenshots of the PLCAM system executed on a PDA according to diverse user needs. Figures8a and b illustrate adapted content based on the same HP and LP with different adaptation parameters under different bandwidth values, respectively. As stated in Example 1, the attributes of the LP and HP can be extended to meet the various requirements. Thus, a new attribute in the LP, called Preferred Picture Property Ordering (PPPO), includes three properties: Dimension (D), Color Depth (C), and Quality (Q). This attribute is used to define the learner’s preferred order of image properties. For instance, like the attribute PPFO, a string, DQC, denotes that the order of image priorities is D > C > Q. Hence, we added the PPPO attribute into the LP and changed several parameters, e.g., Delivery Time (DT) and Audio Switch (AS), to test the results of the learning content adaptation process. As shown in Fig.8c, according to the new LP setting and original HP, the property of the picture was changed and the audio, background picture, and icon were replaced by hyperlinks with annotation text. We further evaluated the learning content adaptation capability by changing the screen’s horizontal size and color depth of the

(26)

Fig. 8 Screenshots of the learning content adaptation process performed by the PLCAM system

HP while using the same LP in Fig.8c. The desired adapted learning content was delivered by the PLCAM system, as shown in Fig.8d.

The PLCAM system also includes a monitoring interface of the LCAMS Web server, used to monitor the latest system status and maintain the CADT. As shown in Fig.9, the “Assign Cluster Label” function button can be used to perform the LP Clustering Algorithm (LPCALgo) for grouping the historical CARs into several clus-ters according to the learners’ LPs, where the resultant clustered information of the LPCALgo will be shown in the bottom-left part of Fig.9. Furthermore, the CADT can be reconstructed by the “Rebuild Decision Tree” function button. Its graphical presentation and rule-based representation will be automatically shown in the top-left part of Fig.9. The right-hand side of Fig.9will list all of the CARs.

(27)

Fig. 9 Monitoring interface screen of the LCAMS Web server in the PLCAM system

6.2 The experimental results

To evaluate the PLCAM system, actual experiments and simulated experiments were performed. The details and results are described in the following subsections.

6.2.1 Results of actual experiments

In the actual experiments, performance of the prototypical PLCAM system was eval-uated by the experimenters in terms of the personalized learning content delivering process, including the Content Adaptation Process (CAP), the Adaptation Decision Process (ADP), and the Learning Content Synthesizer (LCS), and the dynamic band-width detection scheme, based on the SCORM-compliant learning content in relation to “the plants in campus” shown in Sect. 6.1. These characteristics were tested to observe and evaluate the resultant Delivery Time (DT) and the transmission data size according to various bandwidth settings for different requests. The performance of the PLCAM in terms of DT compared with inadaptation and static adaptation approaches was evaluated as well.

As mentioned in Sects.1and 2, the DT plays an important role in affecting one’s learning performance in mobile learning environments. Therefore, a bandwidth detec-tion scheme was developed to automatically detect the latest network bandwidth for providing the learner with more precise personalized learning content with higher fidelity. As shown in Fig.10a, with the decrease of “actual” network bandwidth, the

(28)

Fig. 10 Experiment results of the automatic dynamic bandwidth detection scheme

bandwidth that a user can consume will decrease as well by means of monitoring each transmission time compared with the setting of the user’s desired maximum DT. For example, in Fig. 10b, a user’s functional bandwidth has been updated at 3, 5, and 7 (times) due to the detection of the long delivery latency at 2, 4, and 6 (times) (maximum DT= 5 s).

Figure7shows the effectiveness of the PLCAM in delivering proper personalized adapted learning content that meets a similar LR without waiting for the CAP to be completed. Therefore, Fig.11illustrates the DT, which consists of the data transmis-sion time and the PLCAM adaptation time, in terms of the different requests and the size of transmission data based on the various bandwidth settings. In Fig.11a, assume the definition of the maximum DT is 5 s, during the first request for a learning object, the average DT is about 8.416 s, including the content adaptation process (about 3 s) and the actual content delivery. On the contrary, the average DT during the second request can be controlled around 5.238 s to meet the constraint of the maximum DT without repeating the content adaptation process. Figure11b shows that the size and quality of transmission data can be increased gradually with the increase of usable bandwidth by the aforementioned dynamic bandwidth detection scheme based on the definition of maximum DT.

Assume that there is a learning object, which contains a Waveform Audio Format (WAV) file with 660,768 bytes and six pictures with 1,016,392 bytes. The original size is about 1.8 Mbytes. The learner specified the maximum tolerable DT to be 5 s We observed the transmission results based on the various bandwidth settings in terms of the approaches, which included the inadaptation, static adaptation, and PLCAM. In

數據

Fig. 1 Architecture of PLCAM
Table 1 Example of CAR for the same page (p j ) in an LO (LO i ), (LO i j )
Table 2 Result of applying LP
Table 4 HP in CAR data with
+7

參考文獻

相關文件

 Promote project learning, mathematical modeling, and problem-based learning to strengthen the ability to integrate and apply knowledge and skills, and make. calculated

← This allows teachers to adapt the school-based English Language curriculum and devise learning/teaching materials that better suit the diverse abilities, needs

Now, nearly all of the current flows through wire S since it has a much lower resistance than the light bulb. The light bulb does not glow because the current flowing through it

Students are asked to collect information (including materials from books, pamphlet from Environmental Protection Department...etc.) of the possible effects of pollution on our

The temperature angular power spectrum of the primary CMB from Planck, showing a precise measurement of seven acoustic peaks, that are well fit by a simple six-parameter

This kind of algorithm has also been a powerful tool for solving many other optimization problems, including symmetric cone complementarity problems [15, 16, 20–22], symmetric

In order to achieve the learning objectives of the OLE – providing students with a broad and balanced curriculum with diverse learning experiences to foster whole-person development

• To consider the purpose of the task-based approach and the inductive approach in the learning and teaching of grammar at the secondary level.. • To take part in demonstrations