Lecture Notes in Computer Science: Vol_20_No_2.files/JO… · Web viewJason C. Hung, Kuo-Feng...

Adaptive Caching and Presentation Mechanismto Enhance M-Learning Market

Jason C. Hung, Kuo-Feng Hwang, and Neil Y. Yen*

Department of Information Management

Overseas Chinese University

No:100, Chiao Kwang Rd., Taichung 407, Taiwan, R.O.C.

{jhung, kfhwang}@ocit.edu.tw

Department of Computer Science and Information Engineering

Tamkang University

Tamsui, Taipei Hsien, Taiwan 251, R.O.C.

[email protected]

Received 30 May 2009; Revised 30 June 2009; Accepted 8 July 2009

Abstract. Many innovative learning systems and activities have also been proposed to promote e-learning, as well as exploited the ways of e-learning. Although the variety of mobile platforms provide more flexi -ble and extendable learning experience, the various hardware conditions and restrictions consequently be -comes the challenges and barriers we need to overcome. To enhance the mobility in e-learning market, we propose an adaptive caching and presentation mechanism to meet the need. In our mechanism, we utilize the benefits from the use of SCORM (Sharable Content Object Reference Model) to deal with the problem could happen in content extraction. Besides, we also take the multimedia resources into consideration and utilize the non-synchronous caching algorithm to solve the streaming problems. The mechanism we pro -posed will promote the application of ubiquitous devices in e-learning market.

Keywords: M-Learning, caching, adaptive presentation, content prefetch, SCORM

1 Introduction

Mobile Commerce, or m-Commerce, is about the explosion of applications and services that are becoming ac -cessible from Internet-enabled ubiquitous devices. It involves new technologies, services and business models. It is quite different from traditional e-Commerce. Ubiquitous devices, especially mobile phones impose very different constraints than desktop computers. But they also open the door to a slew of new applications and services. They follow you wherever you go, making it possible to look for a nearby restaurant, stay in touch with colleagues, or pay for items at a store.

The use of ubiquitous technologies to facilitate the learning process has great impact on improving the fea -sibility of e-learning market. The emergence of ubiquitous technology and handheld affordances has created new opportunities for anytime and anywhere learning paradigm, but it is also associated with essential limita -tions and problems when involving these technologies to construct a practical mobile learning system (Tan and Liu 2004). Su and Seong (2006) proposed the usability guidelines for designing mobile learning portals, indi -cating the limitation of screen size, the presentation of mobile contents, and adaptation of the information to the sensitivity of context (Wang 2004) and devices influence the efficiency and effectiveness when learning via the mobile devices.

The use of ubiquitous devices like pocket PCs and smartphones to support learning is not a new concept. Many innovative learning platforms and activities have also been proposed to promote e-learning, as well as exploited the ways of e-learning. Although the variety of ubiquitous platforms provide more flexible and ex -tendable learning experience, the various hardware conditions and restrictions consequently becomes the chal -lenges and barriers we need to overcome.

* Correspondence Author

mailto:[email protected]

Journal of Computers Vol.20, No.2, July 2009

1.1 Problem Definition & Our Contribution

Many systems to improve the m-learning market aim at creating a practical environment in which students ad -equately enjoy the mobility of learning. However, these systems encounter a problem in which learners proba -bly move into an area outside a local area network connection. The problem usually disables the learning func -tions such as the tracing of learning behaviors, content delivery, and data synchronization. Another important issue in this research area is the use of context-aware technique (Syvanen et al. 2005； Malek et al. 2006) to enhance the usability of ubiquitous learning systems. A similar concept is taken into account for designing the course caching strategy that considers the related context information as an important parameter. Furthermore, a GPRS or 3G telecommunication protocol offers continuous connectivity for learning. Table 1 shows the per -formance of a smartphone with GPRS used to browse sample courses released by advanced distributed learn -ing (ADL). These courses are stored in our remote learning management system (LMS) server. As shown in Table 1, the average time costs required to completely displaying 10.5 KB HTML-based content and 7.35 KB HTML-based contents are 17.67 s and 15.7 s, respectively. On the contrary, if the same contents are down-loaded beforehand and stored in the local storage, only 3 seconds are required to completely display both cour -ses, as shown in Table 2.

Table 1. Loading time of display online learning content with smartphone

Photoshop_Compatancy First Second Third AverageLesson 2 (10.5 KB) 19 14 20 17.6Lesson 4 (7.35 KB) 18 16 13 15.7

Unit: second (s)

Table 2. Loading time of local course with smarthpone

Photoshop_Compatancy First Second Third AverageLesson 2 (10.5 KB) 3 3 3 3Lesson 4 ( 7.35 KB) 3 3 3 3

Unit: second (s)

Therefore, if the caching mechanism can correctly/accurately predict the access of courses and to download them before learning begins, not only the learning process can enjoy the reduced waiting time significantly but also it offers the downloaded content for offline learning. On the other hand, a large number of excellent digi -tal materials have been created and distributed for attracting more e-learning customers. These contents are mostly designed for reading on regular PCs that have big screens, powerful computing, large storage and wide bandwidth compared to ubiquitous devices. The improper presentation leads to incontinence of reading that significantly degrades the learning quality and desire. As the problem we stated above, we proposed an adap-tion mechanism that allows system automatically and efficiently reproduce the high-quality learning content for specific devices in m-learning market. Through this mechanism, the m-learning companies can continually focus on the generation of various kinds of learning materials, and they do not have to worry about the internet transmission capability and the presentation of the materials. Besides, the cost of time in downloading and pre-viewing the learning content can be greatly shortened. The mechanism we proposed can seriously improve the quality of m-learning market.

The remainder of this paper is organized as follows: Section 2 presents a brief survey on the related works, starting from the introduction of m-commerce environment; meanwhile a few relevant technologies are pointed out the current important issues for m-learning. In section 3, we detail each proposed module includ -ing the adopted technologies, strategies and goals to support constructing an adaptive m-learning environment. Section 4 illustrates the demonstration systems that include; the content adaptation authoring tool that verify the practicability of the proposed architecture and interoperability among these modules. Section 5 examines the experimental results and analysis of our proposed ideas. Finally, the conclusion and the future work are shown in Section 6.

2 Related Works

2.1 M-Learning Research Issues

The emergence of mobile devices and wireless communication has opened up another alternative way for e-education that employs mobile technologies as a mean to transfer knowledge to learners (Buchanan et al., 2001). Mobile learning has offered flexibility in learning and present unique educational advantages. The 30

Hung et al: Adaptive Caching and Presentation Mechanism to Enhance M-Learning Market

widespread deployment of wireless infrastructure and rapid adoption of handheld computing devices poten-tially lead many researchers and educationalists to move from web-based and e-Learning to mobile learning which explores more individual and easy ways of learning.

Wagner (2005) mentioned a rich mobile Internet experience to enable mobile learning should include the following attributes: Ubiquity, Access, Richness, Efficiency, Flexibility, Security, Reliability, and Interactiv -ity.

2.2 Acknowledged Research Literatures

The design guidelines for such PDA devices are also introduced and discussed in (Kärkkäinen and Laarni, 2002). These guidelines can be classified according to which aspect of the Web media they are related: soft -ware/hardware, content and its organization, or aesthetic and layout. The need to adapt content for use on handhelds has been long recognized (Katz 1994, Kindberg, and Fox 2002; Narayanan 2000 et al.), and both manual and automatic approaches to implement the content adaptation have been proposed. This research (Chen, Ma and Zhang 2003; Ramaswamy et al. 2004; Zhao and Yang 2005) mostly focused on adapting nor-mal Web Pages such as commercial web sites or portal sites. There have been a lot of automatic approaches designed to provide a real time content adaptation system for browsing Web Pages on handhelds. On the other hand, manual adaptation techniques, such as WAP (Kaasinen 2000; http://www.wapforum.org/what/technical/arch-30-apr-98.pdf), have high cost for data producers who are required to maintain multiple ver-sions of the content.

Adaptation is a well-studied topic in mobile and pervasive computing for years. Hwang et al. proposed a transcoding framework (Hwang, Kim, and Seo, 2003； Han et al. 1998) that represents a Web page as a modi-fied tree structure to efficiently analyze and transcode pages. It is based on html syntax analysis and structure-aware techniques that intend to make complex Web pages accessible and reflect the relative importance of Web components during the transcoding process. On the other hand, refer to textual web content summariza-tion (Orkut et al. 2002), the methods for summarizing are introduced to handle the textual Web pages and HTML forms. A Web page is separated into text units that can each be hidden or partially displayed.

Other related research put their efforts on attempting to restructure the layout of web pages (Chen et al. 2003). Wireless Application Protocols (WAP) such as Wireless Markup Language (WML) and Handheld De-vice Markup Language (HDML) are popular approaches to customize the appearance of web content for mo -bile devices. These protocols can be used to split an HTML page into small cards such that each card can fit into a single screen. Links are made between cards to enable browsing of the page. The two main weaknesses of this approach are that we lose the layout and presentation style of the original web page and that navigating the cards can be a cumbersome process requiring a large number of selections to retrieve desired information.

Usage-AwaRe Interactive Content Adaptation (UARICA) and Feedback-driven Context Selection (FCS) (Mohomed et al. 2006) made adaptation prediction for a user based on the history of the community of users and reflect both the user’s context and content’s usage semantics. Iqbal et al. think optimal adaptation is a challenging problem because it often depends on the usage semantic of content, as well as the context of users (e.g., screen size of device being used, network connectivity, location, etc.) Their works included an automatic techniques, UARICA that allows a user who is unsatisfied with the adaptation prediction to take control of the adaptation process and make changes until the content is suitably adapted for his/her purpose. Moreover, FCS takes advantages of user interaction to determine those contextual characteristics that have the most impact on the adaptation requirements of an object, and therefore should be the basis of grouping users into communities.

Orkut (2002) proposed six different display modes are introduced that utilizes the progressive displaying textual units with keyword extraction and paragraph summarization to gain an overview of a page. They found that the combination of keywords and summaries provides the most significant improvements in access time and number of required pen actions. In this research, TF/IDF was adopted to handle the adaptation of text body, the following is the detail introduction of TF/IDF:

The TF/IDF weight (Zhou, Tang and Wang 2007) (Term Frequency–Inverse Document Frequency) is a weight often used in information retrieval and text mining. This weight is a statistical measure used to evaluate how important a word is to a document in a collection or corpus. The importance increases proportionally to the number of times a word appears in the document but is offset by the frequency of the word in the corpus. Variations of the TF/IDF weighting scheme are often used by search engines as a central tool in scoring and ranking a document's relevance given a user query.

3 System Framework

There are four main modules proposed to carry out the course caching strategy and presentation adaptation methods that aim at promoting the immediacy, accessibility, and interactivity for SCORM-compliance mobile

31


learning environment. The Fig. 1 depicts the entire framework which based on Web Service and Context-Aware technologies to develop the four modules.

Fig. 1. Proposed system modules

3.1 Course Segmentation Module

As shown in Fig. 2, a course activity tree can be separated into several smaller cluster units which considered as a basic building block of learning activity. SCORM sequencing is especially applied to clusters. The cluster includes a parent learning activity and its immediate children. The parent activity of the cluster will contain the learning sequencing information. The children can be a set of leaf learning activities which are the physical learning resources for delivering to learners.

Fig. 2. An illustration of clusters within a course

32


In our proposed strategies, a cluster can be considered as the basic unit to be manipulated. Furthermore, in the sequencing definition model, there are many elements to describe and affect the learning behaviors of learners, such as Sequencing Control Modes for specifying the sequencing behavior, Limit Conditions for set -ting criteria of learning competence, Rollup Rules for gathering statuses from the child activities, etc.

The following is an example of how to segment a SCORM-compliant course which was released by ADL sample course; Photoshop_Compatancy Choice. The Fig. 3 shows the entire course aggregation tree which is constructed according the manifest file attached in the course content package. According to the definition of a cluster, it includes a parent learning activity and its immediate children. The parent activity of the cluster will contain the learning sequencing information. Thus, there are six clusters can be separated from the navigation tree. Each cluster may contain various content resources that result in the size of each cluster may be varied to each other.

Fig. 3. An example of course segmentation

3.2 Course Caching Module

The caching strategy includes two modes— VMM (Virtual Memory Management) mode and COD (Caching of Disk) mode—for dealing with different device conditions and network requirements. Due to the availability of the Internet and the adopted learning device, the system will switch between these two modes in order to meet the requirements of an efficient ubiquitous learning system. The concept of the VMM mode is similar to that of virtual memory management in conventional operating systems. This mode can provide a virtual net -work environment so that the cached clusters are sufficient for learner’s requirements even stays at offline sta -tus. It mainly addresses the suspension due to the intermittent network connection. In contrast, the concept of the COD mode is similar to caching of disk. This mode can prepare the possible incoming learning resources in advance to avoid the time-consuming process of downloading the required contents.

Both the VMM and COD modes are implemented according to the prediction of the learning behavior of the learner. In this regard, it is important to determine the parts of the course contents that should be cached. The prediction results will directly influence the performance of the course caching system. In order to im -prove the accuracy of the prediction, we consider the SCORM S&N information and other available informa -tion collected from devices and users. The details of the factors for prediction and the course caching strategy are provided in the following subsections.

3.2.1 The Prediction Factors

The caching strategy predicts a user’s action and pre-downloads the most required clusters of the course into the specific learning devices. This establishes the need for users to read the learning content without requiring

33


either a permanent Internet connection or waiting for a long latency period to display the content on the de -vice. There are three types of information that are taken into account while carrying out the caching strategy.

3.2.2 Factors of Individual Course Information and Sequencing Setting

A set of factors is evaluated in the proposed course caching strategy to predict the learning resources that will be downloaded prior to the start of the learning activities. Some of these possible factors can be derived from the essence of the course content. These factors have a strong relationship with the sequencing information, which is specified in each cluster. It should be noted that the basic unit to be downloaded or to be dropped in the proposed caching strategy is a cluster defined under SCORM specifications rather than an individual learn -ing asset or SCO (i.e., an elementary element in SCORM courses). Each cluster has its own identifier and con -tains several learning resources. The size of a specific cluster is the sum of the sizes of the physical learning resources within it. Therefore, the size of a cluster can also be considered to be an important factor for the pre -diction in the caching strategy due to the limited storage capacity of mobile learning devices.

The course caching system in the VMM mode prefers to download as many clusters as possible to increase the hit ratio. The hit ratio can be calculated according to the availability of the selected lesson on the mobile learning device at that time. The hit ratio will increase if the selected lesson has already been downloaded to the mobile device; otherwise, it will decrease if the selected lesson needs to be downloaded from a remote server to the mobile device. However, the course caching system in the COD mode prefers to download large-sized resources first in order to reduce the waiting time when a learner is reading online.

In order to measure the relationship between two clusters, we quantify a simple value termed the “path length.” This factor based on the physical distance in a tree structure to determine when the relationship be -tween two clusters is strong or weak. If a cluster’s path length is close to that of the current cluster, there might be a higher probability for the learner to read after finishing the current learning activity. The path length is defined by (1), and Fig. 4 shows an example of the path length derivation.

PLij = Li + Lj - 2．Lc ()

PLij: path length from cluster i to cluster j Li: level of cluster i Lj: level of cluster j Lc: level of the first common parent of cluster i and cluster j

Fig. 4. Path length between two individual clusters

3.2.3 Factors of Specific Devices

The second type of factor is related to the characteristics of various learning devices. One of the most impor -tant factors is the capability of the network connection. This factor determines the mode that should be used in the course caching system. If the specific learning device can access the Internet all the time with sustained connectivity, the system uses the COD mode to reduce the waiting time. However, if the connectivity is inter-mittent or if it fails, the system employs the VMM mode to download as many clusters as possible once in or -der to reduce the network reconnection requests.

In general, the storage capacity of mobile devices that support mobile learning is insufficient to store all the learning materials. Thus, the factor can be applied to influence either the number of clusters to be preloaded into the mobile device or the number of clusters to be replaced.

34


3.2.4 Factors of Learning Record

A learning record is maintained by a learning system when the user starts learning. These factors are currently most referenced when the course caching system requires the replacement of some defunct clusters with newly downloaded clusters.

There are three types of learning record factors: reference count, last access time, and downloading time. If a cluster is referenced frequently, its demand might be higher than that of a cluster that is seldom accessed. Moreover, if a cluster is recently read or is newly downloaded, it should be dropped after the other clusters that have already existed in the client device for a specific period of time.

3.2.5 Automata of Course Caching Strategy in Learning System

Fig. 5 illustrates the diagram of the course caching module executing the caching strategy. Generally, the user’s handheld device would not be used only as a digital learning platform. The available storage capacity should be assigned by the users themselves for the course caching operations.

Fig. 5. Automata of the proposed course caching system

3.3 Course Presentation Adaptation Module

Our proposed adaptation module is composed of three main adaptation models. First, the textual adaptation model is responsible for handling the complicated textual body that may make users feel confused or lost while reading on a restricted small screen. Precisely speaking, our content adaptation tool will summarize tex-tual body and utilize progressive disclosure presentation to revel the original content. We referenced Orkut’s progressive disclosure for text, it combined with keywords and a summary help present the original content in -crementally, has the best improvement of average I/O expenditure and completion time across all tasks. The progressive displaying steps are shown in Fig. 6.

Second, the image adaptation model takes into account the requirement for adapting image size when dis -playing it on a handheld. Briefly speaking, an image will be automatically shrunk if it is too big to display on a handheld, as well as expanded if it is too small. Finally, the layout adaptation model is able to reorganize the layout of adapted elements properly according to the display ability of different handheld devices. But there are a few questions associated with previous descriptions. Exactly which textual body is needed to be summa -rized? How do we evaluate and decide whether a picture is required to be shrunk or expanded? Consequently, before we continue to detail each adaptation module, the presentation unit (PU) and screen unit (SU) will first need to be introduced.

35


Fig. 6. A PU progressively displays the text in three states

3.3.1 Semantic Unit (SU) and Presentation Unit (PU)

The adaptation process begins by partitioning the content into presentation unites (PUs). A content page will be separated into several PUs, which instead of presenting in the actual HTML, each PU is a rectangle around a section which typically presents a paragraph, list, table, image, etc. Accordingly, each PU is considered as a basic unit of the adaptation process. Because each PU contains various contents, the question is which SU should execute the adaptation process. We will define the other unit, namely screen unit (SU), that helps us to evaluate and decide which PU is required to be adapted.

The display area size of a PU is varies because it may contain a paragraph or an image; a SU is a virtual rectangle presentation block where the boundary is fixed according to different handhelds displaying ability. Precisely speaking, the size of a SU is matched to correspond to the screen sizes of handhelds For example, pocket PC’s and smartphone’s in which the typical resolution are 240*320 and 176*220 respectively. The con -tent of each PU will be retrieved then filled into each SU. The main concept is that it does not need to adapt a text or image within a PU if it can be entirely displayed within a single SU without additional scrolling. Math -ematically, we evaluated whether a PU is required to be adapted with a simple formula that calculates a threshold value. We defined a value, textual information density (TID) as follows:

TID = number of words in a PU / area of a SU (2)

The area of a SU is constant according to which adapted target platform is required by a user. The default value of TID is allowing a PU to present the maximum number of words without additional scrolling. Users are also allowed to adjust the TID, which will affect which PU is required to be adapted. For example, a larger TID allows a good deal of textual information located in a PU without any summarization so that the user may need more necessary scrolling actions for reading. The size of a picture will be shrunk or expanded for image displaying, The image in a PU will be adjusted proportionately and conserve the original picture’s presentation until it fits in a SU. The following sections will detail each adaptation module.

3.3.2 Textual Adaptation Model

For textual adaptation process, we proposed two different ways to handle text body. First, utilizing the text to speech technology to transform the complicated text body to a voice file attached in the course content. Sec -ond, for text summarization, we referenced Orkut’s approach to extract keywords from web pages. Their con-tent adaptation approach utilizes keywords and summary sentences to partially represent the original text, then disclose information progressively as figure 3.1 shows. The main difference is that our adaptation unit is a PU and only when its TID is higher than the chosen threshold value will it trigger the adaptation process. Next, the details of how to extract keywords and summary sentences will be introduced.

3.3.2.1 Keyword Extraction

36


Keyword extraction from a text body relies on an evaluation of each word’s importance. According to the idea captured in the TF/IDF measure. The importance of a word W is dependent on how often it occurs within the body of text, and how often the word occurs within a larger collection that the text is part of. Intuitively, a word in given text will be considered as the most important one if it occurs frequently within text, but infre -quently in the larger collection. The formula is shown as follows:

Wij = Tfij * log2 N/n whereWij = weight of term Tj in document Di

Tfij = frequency of term Tj in document Di

N = number of documents in collectionn = number of documents where Tj occurs at least once

(3)

The documents of the formula must be modified as an individual package. The N of the collection in our case is learning content database and the parameter n in this formula requires knowledge of all words within the collection that holds the text material of interest.

For calculating each word’s importance, we need to construct a dictionary that contains the information of how frequently it occurs across course packages in learning content database. To construct our learning dictio-nary, we identified word frequencies across all course packages that we had previously stored in our database.

Fig. 7 shows each step of constructing the dictionary of weight words from our learning content repository. It begins from the content parser which fetches learning courses from the repository and extracts all the words from each course, unless the frequent stopped words such as “is”, “are”, “and”, etc.

Fig. 7. A Construct a dictionary of weighted words

Then, each uniquely extracted word will be tagged by a counter module with a number and keeps track of the number of courses where the word occurred. Once the counting is complete, the words that occurred less than a chosen threshold value across all the courses are eliminated. The value is required to be tuned because it depends on the size of the repository. It would conserve too many insignificant words if the value is too large. On the contrary, it is probable to remove rare words that may quite important and have the potential to become keywords. The remaining words are passed through a spell checker and finally, words that have the same grammatical stem are combined into single dictionary entries. For example adaptive and adapted, would share an entry in the dictionary. Accordingly, the size of the dictionary will continually shrink.

When the significant keywords must be extracted from a PU, all the words in the PU are stemmed. For each word, the module will search the dictionary to discover the frequency with which the word occurs in the course. The word’s frequency within the course package that contains the PU is found by scanning the course package in real time. Finally, these values are computed for the word’s TF/IDF weight. Words with a weight beyond the chosen threshold are selected as significant.

A special situation arises when a word is not in the dictionary, either because it was discarded during our dictionary-pruning phase or it was a specific word that has never been shown in other learning contents. Such words are probable rarer than any of ones that survived pruning and were included in the dictionary. Therefore these words are considered as special keywords in this course and as important as any of the words we re -tained.

Finally, notice that our implementation directly extract and store the words as important ones if they are somehow highlighted with bold, italic, different color, specific punctuation marks, etc.

3.3.2.2 Summary Sentence Extraction

Rather than summarizing the input text automatically, we can only pick up a few significant sentences to rep -resent the text summary. Because of the previously revealed keywords in a PU a user intends to explore the portion of the content’s summary due to his/her interesting. A sentence will be intuitively considered signifi -cant if it contains one or more keywords. Therefore, the method of extracting summary sentences is based on keyword extraction result.

37


Fig. 8. Processing of the learning content adaptation

The procedure of summary sentence extraction is as shown in Fig. 8. Each sentence in a PU will be ex-tracted by a sentence dividing manager, and then passed to the summary generator. Meanwhile, the previous extracted keywords are also passed through the summary generator in which each sentence will be extracted and listed in order if it contains the matched keywords.

3.3.3 Image Adaptation Model

Similarly, image adaptation replies on comparing its size to a SU’s. Recall the definition of SU. It is a rectan -gle displaying unit and its presentation area is the same as a physical screen area of required handhelds. An image may not be able to adapt its size to perfectly match the proportion of a SU and reside in it. Hence, a large image (its height and width are all exceed a PU’s) might be shrunk proportionately until its width is fit to display in a PU without additional horizontal scrolling action. The entire adapted content will have the default displaying in a single column where vertical scrolling for browsing is necessary, so the height of a image be-yond a SU’s is acceptable. The height of a PU is consequently extended for the image in our implementation. Accordingly, a small image will be enlarged proportionately until its width, is at least fit, to display in a PU without additional scrolling actions.

3.3.4 Layout Adaptation Model

The presentation adaptation mode provides two main functions for user to reedit the content’s layout. One is allowing users to pick up PUs to delete, the other is let users rearrange PUs’ position manually. The procedure is as illustrated in Fig. 5. After the previous automatic adaptation, each PU should contain appropriate display -ing content-content that has either been adapted or not. A few PUs might be required to be eliminated, because they may present relatively insignificant objects such as pictures decorated only for aesthetic purposes in its original content. Besides, users may decide to delete certain PUs considered unnecessary for learning accord -ing to their editing experience, and consequently reduce the content’s size. On the other hand, the default adapted layout rearranges PUs in an orderly single column presentation. Kärkkäinen mentioned a design guideline for a small display screen; put as much important content as close to the top of the hierarchy as pos-sible. Each PU is a unit of presentation, such that the PU deployment manager allows users to manually rear-range each PU’s displaying position to satisfy user’s specific requirement and construct a preference layout.

3.4 Course Rich-presentation Packaging/Delivering Module

The purpose of Course Presentation Adaptation Module is producing the adapted content that can be appropri -ately displayed on specific mobile learning platforms. Therefore after the editing phase, the author may create plural content versions such as pocket PC version and smartphone version. Therefore, Course Rich-presenta-tion Packaging/Delivering Module allows the author to save the produced plural course versions as a rich-pre -

38


sentation course package conforming Common Cartridge Standard. Meanwhile, the rich-presentation informa-tion is also properly described in cartridge meta-data file in which the available device versions are specifi -cally listed. Thus, LMS can recognize whether the course exist a corresponding presentation version to the leaner who is requesting the course for learning.

3.4.1 Producing and Delivering a Rich-presentation Course Package

The Course Rich-presentation packaging utilizes the significant characteristic of Common Cartridge to com -pose a rich-version course package. The Fig. 9 shows the Common Cartridge file structure having a learning application object folder that includes three different SCORM course versions for corresponding learning plat -forms: pocket PC, smarthpone and regular PC.

Fig. 9. An example of rich-presentation course package

For the sake of recording the information of existing versions of a rich-presentation course package, we add a tag <coverage> that is defined at the nine categories of IEEE LOM Metadata. In the first category, “Gen -eral”, that includes the related information of the learning object such as regular data, title, keywords and iden -tity.

In terms of describing a learning object, most general information are recorded at this category. Briefly speaking, the object information of this part of metadata is traced at the process beginning. There is a subclass namely coverage, which is defined to describe the proper learning scopes of this object. According to the char -acteristic of the subclass, we insert the information of different presentation versions existed in this rich-pre -sentation course package. The Fig. 10 illustrate the metadata of the rich-presentation course package in which the tag <coverage> attached R (regular PC version, PDA (PDA version) and SP (smarthpone version) that indi-cates the course package can offer the three different presentation versions. In the manifest file, it also illus -trates the accessing path of the three different presentation version stored in the back-end server. System can trace the resource location with the <identifier> to retrieve the corresponding terms in <resources> tag.

Fig. 10. The imsmanifest.xml in the Rich-presentation course package

39


5 Experiment Results

In this section, we focus on evaluating the performance of the proposed course caching strategy and verify the great improvement on system accessibility and immediacy for mobile learners. Firstly, we design a simulator that allows the user to input the relevant parameters to simulate an m-learning environment such as the spe-cific handheld’s specification and wireless network conditions, as well as the user can monitor and trace how the caching process manipulates course clusters according to the customized configures space step by step.

5.1 Simulator

In order to evaluate the performance of the course caching strategy, we developed a simulator that consists of the following six components as shown in Fig. 11.

Fig. 11. Simulator components

1. Device specificationThe user can select a specific handheld device to make evaluation, where the hardware specifications have been built into the program in advance. Alternatively, the device profile can be edited or created by the user to modify the default values such as CPU frequency, storage size, and network capability. Furthermore, the user also allows deciding the whole simulation process will execute under which caching mode; VMM or COD mode.

2. Execution speed controlThe control allows users to adjust the simulation execution speed so that they can observe the internal clus-ter operations and storage usage.

3. Learning order creation modelThe model can produce many possible learning orders according to the testing course structure and associ -ated SCORM sequencing information. The model associates with an advanced learning order creation appli -cation that can produce the valid and possible learning order according to SCORM S&N information at -tached in the selected course. Therefore, the user is allowing selecting the SCORM-compliant course to ex -amine the caching performance. After loading the selected course, the learning order creator will parse the manifest file in which describe the whole information about sequencing information then retrieving the con -figure parameters to calculate the possible learning order.

4. Storage monitorIt is a graphical interface that dynamically displays the status of the current storage usage.

5. Cluster informationIt is used to build a simple SCORM course aggregation tree with cluster index that illustrates the swapping operation of the current status.

6. Statistical analysis dataThis area shows the statistical information of the experimental results that include the size of the current accumulated download contents, size of the dropped content, waiting time, network reconnection frequency, and hit ratio.

5.2 Testing Courses

40


In this section, we will introduce the courses which are adopted to evaluate the caching strategy. In our experi -ment, using three test courses—two standard HTML-based courses released by ADL and a customized course that includes various multimedia resources produced by our authoring tool. The features of these courses are as follows.Course 1: Photoshop_None HTML-based learning content Simple course cluster structure Without SCORM sequencing informationCourse 2: Photoshop_KnowledgePaced HTML-based learning content Normal course cluster structure With SCORM sequencing informationCourse 3: Customized Course Hybrid learning content Complex course cluster structure With SCORM sequencing information

Five evaluation modes were used: No caching strategy (NONE) COD with FIFO (first in first out) algorithm (COD-F) COD with our algorithm (COD-O) VMM with FIFO algorithm (VMM-F) VMM with our algorithm (VMM-O)

5.3 Results

Criterion of COD mode: The aim of this mode is to avoid wastage of time for displaying content due to limited network bandwidth, e.g., wastage that occurs when a smartphone with GPRS is used as the learning platform. Therefore, reduction in the latency during the learning process is the primary task of this mode. In other words, the criterion of this mode is just the waiting time.

Criterion of VMM mode: This mode mainly addresses the suspending due to the intermittent network con-nection such as using a Pocket PC with WiFi as learning platform. Thus, a decrease in the frequency of Inter-net reconnection is the criterion of this mode.

Fig. 12. Total Waiting Time (s) for Learning Pho-toshop_None Course

Fig. 13. Internet Reconnection Frequency for Learn-ing Photoshop_None Course

Testing Course 1: Photoshop_NoneThis sample course released by ADL is composed of a few simple small clusters without SCORM se -quencing information. Fig. 12 and Fig. 13 show the COD mode and VMM modes, respectively; significant improvements are observed in the waiting time and request for internet reconnection. However, it can be inferred that the proposed algorithm is very similar to FIFO under the course condition.

Testing Course 2: Photoshop_KnowledgePacedThis test course has also been released by ADL. It has a normal course structure and appropriately associ -ated sequencing information for each cluster. In this case, the course caching strategy represents the best improvement as compared to that in the other evaluated courses. As shown in Fig. 14, the COD-O event reduces the latency period by 95% during the learning process. As shown in Fig. 15, the VMM-O also sig -nificantly decreases the latency period by 74% and request numbers of network reconnection. Although the performance is much better than that of NONE, the execution result is still similar between the caching models adopted by the FIFO method and our proposed algorithm.

41


Fig. 14. Total Waiting Time (s) for Learning Photoshop_KnowledgePaced

Fig. 15. Internet Reconnection Frequency for Learn-ing Photoshop_KnowledgePaced

Testing Course 3: Customized CourseIt is the third test course created by a SCORM authoring tool. It expands the sample course released by ADL by adding various multimedia contents and relatively complicated sequencing information. The course structure is certainly more complex than that of the previous two courses. The entire course could be separated into many independent or dependent clusters, which include large-sized files such as images or videos. This case represents the performance difference between FIFO and our algorithm adopted in the caching modes. The Figure 16 shows that the COD-O mode significantly reduced 29% waiting time better than the COD-F mode 15% waiting time. Figure 17 shows the VMM-O performance shown an improve-ment of 57% over that of NONE and approximately 6% over that of VMM-F.

Fig. 16. Total Waiting Time (s) for Learning Customized Course

Fig. 17. Internet Reconnection Frequency for Learn-ing Customized Course

5.4 Discussion

From the experimental results, we found that the proposed caching strategy shows the best improvement when the applied course is composed of a normal course aggregation structure and explicit SCORM sequencing in-formation. Both the COD and VMM modes have very similar performances irrespective of the algorithm adopted as shown in Fig. 18 and Fig. 19.

Fig. 18. Accuracy of VMM Course Caching Strategy during Learning Customized Course

Fig. 19. Accuracy of COD Course Caching Strat-egy during Learning Customized Course

One of the possible reasons that influences the performance of our prediction accuracy might be cased of the user’s learning orders are created by a simulator based on only the SCORM sequencing rules. These testing orders may not violate the rules of SCORM sequencing but no other consideration of user’s learning behaviors and records. Hence, some of the created learning orders may not agree with the normal human reading behav-iors and they may make the important factors for the learning records useless in our algorithm. However, if the assigned course includes various multimedia resources and a relatively complicated course aggregation struc -ture, our caching algorithm gradually performs better than the FIFO method.

42


6 Conclusion & Future Works

The main contributions of our research are exploring the two adaptive strategies: course caching and course presentation to facilitate efficient m-learning systems conformed to international distance learning standards. The demonstration learning systems based on proposed m-learning modules showing the significant learning performance compared to traditional learning environment in which the learning process may be degraded or interrupted due to difficult network situations. On the other hand, we propose the concept of adaptation tem-plate to facilitate the presentation transformation of html-based learning materials composed of texts and im-ages. Instead of real-time adapting general Web pages on the Internet, we believe automatic and manual adap -tations are equally important for learning contents. Therefore, our course presentation adaptation authoring tool allows authors can not only utilize the proposed adaptation templates automatically and efficiently to adapt the learning content for handhelds but also adjust the template parameters to influence the adapted re -sult, as well as the educational quality is assured by authors themselves. To implement our proposed method-ologies, we believe that could help the raise of the population of m-learning market.

The future work issues we considered include what types of context information are helpful for us and how to improve the course caching performance with retrieving significant information such as learning history, learning progress, user’s background, user’s preference, and user location. Meanwhile, we are upgrading the customized version of Pock SCORM with the course caching strategy to facilitate the mobile learning project for employee training in domestic airways. Although, the mobile devices are becoming more and more power -ful, there still have been a lot of various multi-media resources can not be displayed properly on handhelds such as flash files and specific format videos. As a result, how to adequately adapt or represent the multi-me -dia resources included in learning contents is our another main future work.

References

[1] O. Buyukkokten, H. Garcia-Molina, A. Paepcke, “Accordion Summarization for End-game Browsing on PDAs and

Cellular Phones,” The SIGCHI Conference on Human Factors in Computing Systems, pp. 213–220, 2001.

[2] H. P. Chang, C. C. Wang, K. H. Jan, T. K. Shih, “SCORM Sequencing Testing for Sequencing Control Mode,” in

Proceedings of the 20th Advanced Information Networking and Applications Conference (AINA), pp. 164-169, 2006.

[3] H. P. Chang, W. C. Chang, Y. L. Sie, N. H. Lin, C. H. Huang, T. K. Shih, Q. Jin, “Ubiquitous Learning on Pocket

SCORM,” The Second International Symposium on Ubiquitous Intelligence and Smart Worlds(UISW) , pp. 85-89,

2005.

[4] Y. Chen, W. Y. Ma, H.J. Zhang, “Detecting Web Page Structure for Adaptive Viewing on Small form Factor De-

vices,” in Proceedings of the 12th International Conference on World Wide Web, pp. 225-233, 2003.

[5] G. Buchanan, S. Farrant, M. Jones, H. Thimbleby, G. Marsden, and M. J. Pazzani, “Improving Mobile Internet Us-

ability,” in Proceedings of the 10th International Conference on World Wide Web, pp. 673-680, 2001.

[6] R. Han, P. Bhagwat, R. LaMaire, T. Mummert, V. Perret, J. Rubas, “Dynamic Adaptation in An Image Transcoding

Proxy for Mobile Web Browsing,” in Proceedings of the IEEE Personal Communications, pp. 8-17, 1998.

[7] L. Kärkkäinen and J. Laarni, “Designing for Small Display Screens,” in Proceedings of the 2nd Nordic Conference

on Human-computer Interaction, pp. 227-230, 2002.

[8] E. Kaasinen (2000), “Two Approaches to Bringing Internet Services to WAP Devices,” in Proceedings of the 9th In-

ternational Conference on World Wide Web, pp. 231-246, 2000.

[9] S. A. Kazi, “A Conceptual Framework for Web-based Intelligent Learning Environments Using SCORM-2004,” in

Proceedings of the IEEE International Conference on Advanced Learning Technologies, pp. 12-15, 2004.

[10] R. H. Katz, “Adaptation and Mobility in Wireless Information Systems,” in Proceedings of the IEEE Personal Com-

munications, pp. 6–17, 1994.

[11] T. Kindberg and A. Fox, “System Software for Ubiquitous Computing,” The IEEE Pervasive Computing, 2002.

43


[12] L. Ramaswamy, A. Iyengar, L. Liu, F. Douglis, “Automatic Detection of Fragments in Dynamically Generated Web

Pages,” in Proceedings of the 13th International Conference on World Wide Web, pp. 443–454, 2004.

[13] S. T. Li and C. H. Lin, “On the Distributed Management of SCORM –Compliant Course Contents,” in Proceedings

of the 19th International Conference on Advanced Information Networking and Applications , pp. 221-226, 2005.

[14] N. H. Lin, T. K. Shih, H. H. Hsu, H. P. Chang, H. B. Chang, W. C. Ko, L. J. Lin, “Pocket SCORM,” in Proceedings

of the 24th International Conference on Distributed Computing Systems, pp. 274-279, 2004.

[15] I. Mohomed, J. C. Cai, S. Chavoshi, E. de. Lara, “Applications: Context-Aware Interactive Content Adaptation,” In-

ternational Conference on Mobile Systems, Applications and Services MobiSys, pp. 42 – 55, 2006.

[16] J. Malek, M. Laroussi, A. Derycke, “A Multi-Layer Ubiquitous Middleware for Bijective Adaptation between Con-

text and Activity in a Mobile and Collaborative learning,” in Proceedings of the International Conference on Systems

and Networks Communication, pp. 36 – 39, 2006.

[17] D. Narayanan, J. Flinn, M. Satyanarayanan, “Using History to Improve Mobile Application Adaptation,” in Proceed-

ings of the 3th IEEE Workshop on Mobile Computing Systems and Applications, 2000.

[18] B. Orkut, K. Oliver, G. M. Hector, P. Andreas W. Terry, “Efficient Web Browsing on Handheld Devices Using Page

and Form Summarization,” Journal of ACM Transactions on Information Systems, 20(1), pp. 82-115, 2002.

[19] D. Su and K. Seong, “Usability Guidelines for Designing Mobile Learning Portals,” in Proceedings of the 3rd Inter-

national Conference on Mobile Technology, Applications & Systems, 2006

[20] A. Syvanen, R. Beale, M. Sharples, M. Ahonen, P. Lonsdale, “Supporting Pervasive Learning Environments: Adapt-

ability and Context Awareness in Mobile Learning,” in Proceedings of the IEEE International Workshop on Wireless

and Mobile Technologies in Education, pp. 28-30, 2005.

[21] J. Su, S. Tseng, J. Weng, K. Chen, Y. Liu, Y. Tsai, “An Object Based Authoring Tool for Creating SCORM Compli-

ant Course,” The 19th International Conference on Advanced Information Networking and Applications , pp. 209-214,

2005.

[22] T. K.Shih, W. C. Chang, N. H. Lin, L. H. Lin, H. H. Hsu, C. T. Hsieh, “Using SOAP and .NET Web Service to build

SCORM RTE and LMS,” in Proceedings of the 7th International Conference on Advanced Information Networking

and Applications, pp.408-413, 2003.

[23] Y. Hwang, J. Kim, E. Seo, “Structure-Aware Web Transcoding for Mobile Devices,” in Proceedings of the IEEE In-

ternet computing, pp. 14–21, 2003

[24] T. H. Tan and T. Y. Liu, “The Mobile-based Interactive Learning Environment (MOBILE) and A Case Study for As-

sisting Elementary School English Learning,” in Proceedings of IEEE International Conference on Advanced Learn-

ing Technologies, 530-534, 2004.

[25] Y. K. Wang, “Context Awareness and Adaptation in Mobile Learning,” in Proceedings of the 2nd IEEE International

Workshop on Wireless and Mobile Technologies in Education, pp. 154-158, 2004.

[26] J.T. D. Yang, C. H.Chiu, C. Y. Tsai, T. H. Wu, “Visualized Online Simple Sequencing Authoring Tool for SCORM-

Compliant Content Package,” in Proceedings of the IEEE international Conference on Advanced Learning Technolo-

gies, pp. 609-613, 2004.

[27] G.Zhao and Z. Yang, “Learning Resource Adaptation and Delivery Framework for Mobile Learning,” in Proceedings

of the 35th Annual Conference on Frontiers in Education, F1H-18 - F1H-24, 2005.

44


[28] Y. Zhou, J. Tang, J. Wang, “An Improved TFIDF Feature Selection Algorithm Based On Information Entropy,” in

Proceedings of the Control Conference, pp. 312-315, 2007.

45

Lecture Notes in Computer Science: Vol_20_No_2.files/JO… · Web viewJason C. Hung, Kuo-Feng...

Documents

Transcript of Lecture Notes in Computer Science: Vol_20_No_2.files/JO… · Web viewJason C. Hung, Kuo-Feng...