Data Webhouse to support web intelligence in

e-learning environmentsMarta Zorrilla, Universidad de Cantabria, Spain. Socorro Millan, Universidad del Valle, Cali, Colombia. Ernestina

Menasalvas, Universidad Politecnica de Madrid, Spain.

Abstract- The amazing usage of internet as a communicationchannel has changed the very structure of companies activities,being education a clear example of this situation.

Though data mining has been successfully used to improvethe 1-to-1 relationship in e-commerce sites, when applied to e-learning environments the results are often limited to provide theeducator with access summary information or primitive patterns.In any case, information being obtained is not being used pro-actively. For enviroments to profit from the information obtainedit has to be integrated in a global infrastructure. This paperpresents the design of a data e-learning web-house as supportingstructure for future e-learning personalized systems.

I. INTRODUCTION

The explosive growth of Internet as a communication chan-nel has changed the very structure of companies activities,being e-commerce a clear example of this situation. However,no matter what the activity being develop through the internetbe, the direct contact with the service recipient (client in thecase of e-commerce) has been lost and, therefore, it is more

necessary than ever to analyze the use potential navigatorsmight make of this new way of communication.

Education is an example of an activity that is increasinlybeing offered through the web. Students and teachers are

provided with easy to use tools and are not subject to any

temporal constraint. But once again, e-learning environmentslack a closer student-educator relationship. The lack of thisrelation is evident by the fact that a teacher does not reallycontrol the evolution of his students, and students cannotexpress their problems and deficiencies in a natural way. Thisproblem is found in web-based learning environments suchas Virtual-U [25] and WebCT [10], including course contentdelivery tools, virtual workspaces for sharing resources, whiteboards and other tracking tools except specific tools to trackand analyze student behavior and their evolution.

In the case of e-learning, identity of the navigator is notoften a problem as users are almost always (students andteachers) autentified by the system. However, this is nottranslated into a successful 1-to-i relationship. It is verydifficult and time consuming for educators to thoroughly trackand assess all the activities performed by all learners on allthe e-learning supporting tools. Moreover, it is hard to evaluatethe structure of the course content and its effectiveness on thelearning process.

This work has been partially supported by Spanish Mininstry of Educationunder proect N2004-05873" and "TIC2002-01306"

0-7803-9017-2/05/$20.00 ©2005 IEEE

Although data mining has been successfully applied in e-commerce sites to improve the lost 1-to-I relationship, in e-learning environment the results are not yet mature enough(see Section II for a discussion of the effort made in thisdirection). Web-based learning systems can profit from theintegration of data mining and e-learning technology [8].

Application of data mining techniques to e-learning environ-ments is consequently, a challenging activity as most of thework done has to be reviewed in order to be adapted to thisnew kind of activities. To begin with, as a consequence of verydifferent semantics underlying the behavior of navigators andactivities perfomed, a preprocessing stage has to be adapted.Information related to courses, topics, and activities(exams,exercises, ...) will have to be integrated with navigation infor-mation coming from the web server log prior to the applicationof data mining algorithms in order to obtain patterns.

Note that e-learning can be analyzed from different view-points: students, professors, activities leading to useful patternsin any case.

Nonetheless, just knowing most frequent patterns is notenough: it is of vital importance to integrate this informationinto the system so that this information is used proactivelywhen a user (student, teacher) is connected .

Thus, the e-learning platform acting according to user prefe-rences is the unavoidable commitment that e-learning sponsorsmust face except that when doing so, complete informationof courses, events, and in general any information on theeducational organization behind cannot be neglected.

Architectures have been proposed (see Section II) for colla-borative e-systems. In particular, for CRM, Gartnet Group[7] presented a three component (operational, collaborativeand analytical) architecture that can be easily translated whendealing with e-learning environments. In this architecture, it isassumed that the collaborative module actions are the result ofthe patterns obtained by the analysis component that is in turnconnected to the operational module. For all this interchange ofinformation to properly work, data webhouse containing bothinformation of the web and of the application are needed. Thebetter the information the better the results.

In this paper, we present the design of a webhouse adaptedto support collaborative action in an e-learning environment.We also present a prototype of the proposed e-learning da-tawebhouse in which the sole application of OLAP analysishas shown how heuristics to calculate sessions have to beadapted to the navigators in each case.The remainder of the paper is organized as follows. Section

II presents the related work where prior work on the field722

shows on the one hand the inmaturity of research and on theother the needs for support structures so that integrated colla-borative e-learning platforms can be a reality. Consequently,in Section Im we briefly describe the kind of information thatis required as a minimun to be stored, to present in Section IVthe design of the e-learning webhouse proposed. Preliminaryresults of the implementation and use of this webhouse in aproject being developed at Universidad de Cantabria (Spain)as well as future work in presented in Section V.

II. RELATED WORKThe WWW have been used, over the past few years, for

teaching and learning. In this kind of system both studentsand professors are provided with easy to use tools and are notsubject to any temporal constraint. However, it is necesary tobuild a 1-to-I relationship between student and professor andprovide an adequate environment to facilitate learning.A great challenge in this context is to improve both instruc-

tional productivity and learning quality [22]. Active learningstrategies have been incorporated into teaching: expandinglearning experiences, taking advantage of the power of interac-tion and creating a dialectic between experience and dialogue[22].Many proposals have been presented in order to contribute

to improve this relationship. In Oppermann and Rashev [19]the concepts of adaptability and adaptivity on the learningsystems context are discused. Adaptable and adaptive learningsystems make it possible for the student to change someparameters and to adapt to the user based on his/her needs.To make learning systems eficient to the students, adaptableand adaptive, it is neccesary to adapt the learnig environmentaccording to student's goals and capabilities. Some adaptablefeatures are related to the interface, functionality, messagesand feedback. Adaptivity features are related to, for example,user level and task level.

In [11] three types of intelligent agents are introducedin order to assist teachers and students: the digital teachingassistant (DTA), the digital tutor(DT) and the digital secretary(DS). The DTA assists the teacher in teaching functions andthe DT assists students with specific learning needs. The DTcould learn and become more expert to help students and it hasaccess to students learning profiles. Learning profiles includedata about students learning credentials, learning preferences,learning styles and learnig habits. Data to build profiles mayinclude student grades and performance, students usage logsand accomplishments. The DS assists teachers and studentsin logistical and administrative needs. [11] also describestwo major obstacles to obtain these data: the technology andsoftware necessary to collect and analyze learning data, andthe legal challenges to educational institutions for collectingstudent learning data.

According to Zaiane [30], in spite of the significant researcheffort made in e-commerce systems, little has been done inweb-based learning environments. Educators have very littlesupport for evaluating student activities and for identifying

different student on-line behavior. Data mining techniquescould be applied to web-based distance learning in order totrack student activities in a course web site to extract patternsand behavior profiles that help teachers to improve learningresults.

It is necessary, in order to improve web-based learning en-vironment, to take student behavior profiles into account. Usermodeling [20], [23], [1] includes individual preferences whichdetermine the users behavior and personal information abouthim/her. Web usage mining and personalization approachescan also be used to build user profiles [18], [12], [17], [28],[5], [3], [26], [21], [2]in learning domains.Machine learning techniques have been applied to user mo-

deling problems. In [22], IDEAL, an intelligent agent assistedsystem is presented. Student learning-related profiles providestudents with instructional content and facilitate automaticevaluation. A unique personal agent is assigned, in IDEAL,to each student in order to manage his/her profile, includingknowledge background, learning styles, interests and coursesenrolled in. Bayesian belief networks are used to infer astudent model from the performance data.

In [15] web log data generated by course managementsystem are analized to track student learning. An informationvisualization technique is introduced to help instructors tounderstand student behavior and to become aware of whatis happenning in distance education classes. CourseVis, a vi-sualization tool, generates graphical representations of studentdata.Based on data extracted from log data in an education web-

based system, Minaei-Bidgoli and Punch [16], use geneticalgorithms in order to classify students. Using data related toeducational resources (e.g. web pages, demonstrations, simu-lations, homework assignments, quizzes)and user information,data mining methods are analized for extracting knowledge toidentify types of students and problems assigned to them.An evolving e-learning system is described in [27] which

can adapt itself to its users and to the open web based on theusage of its learnig materials. The system users are clusteredbased on their learning interests.

Lei et al. [14] apply web usage mining techniques togetherwith an analytic model in order to both study the learnerinteraction with content and improve instructional designs.To evaluate learning behavior, Wu and Leung [29] presentpreliminary results on learner behavior in which data miningtechniques have been applied.

III. E-LEARNING INTELLIGENCE SUPPORT SYSTEM

Business Intelligence is a popularized, umbrella term intro-duced in 1989 (Gartner Group) to describe a set of conceptsand methods to improve business decision making by usingfact-based support systems [6]. Consequently, it describes allthe set of applications, and technologies for gathering, storing,analyzing, and providing access to data to help enterpriseusers make better business decisions [4]. When translatingthis into education developed through internet, we end up withapplications as the ones described in II to help understand andmake decision, related to students in an e-learning platform.

Most of the work described has to do with applications,algorithms to obtain patterns or user profiles. But profiles aloneare not of interest in an e-learning platform where actionshave to be taken by the platform to stimulate the studentbehavior as if the teacher were observing. The aim of applyingdata mining to data captured by an e-learning tool is to builde-learning intelligent systems. Intelligence understood as theability to understand and profit from experience means thatwhen translated into an automated system, information has tobe stored. This is to say that profiling information and pastbehavior of users has to be stored so to profit in each futureaction of a connected student.

Consequently our main purpose is to discover the informa-tion to be stored for tools to intelligent behave. The morebusiness related information we gather the more profit we canmake out of it, and ,consequelty, the most intelligent decisionswill be made.We distinguish three different nature kinds of information

that have to be stored:. Information gathered by the operational systems of in-

terest for users profiling: information related to students,faculty, enrollment, exams and qualifications on the onehand and information related to courses structure and re-quierements on the other. This is the information availableindependently of the communication channel chosen tolearn (face-to-face learning or e-learning)

. Navigational information: all the information related tousers navigations gathered by the web server: (sessions,time of stay in pages,...).

. Web site structure and content information: informationabout pages, lessons of the courses referenced by pages,

. Analysis information: Information, knowledge obtainedby the data mining algorithms and/or other analyticaltools (profiles, scores, ...)

Note that all but the analytical information prior to be storedwill have to be preprocessed, transformed, and integratedaccording to the data semantics.

In [9], Hu et al. propose a framework for web usage miningand business intelligence reporting which consists of fourphases: data capture, webhouse construction, pattern discoveryand pattern evaluation. Related to the second phase, theypresent a suitable webhouse star schema in which clickstreamand business data can be stored.

In what follows we present the star schema for the caseof e-learning. It is important to say that in our case, thepresented schema supports not only pattern discovery butalso the preprocessing stage as the data web-house proposedwill store information that properly analyzed can be used forpreprocessing. On the other hand as it was suggested above,the webhouse can be used for discovery purposes and foronline application of this knowledge.

IV. LEARNING DATA WEBHOUSE

We present a twofold webhouse structure: on the one hand,it has to store information so that intelligent algorithms can beapplied and patterns are obtained; on the other hand, such an

structure has to be the repository of patterns applied to dataso it stores information about users (teachers and students)behavior and navigation patterns.

In addition to the facts analyzed by Kimball [13] and Huand Cercone [9] we add all the information related to coursesand translate the information of users (customers in businessenviroments and students in this case) in order to adapt themodel to the e-learning enviroment see Figure 1 shown. In fact,as we are mainly interesed in two kinds of users, students andteachers, we propose to analyze them separately. According to[11] we should also store information related to administrativepersonel. However, in the present design we just concentrateon Faculty and students. On the other hand, we will add all thetables needed to gather all the information related to courses,lessons, exams, exercices and any relevant information. Thefollowing data marts are obtained:

. Navigational data mart (see figure 2). This will make itpossible to store all the information that will be needed byalgorithms when calculating navigational patterns (bothrelated to pages or users or both)

. Learning data mart (see figure 3). In this data martstudents are not seen as dimensions but as fact tablesso all the history as student and as navigator is stored.The information in this datamart can be used to calculateincremental patterns, but most importantly it can alsobe used to analyze when the user changes behavior orexhibits certain modifications in his normal patterns. Asthis data mart stores information about the user profile,this is the supporting structure for proactiveness.

* Teaching datamart (see figure 4). Similarly to the learningdata mart, the teaching data mart stores information aboutteachers and their behavior

. Course datamart (see figure 5). It is the data mart inwhich courses as seen as fact tables to be able to storeall the information about courses. It will be the one tosupport of all the operations of analysis of courses forappropiateness, contents, and structure

We present a further analysis of the dimensions and factsin what follows.

A. Dimension tablesMost of the dimensions of the design are common to those

of [13]. Thus only differences among them are highlightedhere.

. Session dimension: In this context, a session must beunderstood as the set of the page references for a givenuser and course, because a student can be enrolled inmore than one course in the same period of time. Becausethe student must generally identify himself to the systemthe start of the session is always known but, as theHTTP protocol is stateless, it is impossible to know whenhe abandons the site. Consequently some assumptionshave been made in order to identify sessions [24], [31].However, the data webhouse itself in this case will makeit possible in the future, according to user behavior indifferent courses, to establish parameters to calculatesessions and other preprocessing parameters.

. User dimension: This dimension gathers only the iden-tification of the users in the plataform and is usefulfor analyzing the clickstream. It serves to relate theinformation about navigation to the demographic data ofeach student and teacher, which is generally stored inother data sources.

. Course and resource dimensions: In order to analyzethe student behavior in the web and compare it tothe behavior of other students in other courses is veryimportant to gather not only the course in which thestudent is connected but also the visited pages, exams andquizzes done and the used collaborative tools. This is thereason our webhouse uses two dimension tables: coursesand resources. The first one collects all the characteristicsof the course (name, URL,...) and the second gathers theresources (content pages, download files, quizzes, exams,news, etc). On the other hand, if a personalization of theenviroment is chased, it can be convenient to group thecourse content and assessment resources in modules. Inthis way, for example, an analysis of the more frequentlyfollowed paths at this level would allow the system torecommend the student the activities and/or resources thatwould favor and improve learning and to help teachers toevaluate their courses structure effectiveness.

* Temporal dimensions: Temporal dimensions are alwayspresented in a data warehouse because each fact tableis a time series of observations of some sort. In thisparticular case, where several hits can be done at eachminute, it seems suitable to create two different temporaldimensions, one for date and another for time. The lattercan be grouped in periods such as early morning, after-noon, lunch hour, . . . which becomes useful for generatingderived variables.

* Event dimension: This dimension collects the specialevents that happened during the course. Because humansbehave different when something special happens, thisinformation is relevant to analyze student and teacherbehavior.

. Referrer dimension: This information is useful to knowthe path followed in the users navegation and analyze ifthe hyperlinks defined in pages and the structure of thecourse are suitable or not.

B. Navigation fact tableThe goal of this fact table is gathering the meaning variables

which are going to be used as input in data mining algorithmsas the corresponding cluster obtained from them (SpecificNavigation Cluster). The star schema for this data mart can beobserved in figure 2. We have established two categories ofvariables: navigation variables such as number of connectionsin certain period or at a certain time of the day, and otherrelations to the use of resources like how often the studentuses forums, or takes quizzes.

C. Learning fact tableIn this table we are interested in collecting data which

allow us to build clusters according to the student in his

behaviour analyzing lessons. We classify the variables asLearning variables and Navigation characteristics variables.Variables from the first group can be obtained from surveyswhich gather the students opinion about the course (struc-ture, aestetic, navigation,...) and from clickstream fact tableextracting information about activities done in the course,results and happenings, if any. Variables from the secondgroup, Navigation characteristics, must be derivated from theclickstream fact table too. Some examples can be often-connected student indicator, morning-studying student and soon. Additionally, this fact table presents another attribute, theSpecific Learning Cluster, which gathers the cluster whichcorrespond to the row, once clusters have been calculated.

D. Course fact tableIn this data mart (see figure 5) we are interested in gathering

data which allow us to analyze courses in the sense ofdiscovering their efficiency in relation to their contents andstructure. Thus, variables about student performance in thecourse and aesthetic and structure characteristics of itselt mustbe considered. As in previous fact tables, the Specific CourseCluster attribute indicates the cluster it belongs to.

V. CONCLUSION AND OUTLOOK

Adding Intelligence to e-learning supporting tools means gi-ving tools the ability to understand and profit from experience.Consequently, in this paper we have presented the design ofa learning data webhouse that will not only gather and storeinformation coming from operation and navigational systemsfor future analysis but the results of the analysis themselves.A prototype of the presented design has been already

implemented and tested at Universidad de Cantabria (see [31])and results are promising. In fact results have so far made itpossible to customize the preprocessing stage. The parameterused for session establishment, can be now selected dependingon courses and kinds of users.As a consequence, the datawarehouse presented here is an

added value to the knowledge discovery process for e-learning.It is also remarkable that the design can be used as a

supporting structure for any data mining architecture. Thekind of attributes that will have to be stored in the apropriatetable depends on the pattern discovery technique used. Anotherthing that has to be customized depending on each particulare-learning enviroment is the structure and the informationavailable on courses and events related to them.Our next step is to adapt the general design presented here

to establish well-known discovery processes in e-learning. Infact, we are working on adapting this design as an automatedincremental repository for typologies of students dependantand non depending on the course they are enrolled in.

REFERENCES[1] Peter Brusilovsky. Knowledge tree: A distributed architecture for

adaptive e-leaming. In Proccedings of WWWO4, May 17-22, New York,pages 104-113, 2004.

[2] Ricardo Carreira, Jaime Crato, Daniel Goncalves, and Joaquim A Jorge.Evaluating adaptive user profiles for nwes classification. In Proccedingsof IUl04, January 13-16, Madeira, Portugal, pages 206-212, 2004.

[3] Yoon Ho Cho, Jae Kyeong Kim, and Soung Hie Kim. A personalizedrecommender system based on web usage mining and decision treeinduction. Expert Systems with Applications, 23:329-342, 2002.

[4] Informatica Corp. Technical glossary. http://www.informatica.com/.[5] Magdalini Eirinaki, Charalampos Lampos, Stratos Paulakis, and -

Michalis Vazirgiannis. Web personalization integrating content seman-tics and navigational patterns. In Proccedings of WIDM04, November12-13, Washington, pages 72-79, 2004.

[6] eLiteral. Decision support systems glossary. http://services.eliteral.com/.[7] Inc. Elizabeth Shahnam. META Group. The customer relationship

management ecosystem.[8] M. Hanna. Data mining in the e-learning domain. In Campus Wide

Information System, pages vol. 21 no. 1 pp. 29-34(6), 2004.[9] Xiaohua Hu and Nick Cercone. An olam framework for web usage

mining and business intelligence reporting. In Proceedings of the 2002IEEE International Conference on Fuzzy Systems, volume 2 (7), pages950-955, 2002.

[10] WebCT Inc. Webct. http://www.webct.com/.[11] Ali Jafari. Conceptualizing intelligent agents for teaching and learning.

EDUCAUSE QUARTERLY, 3:28-34, 2002.[12] Soren Jespersen, Torben Bach Pedersen, and Jesper Thorhauge. Evalua-

ting the markov assupmtion for web usage mining. In Proccedings ofWIDM03, November 7-8, New Orleans, pages 82-89, 2003.

[13] Ralph Kimball and Richard Merz. The Data Webhouse Toolkit: Buildingthe Web-Enabled Data Warehouse. John Wiley and Sons, 2002.

[14] Xu Lei, Claus Pahl, and Dave Donnellan. An evaluation techniques forcontent interaction in web-based teaching and learning environments.In Proccedings of The 3rd IEEE International Conference on AdvancedLearning Technologies, 2003.

[15] Riccardo Mazza and Vania Dimitrova. Visualising student tracking datato support instructors in web-based distance education. In Proceedingsof WWW 2004, May 17-22, 2004, Nwe York, pages 154-160, 2004.

[16] Behrouz Minae-Bigdoli and William F. Punch HI. Using geneticalgorithm for data mining optimzation in an educational web-basedsystem. In Proccedings of WWW 2004, May 17-22, New York, pages675-684, 2004.

[17] Bamshad Mobasher. Webpersonalizer: A server-side recommender sys-tem based on web usage mining. In the Proceedings of the 9th Workshopon Information Technologies and Systems (WITS'99), Charlotte, NC,Dec. 1999, 1999.

[18] Bamshad Mobasher, Honghua Dai, Tao Luo, and Miki Nakagawa.Discovery and evaluation of aggregate usage profiles for web personali-zation. In In Data Mining and Knowledge Discovery, Kluwer PublishingVol. 6, No. 1, pages 61-82, 2002.

[19] Reinhard Oppernann and Rossen Rashev. Adaptability and adaptivityin learning systems. In: A. Behrooz (ed.): Knowledge Transfer (Vol. II).Proceedings on Knowledge Transfer, July 14 - 16, 1997 in London, UK,3:173 - 179, 1997.

[20] Christos Papatheodorou. Machine learning in user modeling. LectureNotes in Artificial Intelligence (LNAI,) Springer-Verlag, 2001, 2049:286-294, 2001.

[21] Michael Pazzani and Daniel Billsus. Learning and revising user profiles:The identification of interesting web sites. Machine Learning, 27:313-331, 1997.

[22] Yi shang, Hongchi Shi, and Su-Shing Chen. An intelligent distributedenvironment for active learnig. In Proceedings of WWWO1, May1-5, 2001, Hong Kong. [http://www10.org/cdrom/papers/207/WWW10-207.htmlJ, 2001.

[23] Goran Simic, Dragan Gasevic, and Vladan Devedzic. Semantic web andintelligent learning management systems. In Proceedings of ITS 2004:International Conference on Intelligent Tutoring Systems. 30 August 3September 2004, Macei-Alagoas, Brazil, 2004.

[24] M. Spiliopoulou, B. Mobasher, B. Berendt, and M. Nakagawa. Aframework for the evaluation of session reconstruction heuristics in web-usage analysis. In INFORMS Journal on Computing, volume 15 (7),pages 171-190, 2003.

[25] Denise Stockley, Chris Groeneboer, Tom Calvert, and Linda Harasim.Virtual-u: An online learning environment. In WebNet, 1997.

[26] Kazunari Sugiyama, Kenji Hatano, and Masatoshi Yoshikawa. Adaptiveweb search based on user profile constructed without any effort fromusers. In Proccedings of WWW 2004, May 17-22, New York, pages675-684, 2004.

[27] Tiffany Ya Tang and Gordon McCalla. Smart recommendation for anevolving e-learniing system. In 11th International Conference on Artifi-cial Intelligence in Education (AIED'2003). Workshop on Technologiesfor Electronic Documents for Supporting Learning. July 20-24, Sydney,Australia. 2003, pages 699-710, 2003.

[28] Ahmad M. Ahmad Wasfi. Collecting user access patterns for buildinguser profiles and collaborative filtering. In IUI 99, Redondo Beach,pages 57-64, 1999.

[29] Albert K W Wu and C H Leung. Evaluating learning behavior ofweb-based training (wbt) using web log. In Proccedings of the IEEEInternational Conference on Computers in Education, 2002.

[30] Osmar Zaiane and Jun Luo. Towards evaluating learners' behaviour in aweb-based distance learning environment. In Proc. ofIEEE InternationalConference on Advanced Learning Technologies (ICALTOI) Madison,WI, August 6-8, 2001., pages 357-360, 2001.

[31] M. E. Zorrilla, E. Menasalvas, D. Main, E. Mora, and J. Segovia.Web usage mining project for improving web-based learning sites. InProcedings of the Ninth International Conference on Computer AidedSystems Theory,Las Palmas de Gran Canaria, Spain, 2005.

726

Ai

Ir

Fig. 1. Clickstmam fact table

Fig. 4. Teaching fact table

Fig. 2. Navigation fact table

Fig. 5. Course fact table

Fig. 3. Learning fact table

727

\ lie 21a

10Imag 1--*-

Si/