Knowledge Discovery from E-Learning Activities

Transcription

1 Knowledge Discover fro E-Learning Activities Addisson Salazar, Luis Vergara Universidad Politécnica de Valencia, Instituto de Telecounicaciones Aplicaciones Multiedia Caino de Vera s/n, 46022, Valencia, Spain ABSTRACT This chapter presents a stud applied to the analsis of the utilization of learning web-based resources in a virtual capus. A huge aount of historical web log data fro e-learning activities, such as eail exchange, content consulting, foru participation and chats is processed using a knowledge discover approach. Data ining techniques as clustering, decision rules, independent coponent analsis, and neural networks, are used to search for structures or patterns in the data. The results show the detection of learning stles of the students based on a known educational fraework, and useful knowledge of global and specific content on acadeic perforance success and failure. Fro the discovered knowledge, a set of preliinar acadeic anageent strategies to iprove the e-learning sste is outlined. Kewords Knowledge Discover, Data Mining, Pattern Recognition, Ontologies, Decision Trees, Clustering, Principal Coponent Analsis, Independent Coponent Analsis, Neural Networks, eb-based Applications, Data Preprocessing, Data arehouse, Electronic Learning (E-Learning), Case Stud, Learning Stles, E-learning Resources, Virtual Capus, eb-based Courses, eb-based Education Adinistrative Issues, eb-based Interactions, eb-based Learning, eb-based Teaching, E-ail, Groupware, Collaborative Technologies, Docuent sharing, Electronic eetings. INTRODUCTION This chapter contains a case stud on knowledge discover research carried out on data of graduate and undergraduate courses at the Universidad Politécnica Abierta (UPA) site. This universit is a virtual capus at Universidad Politécnica de Valencia and currentl it has ore than 6000 students registered in about 230 courses, Figure 1 shows a general schea of the virtual capus learning environent at UPA. The stud pursued to obtain knowledge about acadeic perforance success and failure of the students and analzing the e-learning event activit at the capus web to recognize patterns on learning stles of the students. Events covered the personal and collaborative use of the web resources in course activities, including content consulting, eail exchange, foru participation, and so on. The underling hpothesis was there is useful hidden knowledge in data fro e-learning web activities for acadeic anageent and evaluation of the e-learning sste. The chapter describes an integrated ethodolog to extract knowledge fro quantitative and qualitative data; the results obtained its evaluation and a strategic action outline derived fro the discovered knowledge. Different data ining techniques were used to exploit the e- learning data through a knowledge discover approach (Faad, Piatetsk-Shapiro, Sth, & Uthurusa, 1996), (Cabena, Hadjnian, Stadler, Verhees, & Zanasi, 1997), (Maion & Rokach, 2005). Those techniques included, Independent Coponent Analsis (ICA), Neural Networks (NN), clustering, linear regression, and decision trees. ICA allowed distinguishing the independence of the events and detecting learning stles; NN was used to obtain patterns of the student behaviour, linear regression was eploed for nueric analses of the

2 relationship between the student perforance and the event activit levels. Quantitative clustering and qualitative conceptual clustering algoriths were applied for grouping data in hoogeneous datasets. To enable qualitative analsis of the data, nueric data were discretized and descriptions for their interpretation were obtained. Finall, on the descriptive datasets, a ining association rule process was ade b appling the C4.5 decision tree algorith. The obtained decision rules involved global and specific content knowledge that was evaluated b acadeic experts taking into account their validit, novelt, and siplicit. The results were considered as useful for e-learning acadeic anageent. Data fro the use of the UPA web facilities included the following web-log statistics about e- learning event activities: course access, agenda using, news reading, content consulting, eail exchange, chats, workgroup docuent, exercise practice, course achieveent, and foru participation. Date and tie for each event also were available. Besides of the inforation on the web activit, the exercises achieved and grades obtained b the UPA s students were tried in the knowledge discover process. The data were collected fro the virtual capus web in the period fro Januar 2002 to March 2005, totalizing 2 391,003 records. The process of knowledge discover covered the following stages: o Building a reliable datawarehouse, b filtering data inconsistencies, solving data heterogeneit probles and processing data. o Obtaining and interpreting patterns of the student behaviour in e-learning activities b using independent coponent analsis, neural networks and linear regression analsis. o Obtaining hoogeneous data groups b appling clustering processing and selecting data groups, sorted out b research topics, for the definition of decision rules. o Appling a knowledge representation on selected groups using decision trees to obtain the decision rules of the factors that influence on acadeic achieveent success and failure. o Evaluating knowledge findings b experts, fro the point of view of their validit, novelt, and siplicit. o Outlining strategies for the iproveent of acadeic processes. Figure 1. Virtual capus learning environent at UPA The following sections describe the background and context of this work, the results obtained in each stage of the knowledge discover process showing partial findings fro the data ining techniques. Final sections include the global and particular conclusions about acadeic perforance and learning stles and future work. BACKGROUND

3 Knowledge discover in databases (KDD), or just knowledge discover, is a subdiscipline of coputer science, which ais at finding interesting regularities, patterns and concepts in data. Usuall knowledge discover has been related with the global process spans fro data to knowledge using different statistical and heuristic techniques called data ining techniques. However in the current literature knowledge discover, data ining and achine learning are often used interchangeabl. Recentl, the data ining approach has been applied in acadeic research. Those applications include predictive or descriptive odelling on educational data. Traditional sources of data have been databases or questionnaires, and ore recentl data fro the web. Soe of the works in educational predictive odels fro databases or questionnaires are the following: predicting whether the students graduates in six ears (Barker, Trafalis, & Rhoads, 2004), selecting students who would need reedial classes (Ma, Liu, ong, Yu, & Lee, 2000), and predicting individual student s final acadeic achieveent b odelling with decision trees and hierarchical odels (Gasar S., Bohanec M., & Rajkovic V., 2002). Other predicting works on student s success, errors or help request are: predicting the tie spent in solving an exercise task b using neural networks (Beck & oolf, 1998), and predicting in which word the student asks help in reading English, where inforation of the student (gender, approxiated reading test results of the da, help request behaviour) and word (length, frequenc, etc.) were processed (Beck, Jia, Sison, & Mostow, 2003). Regarding to descriptive data odelling techniques applied to educational data, there are several references in subjects such as: analzing factors with affect acadeic success, desertion and retention of students, ining navigation patterns in log data, analzing student s copetence in course topics, analzing student s errors in progra codes, and ining student answers fro a web-based tutoring tool database to get pedagogicall relevant inforation and to provide feedback to the teacher (Shin & Ki, 1999), (Salazar A, Gosalbez J, Bosch I, Miralles R, & Vergara L, 2004), (Kristofic & Bielikova, 2005), (Roero, Ventura, De Bra, & Castro, 2003), (Merceron A. & Yacef K., 2003). Data ining fro web data (webining) is a new research area that pursues to understand the inforation flow at the web b eans of autoated techniques for searching knowledge. This area has a wide range of eergent applications including e-learning, e-coerce, autoated inforation assistants and an applications that operate through the web (Srivastava, Coole, Deshpande, & Tan, 2000). One exaple of webining is the classification of web pages based on understanding the textual content of eails based on hierarchical probabilistic clustering (Larsen J., Hansen L.K., Szkowiak A., Christiansen T, & Kolenda T., 2002). Nowadas there exist an topics open in e-learning concerning to the qualit evaluation of the sste, the knowledge and control of the web activities of the students, and the use of the huge quantit of outcoe inforation fro the e-learning process (Sun, Tsai, Finger, Chen, & Yeh, 2007), (Liu & Yang, 2005), (Pituchs & Lee, 2006), (Reill, 2005), (Seli, 2007), (Pirauthu, 2005), (Kiber, Pilla, & Richards, 2007), (Shee & ang, 2007), (Liaw, Chen, & Huang, 2006). Particularl there is an increasing interest in webining of the e-learning data, soe exaples are: predicting drop-out on deographic data (sex, age, arital status, etc.) and course data in the first half scores of the course (Kotsiantis, Pierrakeas, & Pintelas, 2003), predicting the course score processing success rate, success at first tr, nuber of attepts, tie spent on the proble, etc. (Minaei, Kash, Korteeer, & Punch, 2003), cobining several weak classifiers b boosting to predict final score (Zang & Lin, 2003). Recentl, new holistic webining approaches considering extracting learning stles fro the web navigational behaviour of the students have been outlined (Mor & Minguillón, 2004), (Xenos, 2004), (Garcia, Aandi, Schiaffino, & Capo, 2007).

4 A learning-stle odel classifies students according to where the fit in a nuber of scales corresponding to the was in which the receive and process inforation. One of the ost accepted learning stle taxono for engineering students is (Felder & Silveran, 1988), see Table 1 -one learning stle is confored b the cobination of one feature in each diension, for instance, intuitive-visual-deductive-active-global-. This odel was used in the present research Preferred Learning Stle Sensor Intuitive Perception Visual Auditor Input Inductive Deductive Organization Active Reflective Processing Sequential Global Understanding Corresponding Teaching Stle Concrete Abstract Content Visual Verbal Presentation Inductive Deductive Organization Active Student Passive Participation Sequential Global Perspective Table 1. Diensions of Learning and Teaching Stles (Felder et al., 1988) pp. 675 Our work pursued two objectives: 1) to find patterns on acadeic perforance of the students, and 2) to detect student learning stles underling in the web-data. Thus the fraework is webining for descriptive odelling of the educational data. e contribute with an epirical stud with a huge aount of data to contrast the results. The coplexit of the social studied phenoenon requires of such kind of analsis as we can found in recent literature (Lev, 2007), (Schellens & Valcke, 2006), (Puntabekar, 2006), (Stephenson, Brown, & Griffin, 2006). In contrast of using one technique as Baesian networks, we used an integrated approach with several techniques in order to exploit and copleent the advantages of each one of techniques. This work is developed in a novel research line that pursues to discover student learning stles in order to feedback and iprove the e-learning sste using no-odel based ethods. The understanding of learning stles is ore difficult in e-learning education than traditional education environent, and then research in this topic is alwas a challenge. Mining patterns fro data is a classical task eploed for a long tie and it is especiall useful and suitable in the e-learning context due to the new generating data processes coe fro web technological innovations. The algoriths applied search for patterns in the data and its output could be stored in a knowledge based sste. However the scope of the chapter does not coprise the stage of knowledge storage, so etadata are not included for the output of the algoriths; although fro the decision rules obtained and using, for instance, logic prograing languages, the creation of a knowledge base can be easil undertaken. In addition, the set of decision rules, cluster descriptions and learning stles detected in this work iplicitl define a kind of inforatics ontolog that could be used to share or reuse the discovered knowledge, using an ipleentation forat and language. There are several references coparing inforatics ontologies and statistical approaches as the included in this chapter (Caragea, Pathak, & Honavar, 2004), (Pils, Roussaki, & Stripakou, 2006), (Goez-Perez & Manzano-Macho, 2004). Soe of the definitions of inforatics ontolog are the following: the inforatics ontologies define the kind of things that exists in the application doain, allowing no confusion of the ters and sbols (Sowa, 2000), inforatics ontologies define a kind of explicit specification of a concept set (Gruber, 1993), inforatics ontologies are defined as foral specification of a shared concept set (Borst, 1997). The principal otivation of the inforatics ontologies is allowing sharing and reusing of knowledge bases coputationall, b using a coon vocabular.

5 Inforatics ontolog can be defined in several was, but necessaril it includes a vocabular of ters and soe specification of their eanings. This includes definitions and specifications on relationship between ters, so in general a structure is iposed on the doain and ter interpretation is constrained (eigand, 1997), (Uschold, King, Morales, & Zorgios, 1998), (Fridan & McGuinness D., 2001). Thus, inforatics ontolog can be a logic theor, a seantic foral description, the vocabular of a logic theor and a specification on conceptions. DATA PREPROCESSING A reliable datawarehouse was created fro the historical ( ) log-web data fro the UPA. For this end, different operations were applied to the original data: Filtering of issing and erroneous data, solving data heterogeneit probles due to different data sources, adaptation of variables for data processing and data retrieval for stratified analsis. Figure 2 shows a siplified schee of the data entities at the UPA. Figure 2. Structure of the web data The objective data were collected fro the web activit of the virtual capus in the tables: Grades and Event. The event table contained one record for each event at the virtual capus web, it fields were: group, course, student code, event tie and event class, see kind of events (e-learning activities) at tables 1 to 10. The total nuber of events in the analzed period was of which corresponded to student events and the rest of events were for teachers and sste adinistrator. Fro this table, onl the records corresponding to students belonging to courses with ore than 3 students were selected ( records) as event table for the stud. A new event table was calculated suarizing records for student and kind of event. Resulting table contained the fields: group, course, student course, event class counter (event instance total for that event class) and average event class tie (average tie of the student activit in that event class). It was a records table. In order to build a datawarehouse for data ining, a new table with the projection of the event table on the student code was calculated. It was an 8909 records table. That nuber of records is due to that active and no active student data were contained in the initial event table. For each student, the corresponding total instance counter of each kind of event was calculated, and a noralized value (1-100 scale) of student event activit was calculated with the following equation, event total instancestudent even_activ itstudent 100 (1) instance axiu event

6 The student activit data were added as fields to the datawarehouse. Fro the grades table, the average grade for a student was calculated and added to the datawarehouse. Because all the virtual courses do not have evaluation, onl 1873 of the rows of the datawarehouse had a value for the variable average grade. To get the qualitative descriptions of the student event activit, tables 2 to 11 were defined. Tpe Course access 1 Alost never 2 Occasionall 3 Usuall 4 Ver frequentl Table 2. Course access description Tpe Agenda using 1 Does not use it or use it a little 2 Average use 3 Use it a lot Table 3. Agenda using description Tpe News reading 1 Does not read it or alost never read it 2 Average use it Table 4. News reading description Tpe Content consulting 1 Alost never 2 Occasionall 3 Usuall 4 Ver frequentl Table 5. Content consulting description Tpe Eail exchange 1 Sporadicall exchange it 2 Usuall exchange it 3 Copiousl exchange it Table 6. Eail exchange description Tpe Chats 1 Not ver active 2 Fairl active 3 Ver active Table 7. Chat description Tpe orkgroup docuents 1 Low collaboration 2 Average collaboration 3 High collaboration Table 8. orkgroup docuents Tpe Exercise practice 1 Few exercising 2 Enough exercising 3 Much exercising Table 9. Exercise practice description Tpe Course achieveent 1 Sporadicall 2 Usuall 3 Ver frequentl Table 10. Course achieveent description Tpe Foru participation 1 Little 2 Average 3 High Table 11. Foru participation description The global ean and liits of the event activit were calculated using the following equations: total instance nuberevent eanevent (2) student nuber liit ean ( axiu - iniu) 0.3 (3) event event Equation 3 was applied after checking the noralit of the event instance distributions. The superior liit (supli) for event instance was calculated using plus in equation 3 and inferior liit (infli) was calculated using inus in equation 3. Thus the 60% of the probabilit densit distribution of the event instance was contained between the superior and inferior liits. Given the event total instance nubers (event activit) for a student, the corresponding description values were calculated using the description tables and the event liits. To event

7 allocate a value of the description table fro an event activit value for a student, the following algorith for 4-entries description tables was used: If event_activit < infli event allocate to Tpe 1 If event_activit >= infli event & < ean event allocate to Tpe 2 If event_activit >= ean event & < supli event allocate to Tpe 3 If event_activit >= supli event allocate to Tpe 4 In the case of 2-entries tables, Tpe 1 was allocated when event_activit < ean event and Tpe 2 was allocated when event_activit >= ean event. For tables with 3 entries, Tpe 1 was allocated when event_activit < infli event, Tpe 2 was allocated when event_activit >= infli event & < supli event, and Tpe 3 was allocated when event_activit >= supli event. Besides of assigning description values to event activit values, the instance tie of the events was discretized using the following hour intervals: [0-6) =dawn, [6-12) =orning, [12-18) =evening, and [18-24) =night. The qualitative value for average grade was calculated using the following discretization intervals: [0-5) =unsatisfactor, [5-7] =fair, [7-9) =good and [9-10] =excellent. At the end of the data preprocessing stage, the datawarehouse obtained consisted of a 8909 (student records) x 27 (variables) as follows: 3 identification record variables (group, course, student code), 4 variables for quantitative and qualitative values of average grade and average tie and 20 variables for quantitative and qualitative values of the activit for the different kind of e-learning events. DATA MINING SCHEME Figure 3 shows a general schea of the relationship between the data ining techniques applied on the datawarehouse. Quantitative clustering was ade b appling the fuzz c-eans algorith (Bezdek J.C & Pal S.K, 1992) and qualitative clustering using the conjunctive conceptual algorith (Michalsk R.S & Stepp R.E, 1983). Figure 3. Interconnection of the applied data ining techniques

8 Total population of the datawarehouse was divided accordingl to the course tpes in: graduate (inforal courses), doctorate, and regular acadeic career courses. In addition, each of those population divisions was divided in two subsets: cases with grades and cases with no-grades. Therefore 6 disjunctive data subsets to analze were generated. INDEPENDENT COMPONENT ANALYSIS (ICA) ICA is a powerful statistical technique that has had a successful application in different areas of signal processing (Hvärinen, Karhunen, & Oja, 2001), (Cichocki & Aari, 2001). ICA assues that there is a M-diensional zero-ean vector s ( t ) s ( t),, s ( t) T 1, such that the coponents s (t) M i are utuall independent. The vector s (t) corresponds to M independent scalar-valued source signals s i (t). The ultivariate probabilit densit function (p.d.f.) of the vector can be rewritten as the M product of arginal independent distributions p( s) ( ). A data vector x ( t ) x ( t) x T 1 N ( t) is observed at each tie point t, such that x( t) As( t) where A is called ixture atrix and it is full rank N x M, (Hvärinen, Karhunen, & Oja, 2001). There are several standard ICA algoriths as FastICA (Hvärinen & Oja, 1998), Extended Infoax (Lee, Girolai, & Sejnowski, 1999) or TDSEP (Ziehe & Müller, 1998). Those algoriths rel on assuptions about the source signals, such that ipl a given odel for the source distributions or ake assuptions that are onl fitted to specific applications. e applied standard ICA algoriths and a new non-paraetric ICA algorith proposed in Annex 1. This latter algorith ields the best results, because it was ore adaptable to the data. It does not assue an restriction on the data, since the probabilit distributions are calculated directl fro the training set through a non-paraetric approach, and also focusing the independenc between the source coponents directl fro its definition based on the arginal distributions. i1 p i s i Figure 4. Data of the graduate courses with grades for the 10 events, and average grades and connection tie: e1 (course access) e10 (foru participation)

9 Figure 5. Sources calculated b an ICA algorith for data of Figure 4 ICA was applied on the UPA data in order to identif independent sources (independent event activit), i.e. searching those event activit that can separate b an ICA algorith as a source. Figure 4 shows the estiated activit for the 10 events (see Tables 2-11) on the UPA web plus the average connection tie and average grade for 1072 students of graduate courses with grades. Note that data are displaed as signals (vectors of saples) for Figures 4 and 5. This latter show the sources estiated b an ICA algorith, note that signal of event 8 (exercise practice) in Figure 4 is ver siilar (high correlated) to source 5 in Figure 5, it eans that the activit corresponds to the workgroup docuent event could be recognized as an independent source for this subset of data. After analzing the results fro ICA applied to the different data subsets and considering additional inforation about the courses and students in the capus, we can infer the following conclusions: o Eail exchange was independent in soe cases. It could be due to weakness in teaching strategies for prooting the student interactivit. Then eail exchange is transfored in eail review done as a routine. o In courses with no grades, the workgroup docuent event was independent. The lack of evaluation and grades discourage the participation of students in collaborative tasks. o In soe datasets the content consulting event was independent as reflect of a kind of distributed passive learning (DPL) nature of the web platfor. Thus content consulting becoes a routine consisting in download aterials with no interactive learning process. o Exercise practice and course achieveent also were found as independent events for soe datasets. It could be due to the profile of soe students that includes inforation and telecounications background and knowledge about course contents. For those students participating in those event activities could be irrelevant. PRINCIPAL COMPONENT ANALYSIS (PCA) AND ICA PCA is a ver well known technique that reduces the variable diensionalit in statistical ultivariate analsis (Hardle & Siar, 2006). e applied PCA for grouping the events of the web activit in learning diensions taking into account the Felder s fraework (Felder & Silveran, 1988). PCA reduced 10 web event activities to 5

10 coponents. To solve the proble of detecting learning stles in e-learning we assue that the underling independent sources that generate the web log data are diensions of the learning stles of the students and we observe x linear cobinations of those stles through the use of the facilities b the students at the virtual capus. Then, si,( i 1,,5 learning stle diension) correspond to the perception, input, organization, processing, and understanding diensions (see Table 1); and the ixture atrix A provides the relation between e-learning stle diensions and e- learning event activities, aij,( i 1,,5 learning stle diension), ( j 1,,10 e - learning activit). Table 12 contains the six first sorted contributions of web activities of the ICA ixture atrix for the 5 sources estiated. Each source was associated with one learning diension of Table 1 analzing the weight of the web activities and considering the principal evaluation ethodologies eploed b teachers for graduate courses with grades. -Diension 1 was not detected and diension 5 was detected twice-. The ethodologies assigned grades focusing on: achieveent, individual student participation, or group work. The iplicit teaching stles of the evaluation ethodologies encourage specific learning stles of the students, as we explain below. LSD * Sorted web activit contribution chat foru news eail access exercises eail content wg-doc ** exercises foru chat wg-doc news achieve content chat eail achieve content agenda access foru news access agenda content achieve eail chat Table 12. ICA Mixing Matrix ( * Learning stle diension, ** workgroup docuents) The learning diension 1 (sensor-intuitive) corresponding to perception was not detected in the ICA ixing atrix; it could be because the ephasis of educational strategies did not favour to highlight that diension. Fro Table 12, the relationship between learning stle diensions and web activities can be ade, see Table 13 where we have added a possible web activit cobination for learning diension 1. Note that soe web activities are associated with ore than one diension; it has sense because a web activit could deand several capabilities of the students used in their learning process. Allowing that kind of relationship we can obtain ore real and versatile descriptions of the student learning stles, besides of including all the diensions of the learning fraework. In (Garcia, Aandi, Schiaffino, & Capo, 2007) just three diensions of the Felder s odel were considered and the Baesian network proposed constrained relationship of the web activities with just one diension of the learning odel Learning Stle Sensor- Intuitive Perception Visual- Auditor Input Inductive- Deductive Organization Active- Reflective Processing Sequential- Global Understanding eb event activit chats, foru participation, course access. chats, foru participation, news reading, eail exchange. workgroup docuent, news reading, course achieveent, content consulting. eail exchange, content consulting, workgroup docuent, exercise practice. course access, agenda using, content consulting, course achieveent. Table 13. Association between learning stles and eb activities

11 Figure 6 shows the sources 3, 4, and 5 (organization, processing, understanding) obtained for the grade graduate course dataset. Four labelled characterised zones in the learning stle space are displaed: 1.) Represents the learning stle ore iportant in the population. The learning for the students in this zone ephasizes global understanding, active processing, and deductive logic (natural huan teaching stle), and high grades. 2.) This learning stle is focused on inductive logic (natural huan learning stle), with sequential understanding, and relative active processing. Students within this stle could have natural skills for virtual education. 3.) It is characterised b global understanding, deductive logic, and reflective processing. Students within this stle would have higher abstraction skills that need of teaching. 4) Basicall this cluster represents outliers with individual learning stles. Figure 6. Three sources in a learning stle space for graduate courses with grades e can conclude that diension of understanding enables to project clearl the learning stles, and its principal coponents are achieveent, content, and agenda. This finding confirs the assuption that the ore quickl wa to change the learning stle of the student is to change the assessent stle, i.e., expected evaluation bias how the student learns (Elton & Laurillard, 1979). e ade a cluster validation procedure to deterine best qualit of cluster configuration for data of Figure 6. It consisted in estiate the partition and partition entrop coefficients for different nuber of clusters (Haldiki, Batistakis, & Vazirgiannis, 2001). The best cluster configuration for data of Figure 6 was 4 clusters -a detailed explanation of cluster validation procedure is in cluster analsis section-. Figure 7 shows three sources for graduate courses with no grades. The distribution of the data in Figure 7 does not allow foring learning stle groups and show all the subjects within a unique learning stle. As understanding and organization diensions do not discriinate projection of the learning stles, onl the diension of the processing provides soe discriination. Then the unique learning stle ephasise reflection over actuations, it would be the content consulting and exercise practice coponents of that diension. The conclusion is the lack of assessent does not allow developing student learning stles.

12 Figure 7. Three sources in a learning stle space for graduate courses with no grades Results for regular acadeic career courses were siilar to the graduate courses results finding eaningful learning stles for courses with grades. Results of this section could be analzed as a kind of ontolog. The conceptions are the learning stles detected, related with the diensions of learning (input, organization, processing, understanding), and ultiatel with the web learning activities. Figure 8 shows such kind of ontolog (nubers in the boxes corresponding to the nubers of learning stles in Table 1). Figure 8. Ontolog of learning stles detected REGRESSION ANALYSIS AND NEURAL NETORKS A linear regression odel was designed with the average grade as dependent variable and the ten web event activities plus the average connection tie as independent variables. The adjustent of the linear odel were not statistical significant, so there is soe non-linear relations between the variables. e tried with ore adaptive odels appling the Linear Vector Quantization (LVQ) neural networks to classif the different datasets in for classes depending on the average grade: [0-5) =unsatisfactor, [5-7] =fair, [7-9) =good and [9-10] =excellent. LVQ algorith includes the self-organizing and copetitive stages; four output neurons were

13 defined corresponding each to the target classes. 75 % of the data were used in the training phase and the rest of the data in testing phase. Kohonen learning rate of 0.01 and conscience learning rate of were eploed. Different tests were ade varing the nuber of neurons of the hidden laer fro 4 to 10. The groups found autoaticall b LVQ could be interpreted with siilar contents to the obtained b the fuzz clustering algorith described in next section. However soe dissiilar groups were found b LVQ, e.g., students that onl have a lot of chats and obtain an excellent average grade, students that exchange eails and have chats with excellent average grades. CLUSTERING ANALYSIS In the clustering procedure, a (specified) nuber of clusters are calculated fro a set of objects. A cluster is represented b a cluster center which defines the center point of the cluster in the feature space. A cluster center is thus an (iaginar) object which defines the tpical or ideal representative of its cluster. Figure 9 shows a schee of a data cluster structure, consisting of five clusters and the distances between cluster centroids. Doughnut sizes represent the nuber of records in each cluster and doughnut slides represent the 12 variable values of cluster centroids. Figure 9. Cluster structure The fuzz c-eans was applied using a fuzziness degree of 1.3 (exponent ). Validit easure was the partition coefficient and the axial nuber of classes used in training was 16. The calculation was carried out for all classes and the best nuber of classes was deterined and validated b checking the evolution through the class nuber range of the partition coefficient (pc) vs. the classification entrop (pe), see Figure 10.

14 Figure 10. Cluster validation Figure 10 was calculated with graduate with no-grades data subset and the best portioning is at c=5. The partition coefficient and the partition entrop both tend towards onotone behaviour depending on the nuber of clusters. So as to find the "best" nuber of clusters c* one chooses the nuber where the entrop value c* lies below the rising trend and the value for the partition coefficient lies above the falling trend. On viewing the curve of all the connected values, this point can be identified as a kink (thus the nae "elbow criterion"). The best partitioning of the clusters applies at that point with a value of c to get the highest cluster differentiation (axia of inter-clusters ean distances) with good hoogeneit within cluster ebers (inia of distances between cases and centroids) (Haldiki, Batistakis, & Vazirgiannis, 2001). The application of the fuzz c-eans algorith on the defined 6 data subsets generated 23 clusters or groups. 10 groups for the data subsets graduate and career regular courses with grades, and 13 groups for the sae courses with no-grades. For the doctorate courses no groups were generated due to there were not enough cases to analze. Besides of the fuzz c-eans, the conjunctive conceptual algorith was used on the qualitative variables to get logical conjunctions of relations between the variables. Fro the analsis of the obtained cluster centroids, the following conclusions on acadeic perforance were derived. For the regular career courses with grades: o The student group with the best grades shows a siilar activit level in the different event tpes, except in course achieveent where it has a higher activit than the other groups. o The student group with the worst grades shows a higher exercise practice proportion than the other groups; however the course achieveent is relativel low. o The interediate acadeic perforance groups show an ibalance in the proportion of activities, focusing in eail exchange and agenda using. o This data subset does not use events that require interactivit (chats, foru participation and workgroup docuents) aong several students. o The students with worst grades are devoted ainl to news reading, content consulting and eail exchange but the do not undertake to course achieveents. o The relation cluster-grade follows a noral distribution respecting to the average grade decision variable. Being the best and worst grade groups the less nuerous ones. For the graduate courses with grades:

15 o There are not significant differences in the acadeic perforance of the clusters. o The worst grade group is the highest course achieveent one. o The best grade group shows a siilar proportion in ever event activit, being this group the highest course access. o In this data subset the events that require interactivit aong students were used, but its activit was lower than the events that do not require student interactivit. o The best grade group was the second ost nuerous ones and the worst grade group was the less nuerous ones. o The best grade groups used the eail ore frequentl than the others. For the regular career courses with no-grades: o Ever group exhibited a good utilization of the interactivit events foru and chats but workgroup docuent event. o Ever group showed siilar proportion in exercise practice and course achieveent events. For the graduate courses with grades: o Ever group showed a high utilization of interactivit events; even the workgroup docuents activit was high. o There was a group with high values for the three interactivit activities, but the value for exercise practice event is ver low like in the other groups. Clusters calculated fro the graduate course with no-grades data subset are shown in Figure 11. Event activit values in the cluster representation are noralized. Figure 11. Event activit proportion at calculated clusters fro graduate course data subset MINING DECISION RULES Average grade was defined as outcoe variable for ining decision rules on acadeic perforance. 250 decision rules were obtained appling the C4.5 algorith (Quinlan R.J, 1992) to the clusters of the graduate and regular career course data subsets and to the group consisting of the union of both data subsets. Decision rules involve global content or specific content knowledge. Soe of the ined decision rules are listed below including the success percentage of the rules.

16 For graduate courses: Rule No 9: If Agenda using = Average use and News reading = Does not use it or use it a little and Content consulting = Usuall and Foru participation = Average Then Average Grade GOOD [80.65%] Rule No 16: If Content consulting = Ver frequentl and Course achieveent = Usuall and Foru participation = Average Then Average Grade GOOD [77.24%] Rule No 36: If Course = Environent Sste Manageent and Average tie = Evening and Eail exchange = Copiousl exchange it and Chats = Ver Active Then Average Grade EXCELLENT [71.23%] Rule No 55: If Course = Renewable Energ and Chats = Fairl active Then Average Grade EXCELLENT [71.45%] Rule No 67: If Course = Teledetection Sstes for Environent Risk Prevention and Course access = Ver frequentl and Content consulting = Occasionall Then Average Grade FAIR [80%] Rule No 78: If Course = Basic Environent Technical English and Course achieveent = Null Then Average Grade UNSATISFACTORY [80%] For the regular career courses: Rule No 114: If Course = Econoic and Financial Sste and Exercise practice = Much exercising and Course achieveent = Usuall Then Average Grade FAIR [83.54%] Rule No 115: If Course = Econoic and Financial Sste and Average tie = Evening and Eail exchange = Usuall exchange it and Course achieveent = Ver frequentl Then Average Grade FAIR [73%] Rule No 116: If Course = Econoic and Financial Sste and Average tie = Evening and News reading = Average use it and Content consulting = Usuall and Chats = Fairl active Then

17 Average Grade FAIR [81.5%] Rule No 121: If Course = Econoic and Financial Sste and Average tie = Morning and Agenda using = Average use and Content consulting = Usuall Then Average Grade FAIR [79.57%] Rule No 133: If Course = Manageent I and News reading = Average use it and Exercise practice = Enough exercising Then Average Grade GOOD [94.85%] Rule No 140: If Course = Econoic and Financial Sste and Average tie = Evening and Eail exchange = Usuall exchange it and Chats = Ver active Then Exercise practice = Enough exercising Then Average Grade GOOD [95.27%] Rule No 141: If Average tie = Evening and Chats = Null and Exercise practice = Enough exercising and Foru participation = High Then Average Grade GOOD [84.24%] Rule No 154: If Course = Manageent I and Exercise practice = Much exercising and Course achieveent = Usuall and Average Grade EXCELLENT [90%] The knowledge findings obtained in the research were evaluated b acadeic adinistration experts of the universit in aspects such as validit, novelt, and siplicit, obtaining general score of 8.2 points on a scale fro 1 to 10. Fro the point of view of inforatics ontologies, the trees confored b decision rules could be analzed as concepts about good and bad acadeic achieveent of the students. Those rules provide structure for relationship of the ters that define the student behaviour using the web e-learning activities. In (Silvescu, Reinoso-Castillo, & Honavar, 2001) is explained how iplicit ontologies drive the inforation extraction and data integration procedures used in knowledge acquisition fro data, specificall using decision trees. STRATEGIC ACTION OUTLINE The following are soe preliinar strategies for acadeic anageent that were proposed based on the results of the knowledge discover process. o To epower collaborative inforatics to include practical virtual labs. Soe of the technical subjects to be included are: workgroup, workflow, data ining, searchers, ultiedia, and custoer research anageent.

18 o To collaborate with other educational networks or virtual platfors in order to reinforce teaching qualit and prooting the creation of virtual learning networks. o To design and to ipleent pedagogic course sllabus for students to understand e- learning education. The appropriate utilization of inforation and counications technologies helps to educate ore and better. o To adapt the roles of counsellors, teaching, supporting and adinistrative staffs to the classes in the cberspace. o To proote the knowledge of the virtual capus in all the universit to generate snergic relationship between universit people. This can produce positive feedback to the virtual capus as new students and teachers, e-learning project creation, and counications for iproveents. o To propose special events for diffusion of the virtual universit as conferences and workshops on-line, in order to obtain participation of the students. o To ipleent a virtual librar with references to bibliographic contents of virtual courses. Thus in addition to the basic odules and annexes of the courses, access to bibliographical electronic resources, allowing research activities, would be provided. CONCLUSIONS AND FUTURE ORK Conclusions The proposed ethodolog applied to a real case with huge historical data obtained proise results in detecting student learning stles in an e-learning environent. The diensions of the Felder s learning fraework were odelled using and adaptive approach. The versatilit of the approach consists in integrated descriptive odelling using several data ining techniques for processing quantitative and qualitative data. Non-paraetric independent coponent analsis, a technique norall used in signal processing, has been useful for detecting patterns in e-learning data. Despite of the possible probles of discretization, iproveent of interpretation capabilities has been deonstrated. Modelling learning diensions as a cobination of web event activities enhanced the detection of the student learning stles. The knowledge discover fro e-learning web data found useful knowledge (of global or particular content) on acadeic perforance of the students at the Universidad Politécnica Abierta (UPA). Aong the findings are the following: i.) Events of snchronous interactivit, such as, chats and foru participation and events of asnchronous interactivit epower the student acadeic perforance; ii.) In the courses with grades, acadeic student perforance could be iproved b otivating students to have course achieveent. Soe students show good values for the different event activities, including exercise practice, but do not have evaluations. General results of the research were well evaluated b acadeic experts taking into account the validit, novelt and siplicit of the knowledge. All these knowledge of global and particular contents could be used to iprove the e-learning sste in different aspects. Strategies to encourage interactivit between students, strategies to design an assessent ethodolog that reinforce the student learning stles detected, and global iproveents of different coponents of the e-learning sste towards a ore distributed interactive learning could be proposed. Considering the findings of knowledge, a preliinar set of strategies was outlined. Future work

19 As a prototpe, the stud has ielded encouraging results on the application of knowledge discover to e-learning analsis. Nevertheless, in order to obtain a coplete application of this analsis is necessar to copleent the datawarehouse with ore variables. Thus the coplexit of the analsis of the research topics can be ore realisticall odelled. Aong these variables could be gender, age, location, enrolent date, likes and dislikes, and so on. Besides of teacher s data, such as, course s surve results, research topics. Those variables would be collected through questionnaires, or transferring autoaticall fro databases. The part of teaching stles of the Felder s learning fraework or another educational odel has to be incorporated in the proposed ethodolog. The tuning of the learning and teaching stles to obtain a good perforance in the outcoe of the process would be odelled. Depending on the ixture of learning and teaching stles, several adaptations of the pedagogical e-learning resources could be ade. The results of the enhanced odel could be used to adapt teaching ethodologies, including the critical aspect of the assessent stle, or in general to iprove the e-learning sste, balancing distributed passive learning (DPL) and distributed interactive learning (DIL). The seantic inforation and the iplicit inforatics ontologies defined b clusters descriptions, decision trees, and learning stles conceptions found in the research, could be used to ipleent a knowledge-based sste and/or a standard ontolog of the studied doain. The ontolog could be used to exchange and reuse the knowledge and it would ake eas to increase, foster, and update the knowledge obtained fro the web e-learning activities. The use of a web ontolog language and standard data interchange forats would ake possible the approach to the seantic web. ACKNOLEDGEMENTS Special thanks go to the Universidad Politécnica Abierta personnel for giving the web data and inforation about the virtual capus. This work has been supported b Spanish Adinistration under grant TEC FUTURE RESEARCH DIRECTIONS The chapter has discussed the knowledge discover in e-learning considering several subjects as: e-learning web activities, data preprocessing, data ining techniques, knowledge evaluation, learning and teaching stles, pedagogical innovation, and inforatics ontologies. The balance between interactive and personal activities is a critical factor for e-learning sstes (distributed passive learning (DPL) and distributed interactive learning (DIL)). An interesting area of research is the proposal of new e-learning activities or deterining the suitable ixture of those activities considering, for instance, contents, ultiedia resources, and ubiquitous networks. The qualit of the discovered knowledge is directl proportional to the cleanness and relevance of data. E-learning processes could generate a lot of useless inforation, so efficient algoriths to filtering and suarizing data; to resolve inconsistencies, to estiate issing data, and solving data heterogeneit are valuable for the knowledge discover approach. Pattern recognition is a wide area that includes an kinds of achine learning algoriths. Independent coponent analsis (ICA) algoriths, as the applied in the present chapter, have ielded iportant results in areas as iage filtering and

20 segentation, brain to coputer interface, and electrocardiographic diagnosis. Recentl the ixture of ICAs has eerged as a flexible generating odel to arbitrar data densities using ixtures of Gaussians or Laplacians distributions or non-paraetric distributions for the coponents. Those ICA ixtures could be used to odel data or knowledge in the web. Usuall the evaluation of the knowledge is ade b experts. Nowadas, aspects as novelt or interestingness are estiated b novelt detection algoriths. Those algoriths could be used b intelligent agents in the web in order to ake decisions considering user behaviours. In the field of e-learning it is a novel approach. Recentl, second level patterns in data ining have been studied. Those approaches have been used in protein and ADN research, where the results of a first level of data ining confors a huge knowledge doain. In web applications that kind of ethods would be useful. In addition the autoatic conversion of knowledge or seantic inforation obtained b webining techniques, represented b structures as the decision trees, to ontologies would ake possible the exchange and reuse of the doain knowledge. Methodologies to create hierarchical structures of patterns are suitable to create ontologies, to ake that possible, conversion procedures to translate statistical inforation to standard data interchange forats are needed. All of those approaches a contribute to develop the seantic web. REFERENCES Barker, K., Trafalis, T., & Rhoads, T. R. (2004). Learning fro student odel. Sste and Inforation Engineering Design Sposiu, (pp ). Beck, J. E., Jia, P., Sison, J., & Mostow, J. (2003). Predicting student help-request behavior in an intelligent tutor for reading. 9th International Conference on User Modelling. Beck, J. E. & oolf, B. P. (1998). Using a learning agent with a student odel. Lecture Notes in Coputer Science, 1452, Bezdek J.C & Pal S.K (1992). Fuzz Models for Pattern Recognition: Methods That Search for Structures in Data. New York: IEEE Press. Borst,. N. (1997). Construction of engineering ontologies for knowledge sharing and reuse. Universit of Tweent, NL-Centre for Teleática and Inforation Technolog. Cabena, P., Hadjnian, Stadler, R., Verhees, J., & Zanasi, A. (1997). Discovering Data Mining: fro concept to ipleentation (IBM books). Pearson Education. Caragea, D., Pathak, J., & Honavar, V. (2004). Learning classifiers fro seanticall heterogeneous data. Lecture Notes in Coputer Science, 3291, Cichocki, A. & Aari, S. (2001). Adaptive Blind Signal and Iage Processing: Learning algoriths and applications. New York: ile, John & Sons. Duda, R., Hart, P. E., & Stork, D. G. (2000). Pattern Classification. (2 ed.) ile-interscience. Elton, L. R. B. & Laurillard, D. M. (1979). Trends in research on student learning. Studies in Higher Education, 4(1), Faad, U., Piatetsk-Shapiro, G., Sth, P., & Uthurusa, R. (1996). Advances in Knowledge Discover and Data Mining. New York: The MIT Press. Felder, R. & Silveran, L. (1988). Learning and teaching stles. Journal of Engineering Education, 78(7),