Adaptive Content Delivery System for Ubiquitous Learning Xinyou ZHAO, Toshie NINOMIYA, Mikko VILENIUS, Fumihiko ANMA, Toshio OKAMOTO Abstract Most of learning contents, designed for desktop computers and high-speed network connections today, are not suitable for handheld devices with limited resources and computing capabilities. Also, some delivered materials, irrelevant to learners’ preferences or contextual environment, may slow down learning efficiency, and increase learners’ communication costs and channel burdens. In order to provide adaptive contents based on device capabilities and learner’s context awareness, this paper proposes a functional architecture for ubiquitous learning. The architecture consists of four engines: learner engine, detector engine, adaptation engine, and transcoding engine. After contextual data and learners’ preferences are separately identified by detector engine and learner engine, adaptive contents created by transcoding engine can be dispensed to learners by adaptation engine. Finally, the adaptation content delivery system is experimentally evaluated on standard document learning contents. Keywords: Content Adaptation, Context Awareness, Adaptive Learning, Pervasive Computing

1. Introduction The growing diffusion of mobile devices coupled with Internet access features has improved learning flexibility and quality(1,2,3). Learners may study anyplace, anytime, and with mobile device in a ubiquitous learning environment, which means that the applications and services normally conducted on personal computers should be also applicable on handheld and portable devices. But up to now, access to e-learning contents designed for desktop platform by mobile devices has not become as convenient as expected with mobile browser embedded in mobile devices. The problem is that most of e-learning contents, designed for desktop computers and high-speed network connections, are not suitable for handheld devices, whose capabilities are usually limited in terms of network bandwidth, processing power, storage capability, markup language, and screen sizes, etc(4,5,6,7). Another problem is that massive amount of contents, irrelevant to learners’ preferences or contextual environment, will make learners feel frustrated and dissatisfied (5,8,9,10). They also increase the learners’ communication costs and channel burdens. Consequently, when delivering e-learning contents to mobile learners, it is necessary to suit the context environment (such as device features, network characteristics, location) & preferences of learners in question. The procedure of altering/recoding learning contents in this way to enhance the user experience on particular devices is referred to as content adaptation(3,7,11,12). In regarding to the location where

the adaptation is performed, W3C(4) categorizes three types of content adaptation: server-side approach, proxy-based approach and client-side approach. In this research, we focus on the server-side approach, the major benefit of which is that the server can compute the characteristics of learning contents and preference & needs of learners(3,4,6,11). Current systems on server side only consider the contextual data of learners separately, such as mobile device(1,2,6,11) or learning history(9,10,13). Delivering learning resources designed for tabletop computers to ubiquitous learners is by no means a trivial task. The objective of our proposed solution is to provide adaptation contents in a common dynamic and extensible framework in order to have scalable, flexible and extensible architecture. Contextual data are comprehensively considered with learners’ preference. This will be done by defining an adaptation scheme that recodes/reconstructs learning contents according to context awareness of learners. The reconstructed contents are adapted to learners’ preference, also adapted to learners’ context awareness. The rest of this paper is organized as follows. Section 2 gives a brief overview of related works. Section 3 describes the system architecture we considered, and the detailed descriptions of four engines. A simulation based on PowerPoint file is discussed in Section 4. Concluding remarks and future works are presented in Section 5.

2. Related Works There have been intensive researches on content adaptation on server side, which involve creating

studies, e.g. (6) and (15), have also adopted caching strategy to reduce the transcoding time. When caching content stratifies with transcoding requirement, the server returns the caching contents directly. But this method also increases the storing burden of server. Learners’ context awareness is also considered by (3) and (16). These methodologies can combine the learning infrastructure (educational activity, networks, learning state, etc) to identify the interest & preference of learners and create adaptive contents for learners in a ubiquitous environment.

different suitable content for rendering by mobile device and mobile users from original content such as a web page. Adaptive content delivery can be classified into two main types according to the time when these different content variants are created(1,6): static adaptation and dynamic adaptation. In static adaptation, the server analyzes the learning context and returns the best alternative version from different pre-adapted contents, or filters the original contents (e.g. removes an image or a video)(2). With static adaptation approach, content provider can have a tight control over what is transcoded and how the result is presented(1,6,13). In dynamic adaptation, the desired contents are adapted and delivered on the fly according to dynamic requirements presented to the server. The requirements are based on the current characteristics of the learning environment(6). Many approaches have been proposed based on dynamic adaptation on server. (13) provides personalized contents by analyzing the learning log information. In (9) and (10), colony system was adopted to achieve adaptive contents for learners. Based on an ant colony optimization algorithm, style-based and attributes-based methods help learners find an adaptive learning object more effectively. Furthermore, (7) and (11) proposed adaptation content algorithms to parse the HTML contents for mobile users. In order to keep the characteristic of contents, (1) and (12) present a QoS aware adaptation scheme of multimedia contents to reduce the bandwidth required for delivery. They can adapt multimedia contents in accordance with learner, system and network state to meet end-to-end requirements. Some

3. System Architecture This paper proposes adaptation content delivery architecture for ubiquitous learning environment, which contains four core components: Learner Engine, Detector Engine, Adaptation Engine, and Transcoding Engine. The architecture is shown in Fig.1. Firstly, Device Detector Module in Detector Engine uses http request headers (partial context awareness from mobile device) and device/browser repository to detect capabilities of mobile device. The capabilities will be used to create adaptive contents by Learner Engine and Adaptation Engine. Secondly, Learner Engine recommends adaptive contents for learners. If contents accessed by learners are not suitable for the mobile learning environment, such as a big video or image not supported by mobile device, Learner Engine should recommend a substitute or reconstruct the contents for learners. After analyzing learners’ profile and learning experience under current context awareness,

Ubiquitous Learning Interface (IE, Java, Mobile Web Browser...)

Learner

Adapted Contents Adaptation Engine

Transcoding Engine Text Transcoder

Image Transcoder

Database

Audio Transcoder

Video Transcoder Server

Standard Documents Transcoder ( PPT, DOC, PDF, HTML,…)

Conversion Rules

Content Cache

Data Stream

Learning Context Detector Engine

Markup Language Module

Device Features

Device Detector Module

Learner Engine

Request

Content Negotiation Personalized Contents Module Response

Learning Object

Learner Experience

Access Database

Learner Profile

Database

Personalized Content Module

Browser Repository Running Module

Figure 1 Adaptation Content Delivery Architecture

Device Repository

Personalized Content Module recommends the adaptive contents for learners. Thirdly, Content Negotiation Module in Adaptation Engine negotiates the contents based on device capabilities and network characteristics (context awareness) and recommends constrained version (XML format) of contents. Then, a transcoding request will be sent to Transcoding Engine, which transcodes media into adaptive content based on constrained version of contents. Fourthly, Markup Language Module adds suitable markup language (HTML, WML, XHTML MP or cHTML) with adapted contents after Content Negotiation Module gets the transcoding results from Transcoding Engine. For example, contents of text and image will be recommended when the learner is staying at a noise station. On the contrary, the audio content (should delete texts and images from original content) will be provided while the learner is walking. When the format of image or audio is not supported by mobile device, Transcoding Engine recodes the content by the request from Content Negotiation Module, such as format conversion of image from jpg to gif.

3.1. Detector Engine Detector Engine takes responsibility for detecting capabilities (memory, screen size …) of the client device. The WURFL (Wireless Universal Resource File) (17) model is adopted to define the features of devices and mobile browsers. WURFL is an XML configuration file which contains information about capabilities and features of many mobile devices in the wireless world. However, as there are many different kinds of mobile devices, and new ones appear on the market on an almost daily basis, it is impossible for Detector Engine to keep an inventory of all of these devices and configurations, even for just one mobile service carrier. Therefore, it is important for learning system to provide a mechanism to produce the substitute of capabilities for mobile devices in the case of detection failures of the system. This mechanism is divided into two steps: Step I: Retrieve features from device/browser repository by matching the value in HTTP_USER_AGENT header, which is sent with the http headers from mobile device. The system adopts the algorithm of Levenshtein Distance to match this value to the device repository (Shown in Table 1). Not only the features of the phone, but also the features of the device’s browser are important when adapting contents. In fact, we may use the features of

the browser instead of the features of the mobile device when we can not detect the mobile device (that is, when it is not found in the device repository). Table 1 Levenshtein Distance Input input s and t; // two strings of User Agent set lengths of s, t to n,m; construct a Matrix[n+1,m+1]; Output Matrix[n,m]; S1: initialize the first row, column to 0..n, 0..m; S2: foreach character s[i] in s (i from 1 to n) S3: foreach character t[j] in t (j from 1 to m) S4: if(s[i] == t[j]) cost = 0; S5: else cost =1; S6 endif S7: a= Matrix [i-1,j] + 1; S8: b= Matrix [i,j-1] + 1; S9: c= Matrix [i-1,j-1] + cost; S10: set Matrix[i,j] = Min(a, b, c); S11: endfor S12: endfor

Again we use Levenshtein Distance to match the client device’s browser to the browser repository. For example, based on HTTP_USER_AGENT header (Vodafone/1.0/V705SH/SHJ001 Browser/VF-NetFro nt/3.3 Profile/MIDP-2.0 Configuration/CLDC-1.1) from Softbank 705SH, the system can detect that the browser is VF-NetFront, whose version is 3.3. Although the system can detect or substitute the features of mobile device according to HTTP_USER_AGENT from http request, the system only gets most similar or static capabilities of mobile device at most times. In addition, we may get other headers besides HTTP_USER_AGENT, such as markup language and format of images supported, size of screen, etc. We can get more accurate capabilities of mobile device dynamically based on other headers from http request. Step II: Get more accurate capabilities according to other headers from http request. Different mobile service providers provide different header information for the same feature. For example, HTTP_X_JPHONE_DISPLAY is used for screen size by Softbank, the third largest mobile service provider in Japan, and HTTP_X_UP_DEVCAP_SCREENPIXELS by KDDI, the second largest mobile service provider in Japan. Here, we define different rules to detect features based on different header information. Enhanced algorithm is proposed in Table 2.

3.2. Learner Engine Because of limited features of mobile device and wireless network, we should consider whether

contents can be correctly used or displayed under learning context awareness, which may be learner state (such as aims, interest, location…), educational state (requirements, purpose, results…), and learning environment (weather, place, neighbor, device...). Table 2 Enhanced Algorithm S1: input capability; // get based on user agent S2: get header rules based on capability S3: foreach header[i] in headers S4: if (header[i] in rule) S5: modify capability based on rule S6: else S7: submit to database for future improving S8: endif S9: endfor

We have developed the WebClass RAPSODY e-learning system (13), which recommends the personalized contents for learners by analyzing past learning history stored in learning experience database. In this research, Personalized Content Module recommends personalized contents for learners by combining context awareness with our previous recommendation system in (13). Here, we mainly consider the network characteristics and device capabilities with learning experience as context awareness. Personalized Content Module analyzes the learning style and learning context awareness and recommends adaptive contents. The Ui(t) in Formula I is used to define the ubiquitous learning state at time t for learner i: Ui(t)= [Li (t), Ci (t), Ei (t)] Formula I where: Li(t): the learner’s preference at time t Li(t) = (wi1, wi2, …, win) where, each wij in vector represents a weight for contextual data or content type (such as to image); n is total count of contextual data related to learners. Ci(t): items state in learning content (C) at time t Ci(t) = ( I1, I2 ,…, In) where, Ii is the features (such as format, size) of each item in learning content (C); n is total different item in the learning content. For example, a learning content containing a paragraph text and an image is defined C = (I1, I2), I1 and I2 are features of TEXT and IMAGE separately. Ei(t): context awareness at time t Ei(t) = (li, ti, ui, di, pi, mi) where, li: location (campus, car, train, road, classroom, home); ti: time; ui: upload speed by mobile device; di: download speed by mobile device; pi: network type (GSM, 3G, IEEE802); mi: mobile device type (PC, PDA, Smartphone, Mobile Phone).

Personalized Content Module uses Dice Coefficient to recommend the personalized contents for current ubiquitous learning state (Ui(c)) based on past learning state (Ui(p)). The personalized contents filtered from original contents, such as deleting big video or image, are provided according to the learner past experience and context awareness. Dice Coefficient is a correlation coefficient algorithm for discrete events (Formula II). Dice (U i (c ),U i ( p )) = @

å

å n k =1

n k =1

2 P (U i (c),U i ( p)) P (U i (c)) + P (U i ( p ))

Formula II

wkc wkp

w + å k =1 w 2 kc

n

2 kp

where, P(Ui(c)) and P(Ui(p)) are the probabilities of learning states Ui(c) and Ui(p) separately; P(Ui(c),Ui(p)) is the combined probability for Ui(c) and Ui(p); wkc and wkp are the state variables from Li (c), Ci (c), and Ei (c), and Li (p), Ci (p), and Ei (p), respectively.

3.3. Adaptation Engine Adaptation Engine contains Content Negotiation Module and Markup Language Module. After getting personalized contents and feature of mobile device, Content Negotiation Module matches personalized contents with features of the device. Content Negotiation Module negotiates personalized contents and creates a request for unsuitable items of contents with mobile device to Transcoding Engine, which takes responsibility to complete the transcoding request and then return a transcoding response to Content Negotiation Module. Please refer to (3) for negotiation algorithm in details. Finally, Markup Language Module sends adapted content, which is reconstructed by embedding markup language (HTML, WML, XHTML MP or cHTML) into transcoded & untranscoded contents from Content Negotiation Module. A transcoding request is an XML file, which contains one or more transcoding jobs. For this, OMA standard specification model(18) is adopted. Each transcoding job contains a source and a target. The source represents input a media file (stored on local server) and its characteristics and the target represents the desired output characteristics. Corresponding to request, transcoding response is also a XML file, which contains one or more transcoding result, media location, etc. The model is shown in Fig.2.

3.4. Transcoding Engine Transcoding Engine has many transcoders (also called media processor), each of which takes responsibility to transform one kind of media, such as video transcoder for video, image transcoder for image. After Transcoding Engine has parsed the

Transcoding Job Source -Content Type -Location -Content Attribute -... Target -ProfileID -Location -Transcoding Param eters -...

Transcoding Result -Job ID -Job Return Code -Return String -... Output -Content Type -Location -Size -...

Figure 2 Descriptions of Job & Result request (XML file) sent by Adaptation Engine, it submits the request to the related transcoder. The transcoding procedure is divided into two steps: cache detection and real-time transcoding. Step I: The transcoder checks whether media satisfying request job already exists in the content cache. If yes, it will return that media’s information to Adaptation Engine. Otherwise, it goes into Step II. Step II: The transcoder accesses the original contents, conversion requirement and rules from conversion rules server and then recodes or reconstructs original content into target contents, which are stored in content cache Server. Finally, transcoding results are sent to Adaptation Engine.

4. System Application for PowerPoint Slides According to adaptation content delivery architecture discussed above, a prototype system is developed by PHP5.2.1 language and MySQL5.1 database on Apache2.0.63 HTTP server. The prototype system adopts FFMPEG(19) and ImageMagick(20) to transform the video & audio and image separately based on constrained features from transcoding request. As for standard documents, we have developed a document transcoder: DoCAgent, developed by C# on Visual.net 2005, which converts a standard document into HTML pages, which are stored on same server in parallel with the original document. We evaluate PowerPoint presentation contents on prototype system. DoCAgent creates an HTML page for each slide of PPT file. These HTML pages are divided into two categories: index page and sub-pages. Index page has indexes of all sub-pages (each sub-page for one slide), indexed by their titles (Fig. 3 (a)). When ubiquitous learner accesses a standard document: PowerPoint files, which may contain text, image, audio, animation and video, Learner Engine recommends the personalized contents from a sub-page (one slide) based on learners preference and context awareness, such as deleting image (may not be favorable to learner) or audio (may not be

supported by device). At last, some items of pages are recommended to Adaptation Engine. If device does not support a web browser, Adaptation Engine only sends a transcoding request for text or limited media to Transcoding Engine. At last, Adaptation Engine pushes contents to learners in SMS (only text) or MMS (limited media contents). If a mobile browser is embedded in handheld device, Adaptation Engine negotiates items of personalized contents (sub-page) and sends a transcoding request for unsuitable media. Then, Markup Language Module creates adaptive a mobile page by replacing the HTML tags with mobile tags (such as cHTML, WML). Lastly, mobile page is sent or video or audio content is pushed to the learner. Of course, Transcoding Engine should consider rules for HTML pages, such as text information should be less 160 characters for SMS. Table 3 shows conversion rules for HTML contents. Table 3 Conversion Rules for HTML File Browser Media Transcoding Size Text SMS <=160 char Not support Limited MMS <=30K Limited Mobile Web <=30K Support All — Original File

Snapshots in Fig.3 show some results for one presentation of content for a certain course, which contains 11 slides in PPT files. Fig.3 (a) and Fig.3 (b) are displayed on Softbank Sharp 705H; Fig.3 (c) is displayed on Simulator of OPENWAVE 7.0; Fig.3 (d) is displayed on Dell x50V by IEEE802.11n.

5. Conclusions & Future Works Ubiquitous devices can complement e-Learning or traditional classroom learning. The problem is that certain kind of contents may not be supported by handheld device, which presentation capabilities are limited. Also, a massive amount of irrelevant content presented on handheld devices will make learners feel frustrated and dissatisfied. In this paper, we proposed algorithms for detecting device capabilities and identifying learners’ preferences. According to context awareness of learners, the adaptive content delivery system transcodes original content into adaptive contents. The caching method also introduced reduces the transcoding time on the fly. Lastly, we present an adaptation evaluation on PPT learning content. In future, the system will be applied for an actual university course. The results yielded by this practical experiment will be used to adjust the system for better performance.

(a): Indexes on Phone

(b): One Slide on Phone (c): One Slide on Simulator (d): One Slide on PDA Figure 3 Snapshots on Mobile phone, Simulator, PDA

References (1)

(2)

(3)

(4) (5)

(6)

(7)

(8)

(9)

Rho, S.M., Cho, J.W. and Hwang, E.J.: “Adaptive Multimedia Content Delivery in Ubiquitous Environments”, Proceeding of the International Workshops on Web Information Systems Engineering, pp. 43 - 52 (2005). Kojiri, T., Tanaka, Y. and Watanabe, T.: “Device-independent Learning Contents Management in Ubiquitous Learning Environment”, Proceeding of the World Conference on E-Learning in Corporate, Government, Healthcare, and Higher Education, pp. 991 - 996 (2007). Zhao, X.Y., Anma, F., Ninomiya, T. And Okamoto, T.: “Personalized Adaptive Content System for Context-Aware Mobile Learning”, Int. J. of Computer Science and Network Security, Vol.8, No.8, pp.153 - 161 (2008). W3C: “Mobile Web Best Practices 1.0”, http://www.w3.org/TR/mobile-bp/ (2008). Zhao, X.Y. and Okamoto, T.: “A DeviceIndependent System Architecture for Adaptive Mobile Learning”, Proceeding of the 8th IEEE International Conference on Advanced Learning Technologies, pp. 23 - 25 (2008). Wai, Y.M. and Francis, C.M.L.: “On balancing between transcoding overhead and spatial consumption in content adaptation”, Proceeding of the 8th annual international conference on Mobile computing and networking, pp. 239 - 250 (2002). Ariel, P., Shriram, K. and Michael, P.: “Adapting Content for Wireless Web Services”, Internet Computing, Vol.7, No.5, pp. 79 - 85 (2003). Evangelos, T., Elissavet, G. and Anastasios A.E.: “The design and evaluation of a computerized adaptive test on mobile devices”, Computers & Education, Vol.50, No.4, pp. 1319 - 1330 (2008). Wang, T.I., Wang, T.E. and Huang, Y.M.: “Using a style-based ant colony system for adaptive learning”, Expert Systems with Applications, Vol.34, No.4, pp. 2449 - 2464 (2008).

(10) Yao, J.Y. and Wua, C.N.: “An attribute-based ant colony system for adaptive learning object recommendation”, Expert Systems with Applications (online), Available at www.sciencedirect.com, (2008). (11) Timo, L. and Tapio, H.: “Adapting Web Content to Mobile User Agents”, Internet Computing, Vol.9, No.2, pp. 46 - 53 (2005). (12) Shaha, N., Desai, A. and Parashar, M.: “Multimedia Content Adaptation for QoS Management over Heterogeneous Networks”, Proceeding of the International Conference on Internet Computing, pp. 642 - 648 (2001). (13) Okamoto, T. and Ninomiya, N.: “Organizational knowledge management system for e-learning practice in universities”, Proceeding of the sixth conference on IASTED International Conference Web-Based Education, pp. 528 - 536 (2007). (14) Britton, K. H., Case, R., Citron, A., Floyd, R., Li, Y., Seekamp, C., Topol, B. and Tracey, K.: “Transcoding: extending e-business to new environments”, IBM Systems Journal, Vol.40, No.1, pp. 153 - 178 (2001). (15) Li, C.H., Feng, G.F., Li, W.Z., Gu, T.C., Lu, S.L. and Chen, X.X.: “A Resource-Adaptive Transcoding Proxy Caching Strategy”, Proceeding of the Frontiers of WWW Research and Development, pp. 556 - 567 (2006). (16) Bryan, P., Inmaculada, A.S. and Brendan, T.: “Designing collaborative, constructionist and contextual applications for handheld devices”, Computers & Education, Vol.46, No.3, pp. 294 - 308 (2006). (17) WURFL: http://wurfl.sourceforge.net/. (18) OMA: “Standard Transcoding Interface Specification (2007)”, http://old.openmobilealliance.org/release _prog ram/sti_v10.html (2007). (19) FFMPEG: http://ffmpeg.mplayerhq.hu/. (20) ImageMagick: http://www.imagemagick.org/.

Xinyou ZHAO received his B.Sc. in Computer Science Education Technology from Xinyang Normal University, China in 2000 and Master Degree in Department of Computer Science from Guilin University of Electronic Technology, China in 2003. He worked as an instructor in Guilin University of Electronic Technology from 2003. From 2005 to 2006, he was a researcher in Waseda University. He is currently a PhD candidate at The University of Electro-Communications, Japan. His research interests include mobile learning, data mining, intelligent tutoring system and multimedia technology. He is a member of IEEE and the Japanese Society for Information and System in Education (JSISE). E-mail: [email protected]. Toshie NINOMIYA graduated from Department of Science and Technology, Keio University in 1993 and then worked in Toyota Motor Corporation as an engineer of industrial technology. She is currently an assistant professor in Artificial Intelligence and Knowledge Computing Lab, The University of Electro-Communications, Japan. Now her research interests include Personalized Adaptability of e-Learning Environment and Intelligent Recommendation System. She is a member of the JSISE, and the IEICE. Mikko VILENIUS received his Master's Degree in computer science from the University of Jyvaskyla, Finland in 2006. During the writing of his Master's thesis and after his graduation he worked at the University of Jyvaskyla as a researcher. In October 2006 he entered the University of Electro-Communications of Japan as a Ph. D. candidate. His current research interests include Grid Technology, Collaborative and Cooperative Learning, Bayesian Methods, and Psychometrics. Fumihiko ANMA received his B.E. from Tokyo Institute of Technology, Japan in 2000 and his Master Degree and PhD Degree from Shizuoka University, Japan in 2002 and 2005 respectively. He is currently an assistant professor with the Graduate school of Information Systems, The University of Electro-Communications, Japan. His research interests include Knowledge Computing, E-Learning, Artificial Intelligence, and Software Agent. He is a member of Japanese Society for Information and Systems in

Education (JSISE), Intelligence.

Japanese

Society

for

Artificial

Toshio OKAMOTO received his B.Sc. and Master of Psychology from Kyoto University of Education, Japan in 1971 and 1973 respectively and his PhD Degree from Tokyo Gakugei University, Japan in 1975. From 1981, he worked in Tokyo Gakugei University as an instructor and became an associate professor in 1983. He is currently a Professor and the Chairman of Center for Developing E-Learning in The University of Electro-Communications. His research interests include Intelligent Knowledge, Educational Strategy, E-Learning, e-pedagogy, and collaborative learning. He is a member of the AAAI, the IASTED, the JSISE, the IEICE, the JSET, the IPSJ and the JAEP.