Intelligent web-based learning system with personalized
learning path guidance
Chih-Ming Chen
*Graduate Institute of Library, Information and Archival Studies, National Chengchi University, No. 64, Sec. 2, ZhiNan Rd., Wenshan District, Taipei City 116, Taiwan, ROC
Received 10 April 2007; received in revised form 8 August 2007; accepted 10 August 2007
Abstract
Personalized curriculum sequencing is an important research issue for web-based learning systems because no fixed learning paths will be appropriate for all learners. Therefore, many researchers focused on developing e-learning systems with personalized learning mechanisms to assist on-line web-based learning and adaptively provide learning paths in order to promote the learning performance of individual learners. However, most personalized e-learning systems usually neglect to consider if learner ability and the difficulty level of the recommended courseware are matched to each other while per-forming personalized learning services. Moreover, the problem of concept continuity of learning paths also needs to be considered while implementing personalized curriculum sequencing because smooth learning paths enhance the linked strength between learning concepts. Generally, inappropriate courseware leads to learner cognitive overload or disorien-tation during learning processes, thus reducing learning performance. Therefore, compared to the freely browsing learning mode without any personalized learning path guidance used in most web-based learning systems, this paper assesses whether the proposed genetic-based personalized e-learning system, which can generate appropriate learning paths accord-ing to the incorrect testaccord-ing responses of an individual learner in a pre-test, provides benefits in terms of learnaccord-ing perfor-mance promotion while learning. Based on the results of pre-test, the proposed genetic-based personalized e-learning system can conduct personalized curriculum sequencing through simultaneously considering courseware difficulty level and the concept continuity of learning paths to support web-based learning. Experimental results indicated that applying the proposed genetic-based personalized e-learning system for web-based learning is superior to the freely browsing learn-ing mode because of high quality and concise learnlearn-ing path for individual learners.
Ó 2007 Elsevier Ltd. All rights reserved.
Keywords: Genetic algorithm; Personalized learning path; Intelligent tutoring system; Web-based learning
0360-1315/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.compedu.2007.08.004
* Tel.: +886 2 29393091x88024; fax: +886 2 29384704. E-mail address:chencm@nccu.edu.tw
Computers & Education 51 (2008) 787–814
1. Introduction
Traditional teaching resources, such as textbooks, typically guide the learners to follow fixed sequences to other subject-related sections related to the current one during learning processes. Web-based instruc-tion researchers have given considerable atteninstruc-tion to flexible curriculum sequencing control to provide adaptable, personalized learning programs (Brusilovsky, Eklund, & Schwarz, 1998; Jih, 1996; Lee, 2001; Lin & Hsieh, 2001; Mia & Woolf, 1998; Papanikolaou & Grigoriadou, 2002; Tang et al., 2000; Tang & Mccalla, 2003). Curriculum sequencing aims to provide an optimal learning path to individual learners since every learner has different prior background knowledge, preferences, and often various learning goals (Brusilovsky & Vassileva, 2003; Chen, Lee, & Chen, 2005; Roland, 2000; Weber & Specht, 1997). In an educational adaptive hypermedia system, an optimal learning path aims to maximize a com-bination of the learner’s understanding of courseware and the efficiency of learning the courseware (Roland, 2000).
Moreover, as numerous web-based tutoring systems have been developed, a great quantity of hypermedia in courseware has created information, and cognitive overload and disorientation (Berghel, 1997; Borchers, Herlocker, Konstanand, & Riedl, 1998), such that learners are unable to learn very efficiently. To aid more efficient learning, many powerful personalized/adaptive guidance mechanisms, such as adaptive presentation, adaptive navigation support, curriculum sequencing, and intelligent analysis of student’s solutions, have been proposed (Chen et al., 2005; Papanikolaou & Grigoriadou, 2002; Tang & Mccalla, 2003; Weber & Specht, 1997). Nowadays, most adaptive/personalized tutoring systems (Lee, 2001; Papanikolaou & Grigoriadou, 2002; Tang & Mccalla, 2003) consider learner/user preferences, interests, and browsing behaviors when inves-tigating learner behaviors for personalized services. However, these systems neglect the importance of learner ability when implementing personalized mechanisms. On the other hand, some researchers emphasized that personalization should consider levels of learner knowledge, especially in relation to learning (Chen et al., 2005; Chen, Liu, & Chang, 2006; Papanikolaou & Grigoriadou, 2002). That is, the abilities of individuals may be based on major fields and subjects. Therefore, considering learner ability can promote personalized learning performance.
Over the years, designers of web-based learning have evolved several common lesson structures for different learning occasions. These lesson structures include the classic tutorial lessons, active-centered lessons, learner-customized tutorial lessons, knowledge-placed tutorial lessons, exploratory tutorial lessons, and generated les-sons (Horton, 2000). Among the six kinds of lessons, the generated lessons aim to customize learning for those who have very specific needs and not much time or patience to complete topics they have learned (Horton, 2000). The generated lessons tailors a learning sequence based on the learner’s answers to questions on a pre-test or questionnaire at the start of the lesson (Horton, 2000). To construct a personalized learning path based on simultaneously considering courseware difficulty level and learning concept continuity during ing processes, a genetic-based curriculum sequence scheme is here presented to customize personalized learn-ing path. The proposed approach is based on a pre-test to collect incorrect learnlearn-ing concepts of learners through some randomly selecting testing items (Hsu & Sadock, 1985), then the genetic algorithm is employed to construct a near optimal learning path according to these incorrect response patterns of pre-test. The goal of this study aims to help learners learn more effectively and efficiently by skipping the learning concepts that learner has given correct responses for the corresponding testing items in a pre-test process. Since the fitness function of genetic algorithm is determined by the difficulty parameter of courseware and the concept relation degree between two successive courseware in a generated learning path, the proposed curriculum sequencing scheme can generate high quality learning paths for individual learners. Experimental results indicated that the proposed genetic-based personalized e-learning system with curriculum sequencing mechanism generates appropriate course materials to learners based on individual learners’ requirements, and help them learn more effectively and efficiently in a web-based learning environment.
2. System architecture
This section describes the system architecture, system components, and details of the learning procedures for the proposed genetic-based personalized e-learning system.
2.1. System architecture and components
A personalized e-learning system based on the proposed genetic-based curriculum sequencing scheme, which includes an off-line courseware modeling process, six intelligent agents and four databases, is presented herein. The six intelligent agents are the learning interface agent, pre-test process agent, learning path gener-ation agent, adaptive naviggener-ation support agent, post-test process agent and courseware management agent, respectively. These four databases include the user account database, user profile database, testing items and courseware database and teacher account database. The learner interface agent aims at providing a flex-ible learning interface for learners to interact with the pre-test process agent, adaptive navigation support agent and post-test process agent. The pre-test process agent aims to generate randomly a testing item for the corresponding learning courseware in order to identify the incorrect learning concepts of individual learner according to the incorrect testing responses for personalized curriculum sequencing. In the meanwhile, the pre-test process agent will pass these incorrect pre-testing responses of individual learner to the learning path gener-ation agent to plan a personalized learning path based on the proposed genetic-based curriculum sequencing scheme. Moreover, the adaptive navigation support agent is in charge of guiding the learner’s learning process based on a learning path generated by the learning path generation agent and storing learning records into the user profile database. In addition, the post-test process agent provides a final test while the learner finishes the whole learning process. The courseware management agent with authorized account management mechanism provides a responsive testing items and courseware management interface, aiding teachers to create new test-ing items and course units, upload testtest-ing items and courseware to the testtest-ing items and courseware database and delete or modify testing items and courseware from the testing items and courseware database. The system architecture is shown asFig. 1.
To implement the proposed genetic-based personalized e-learning system by agent techniques, interaction is one of the most important features of an agent-based system (Nwana, 1996). Agent-based systems recurrently interact to share information and to perform tasks to achieve their goals by agent communication language. There are two main approaches including procedural and declarative schemes to designing an agent communication language (Genesereth, 1997). The procedural approach, where communication is based on
Off-line testing items and courseware analysis Testing items / Courseware Database Teacher Account Database Testing items / Courseware Management Agent Testing items / Courseware Modeling Process 1 8 Learner Account Database User Profile Database Learning Interface Agent Pretest Process Agent Learning Path Generate Agent Adaptive Navigation Support Agent Post-test Process Agent 4/10/19/25/32 5 9/18/24/31/35 11 16/22 20/26 30/34 33 6 17/23 15 Off-line testing items and courseware analysis Testing items / Courseware Database Teacher Account Database Testing items / Courseware Management Agent Testing items / Courseware Modeling Process 1 8 Learner Account Database User Profile Database Learning Interface Agent Pretest Process Agent Learning Path Generate Agent Adaptive Navigation Support Agent Post-test Process Agent 4/10/19/25/32 5 9/18/24/31/35 11 30/34 33 6 17/23
Off-line testing items and courseware analysis Testing items / Courseware Database Teacher Account Database Testing items / Courseware Management Agent Testing items / Courseware Modeling Process 1 2 Off-line testing items and
courseware analysis Testing items / Courseware Database Teacher Account Database Testing items / Courseware Management Agent Testing items / Courseware Modeling Process
Off-line testing items and courseware analysis Testing items / Courseware Database Testing items / Courseware Database Teacher Account Database Teacher Account Database Testing items / Courseware Management Agent Testing items / Courseware Modeling Process 1 3 7 13 8 Learner Account Database User Profile Database Learning Interface Agent Pretest Process Agent Learning Path Generate Agent Adaptive Navigation Support Agent Post-test Process Agent 4/10/19/25/32 5 9/18/24/31/35 11 30/34 33 6 17/23 8 8 Learner Account Database User Profile Database Learning Interface Agent Pretest Process Agent Learning Path Generate Agent Adaptive Navigation Support Agent Post-test Process Agent 4/10/19/25/32 5 9/18/24/31/35 11 30/34 33 6 17/23 Learner Account Database User Profile Database Learning Interface Agent Pretest Process Agent Learning Path Generate Agent Adaptive Navigation Support Agent Post-test Process Agent Learner Account Database Learner Account Database User Profile Database User Profile Database Learning Interface Agent Pretest Process Agent Learning Path Generate Agent Adaptive Navigation Support Agent Post-test Process Agent 4/10/19/25/32 5 9/18/24/31/35 11 30/34 33 6 17/23 4/10/19/25/32 4/10/19/25/32 5 5 9/18/24/31/35 9/18/24/31/35 1111 12 14 21/27 28 30/34 30/34 33 33 29 6 6 17/23 17/23
Fig. 1. The system architecture of the genetic-based personalized e-learning system (the numeral marked in this figure represents the system operation procedure).
executable content, could be accomplished using programming languages. The declarative approach is based on declarative statements, such as requesting or commanding, to accomplish agent communication. One of the more popular declarative agent languages is Knowledge Query and Manipulation Language (KQML). To avoid implementing complicated message format and a message handling protocol defined in the declara-tive-based agent communication language, this study used the procedural approach to implement the pro-posed genetic-based personalized e-learning system.
2.2. System operation procedures
Based on the system architecture mentioned-above, the system operation procedures are briefly described as follows:
Step 1. The testing items were designed by courseware experts based on the course materials stored in the testing items and courseware database. According to Item Response Theory (IRT) (Baker Frank, 1992), the difficulty parameters of these testing items can be determined through statis-tical method based on testing results of learners. After that, courseware with web page type can be designed according to the conveying concept of the corresponding testing item. The detailed courseware modeling process is described in Section3.
Step 2. Teachers login the system to upload, delete or revise testing items and courseware in the testing items and courseware database by the legal teachers’ accounts.
Step 3. The designed courseware are maintained and stored into the testing items and courseware data-base through the courseware management agent.
Step 4. Learners login the system through the learning interface agent by the legal learners’ accounts. Step 5. After a learner logs in the system, the learning interface agent checks whether his/her account
stored in the user account database.
Step 6. If the learner has already owned a registered account, the system will get his learning profile from the user profile database and guide the learner to perform the previous unfinished learning courseware; otherwise, the proposed system will treat the learner as a beginner who must accept a pre-test.
Step 7. For a beginner, the proposed system will randomly generate a test sheet based on the testing items stored in the testing items and courseware database for the learner and guide the learner to perform a pre-test.
Step 8. The generated test sheet is transformed to the user interface agent for a learner to conduct a pre-test.
Step 9. The learner performs the pre-test through the user interface agent.
Step 10. The learning interface agent transfers the pre-test results to the pre-test process agent.
Steps 11–12. The pre-test process agent analyzes the pre-test results and conveys the incorrect testing responses to the learning path generation agent for personalized curriculum sequencing.
Steps 13–14. The learning path generation agent plans a learning path according to the incorrect pre-test results of an individual learner transformed from the pre-test process agent based on the pro-posed genetic-based curriculum sequencing approach. Simultaneously, the generated learning path is also stored into the user profile database and conveyed to the adaptive navigation sup-port agent for the courseware learning of individual learner.
Steps 15–16. The adaptive navigation support agent takes charge of guiding the learning path of individual learner according to the generated learning path through the designed control mechanisms in the proposed genetic-based personalized e-learning system.
Steps 17–21. The adaptive navigation support agent communicates with the learning interface agent to guide the learning contents according to the planned learning path for individual learner. Meanwhile, the learning processes of individual learner are also recorded into the user profile database.
Steps 22–27. The learner repeats the same learning procedures mentioned in Steps 15–21 until the learner finishes all courseware planed by the learning path generation agent.
Steps 28–29. After the learner finishes the entire courseware planed by the learning path generation agent, the adaptive navigation support agent will notice the post-test process agent to randomly gen-erate a testing sheet to the learner for performing a post-test in order to evaluate the learning performance.
Steps 30–35. The generated testing sheet in a post-test will be transformed to the learning interface agent, and then displayed to the learner. The post-test results are also provided to the learner for self-examination and stored into the user profile database. So far, the learner finishes the entire learning process for a learning course unit.
3. Courseware modeling process
The courseware modeling process presents a detailed courseware design procedure to establish the difficulty parameters of courseware and courseware contents for personalized courseware generation. This study pre-sents a statistics-based method derived from computerized adaptive testing (CAT) theory (Hsu & Sadock, 1985) through a conscientious test process to determine the difficulty parameters of courseware. The detailed flowchart of the courseware modeling process is illustrated asFig. 2.
To design a course of the course unit ‘‘Fraction’’ of elementary school mathematics in Taiwan as an exam-ple, several experienced teachers were invited as courseware experts to analyze the primary concepts for the course unit ‘‘Fraction’’ in the courseware modeling process. The courseware experts designed the correspond-ing testcorrespond-ing item for each learncorrespond-ing concept. That is, the testcorrespond-ing items are regarded as key characteristic of the corresponding learning content. Additionally, about 500 elementary school examinees who had majored in the course unit ‘‘Fraction’’ were invited to join the exam, which contains 17 testing items to cover those learning concepts. According to the Item Response Theory (Baker Frank, 1992; Hsu & Sadock, 1985) in CAT, their testing data was analyzed by the statistics-based BILOG program to obtain the appropriate difficulty param-eters for these testing items. After that, the web page of courseware was designed following the conveying con-tent of the corresponding testing item. Since the concon-tent of courseware is derived from the concept of the testing item, it is assumed the difficulty of courseware equals the difficulty of the corresponding testing item. That is, each testing item in the testing item database has a corresponding courseware that conveys the learn-ing concept of the correspondlearn-ing testlearn-ing item.
Collect testing data
Obtain difficulty parameters Analyze learning contents Testing items database 3. Exam 5. Record 6. Analyze Design courseware with the corresponding difficulty parameter Courseware database Design testing items for
learning contents 1 2 4 7 course ID course 1 course 2 course 3 difficulty 0.5 0.3 -0.8 : : 8
4. Evaluating concept relation degrees among courseware
In order to facilitate easier courseware concept relation analysis, all courseware in the courseware database has followed the standard of the metadata information model of Sharable Content Object Reference Model (SCORM) 1.2 (SCORM version 1.2-The SCORM Content Aggregation Model, 2001). Restated, each course-ware in the coursecourse-ware database has a corresponding XML binding file to record important SCORM meta-data, which conveys the main courseware concept. In the meanwhile, this study also developed an interface for teachers to maintain the SCORM metadata for the relevant courseware. In order to generate a near optimal learning path for a learner based on the results of pre-test, these SCORM metadata are applied to calculate the concept relation degrees among courseware by using Chinese natural language processing (An Extension Chi-nese Lexicon Scanner, 2006) and information retrieval (Frakes & Baeza-Yates, 1992) methods.Fig. 3 illus-trates the maintained interface of SCORM metadata. Next, how to compute the concept relation degrees for personalized courseware generation will be explained in detail.
4.1. Metadata preprocessing
First, two metadata fields of the corresponding XML binding file of courseware are selected to represent the conveyed learning concept for a courseware. They are keyword and description fields in the SCORM 1.2 meta-data information model shown asFig. 3, respectively. In order to calculate the concept relation degrees for personalized courseware generation, metadata preprocessing is required because the description field in the SCORM 1.2 metadata information model is described by Chinese natural language in this study. Thus, the first phase of metadata preprocessing aims to perform Chinese word segmentation by an ECScanner (An
Extension Chinese Lexicon Scanner, 2006) in order to describe the metadata field of the corresponding XML binding file of courseware so that separated linguistic terms can be obtained. The second phase of metadata preprocessing filters out non-textual words (e.g. numeric data, symbols, notation and ASCII drawings) and one-word terms because they do not carry any usable information for calculating concept relation degrees. Fig. 4shows the details of the metadata preprocessing procedure.
4.2. Estimation of concept relation degree
To estimate the concept relation degree between two courseware, the vector space model (Frakes & Baeza-Yates, 1992) is applied to represent each courseware as vectors in a multidimensional Euclidean space. Each axis in this space corresponds to a linguistic term obtained from Chinese word segmentation process. The coordinate of the ith courseware in the direction corresponding to the kth linguistic term can be determined as follows:
wik ¼ tfik log
N dfk
¼ tfik IDF ð1Þ
where wikrepresents the importance/weight of the kth term in the ith courseware, tfikis term frequency of the
kth term, which appears in the ith courseware; N denotes the total number of courseware in a course unit, dfk
is the document frequency of the kth term, which appears in a course unit.
Assume that there are total m terms under union of all linguistic terms of the ith courseware and jth course-ware. The concept relation degree for the ith and jth courseware can be found by using the cosine-measure, and formulated as follows:
rij ¼ Pm h¼1wihwjh ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pm h¼1w2ih Pm h¼1w2jh q ð2Þ
where ci¼ hwi1; wi2; . . . ; wih; . . . ; wimi and cj¼ hwj1; wj2; . . . ; wjh; . . . ; wjmi, respectively, represent the vectors in
a multidimensional Euclidean space for the ith and jth courseware, rijdenotes the concept relation degree
be-tween the ith and jth courseware.
Assume that there are totally n courseware in the courseware database, the concept relation matrix for all courseware can be expressed by the matrix R, and listed as follows:
R¼ c1 c2 cn c1 c2 .. . cn r11 r12 r1n r21 r22 . . . r2n .. . .. . .. . .. . rn1 rn2 rnn 2 6 6 6 4 3 7 7 7 5 nn ð3Þ XML binding file
Extract the description and keyword fields
Perform the word segmentation by
ECScanner
Filter out the stop words and non-textual words
Store the meaningful terms into datatable
Courseware concept releation
degree table
In summary, the entire procedure of estimating the concept relation degree based on SCORM 1.2 metadata information model can be displayed asFig. 5.
5. Personalized learning path generation based on genetic algorithm
This section explains how to generate a personalized learning path for an individual learner utilizing the genetic algorithm (Rothlauf, 2002).
5.1. Generated courseware for web-based learning
Generated courseware tailor a courseware to each learner based on answers to a pre-test before at the start of the learning course unit.Fig. 6displays the architecture of generated courseware for web-based learning. Generated courseware are helpful to individual learners for performing more efficient learning specially when learners have different needs, varying desires, and different levels of knowledge background. The details of the proposed genetic-based curriculum sequence scheme are presented in the next subsection.
5.2. Genetic algorithm for personalized learning path generation 5.2.1. Definition of individual strings
First, a serial number is assigned to each courseware from 1 to n if there are totally n courseware in the testing items and courseware database for personalized learning path generation. The integer-coded scheme
Courseware concept releation
degree table
Get the terms from courseware datatable
Count each term ’s document frequency in a course unit (dfk)
Courseware concept releation
degree table Compu te each term ’s inverse
document frequency in a course unit (IDF)
Count each term ’s frequency in the ithcoursewares (tf
ik)
Compute each term ’s weight in the ithcoursewares (wik)
Estimate the concept relation degree of the ith, jthcoursewares
(rij)
Estimate the concept relation matrix of all coursewares (R)
Store the concept relation matrix into the datatable
was employed to represent an individual string herein, i.e. a potential solution for the genetic algorithm. Thus, the assigned serial number of each courseware combined with the serial numbers of other courseware forms a chromosome to represent a generated learning path for personalized curriculum sequencing. That is, an indi-vidual consists of a single chromosome herein. The whole indiindi-vidual consisted of the genes of all courseware serial numbers for the genetic algorithm is illustrated asFig. 7. InFig. 7, the assigned serial number of each courseware is viewed as a gene in the chromosome.
5.2.2. Initial population size
Generally, the initial population size can be determined according to the complexity of the solved problem. A larger population size reduces the searching speed of genetic algorithm, but it could increase the probability of finding high quality solution. To plan a learning path with high quality for an individual learner, the initial population size was chosen as one hundred for personalized learning path generation in this study.
5.2.3. Selecting fitness function
Fitness function is a performance index that it was applied to judge the quality of a generated learning path for the genetic algorithm in the study. In order to generate a personalized learning path with high quality for an individual learner based on the pre-test results, the difficulty parameters of courseware and the concept relation degrees of courseware are simultaneously considered to determine the fitness function. In our method, a learning path constructed by the genetic algorithm only considers the mapped courseware that learner gives incorrect pre-test results. Moreover, the courseware with the smallest difficulty parameter is always selected as the first courseware ranked in a generated learning path. The proposed fitness function is formulated as follows:
f ¼X
n
i¼2
ðð1 wÞ rði1Þiþ w ð1 biÞÞ ð4Þ
where f is the proposed fitness function for personalized learning path generation by the genetic algorithm,
r(i1)irepresents the concept relation degree of the (i 1)th courseware with the ith courseware in a generated
learning path, biis the difficulty parameter of the ith courseware, w is a adjustable weight, and n stands for the
total number of courseware considered for personalized learning path generation. Introduction Test
ok
Topic A
Topic B Topic E
Topic C Topic D Topic F Topic G
Topic H Topic I
Summary Test
Fig. 6. The generated courseware for web-based learning (Horton, 2000).
Serial number of courseware 1 Serial number of courseware 2 Serial number of courseware n
The first gene The nth gene
the whole individual
5.2.4. Reproduction operation
In the reproduction operation, the individuals with large fitness function value have a relatively higher probability to reproduce next generation. This operation aims to choose good individuals to achieve the goal of gene evolution. The most common used method of reproduction operation is the weighted roulette selection scheme (Rothlauf, 2002) and the scheme was also employed to perform reproduction operation in this study. 5.2.5. Crossover operation
This operation aims to combine two parent individuals to evolve better child individuals. In this study, the uniform crossover scheme (Rothlauf, 2002) was employed to perform the crossover operation and the prob-ability of crossover is set to be 0.9. Meanwhile, to avoid generating illegal learning paths while performing the crossover operation, i.e. a learning path contains duplicate serial number of courseware or a learning path contains any serial number of courseware that is over the total number of courseware, two randomly selected serial numbers of genes in two individuals exchange genes to each other by probability decision.Fig. 8 illus-trates an example of the proposed crossover operation. Restated, using the proposed crossover operation can guarantee to obtain a logical learning path.
5.2.6. Mutation operation
In the proposed mutation operation, two randomly selected genes in an individual are forced to exchange the gene to each other under probability decision. The proposed mutation operation is similar to the mutation operation of swapping two-points implemented in the standard genetic algorithm. The only difference is that the binary-coded scheme is employed in the standard genetic algorithm, but the integer-coded scheme was employed to represent an individual string, i.e. a potential solution for the genetic algorithm, in the study. This scheme can avoid generating illegal learning paths mentioned in the previous subsection. The mutation oper-ation can evolve some new individuals that might not be produced by the operoper-ations of reproduction and crossover to avoid that the solution traps into the local optimum. Generally, a low probability of mutation can guarantee the convergence of genetic algorithm, but it may lead to poor quality solution. By contrast, a high probability of mutation may lead to the phenomenon of random walk in the genetic algorithm, thus reducing convergence speed. In this paper, the probability of mutation is set to be 0.001. Fig. 9illustrates an example of the proposed mutation operation.
5.2.7. Stop criterion
The genetic algorithm repeatedly runs the reproduction, crossover, mutation and replacement operations until it meets an assigned stop criterion. In this study, the stop criterion is set to be 200 generations because this criterion can obtain satisfied learning paths for personalized learning path generation.
the lth individual
The first gene The nth gene
the mth individual
The first gene The nth gene
crossover serial number of courseware i serial number of courseware i serial number of courseware k serial number of courseware k serial number of courseware 1 serial number of courseware n serial number of courseware 8 serial number of courseware 1 serial number of courseware 5 serial number of courseware n
5.3. Procedure of personalized learning path generation
In summary, the procedures of the proposed genetic-based personalized learning path generation scheme are detailed as follows:
Step 1. A learner performs a pre-test based on randomly selected testing items in a course unit for personal-ized learning path generation.
Step 2. The proposed system collects the incorrect testing items in the pre-test and their corresponding courseware in the testing items and courseware database.
Step 3. The corresponding courseware with the smallest difficulty parameter among the incorrect testing items is selected as the first courseware for personalized learning path generation.
Step 4. The system generates a near optimal learning path for an individual learner utilizing the genetic algo-rithm according to the incorrect response testing items.
Step 5. A learner performs personalized web-based learning according to the generated learning path. Step 6. Terminate the learning process if the learner finishes courseware learning of the generated learning
path; otherwise, return Step 1 for next learning cycle. 6. Experiments
Currently, the proposed genetic-based personalized e-learning system is available on the web to simulta-neously provide both the freely browsing learning mode and the learning mode of curriculum sequencing rec-ommendation. To verify the quality and effectiveness of planned learning path in the learning mode of curriculum sequencing recommendation for personalized web-based instruction, some elementary school stu-dents who were majoring in the course unit of ‘‘Fraction’’ of elementary school mathematics were invited to test this system. The detailed functions of this system and experimental results are described as follows. 6.1. The developmental environment of software and hardware
In this study, AppServ package (AppServ Open Project, 2007) was employed as the development tool to implement the proposed genetic-based personalized e-learning system. The software package can simulta-neously support to install the development tools including apache server, PHP analyzer, MySQL database, and MySQL database management system to promote the development speed of web applications. It is suit-able to be employed to develop web-based learning system with client–server architecture. The designed genetic-based personalized e-learning system contains a server side for learning content services and client side for learner learning. The detailed specifications of software and hardware in both the server and client sides are listed inTable 1.
6.2. The implemented genetic-based personalized e-learning system
To explain how to perform the learning processes using a generated learning path for an individual learner, this section briefly introduces the learning procedure on the implemented genetic-based personalized
The first gene The nth gene
Serial number of courseware i Serial number of courseware 1 Serial number of courseware n Serial number of courseware k Mutation
e-learning system.Fig. 10shows the entire layout of the user’s learning interface. As a learner logins this sys-tem, he/she must conduct a pre-test if he/she is a beginner; otherwise, the system will guide the learners to learn the courseware according to the previous unfinished learning procedures.Fig. 11 shows the interface of performing a pre-test for a beginner. After a beginner finishes a pre-test, the system will analyze the results of pre-test, then the system will generate an appropriate learning path based on the incorrect testing responses of individual learner in a pre-test process.Fig. 12displays a generated learning path with learning priority according to the incorrect testing responses of an individual learner. The system will guide the learner to per-form the learning process according to the learning path generated for the learner. Particularly, a learner must follow a learning path planned by the genetic-based personalized e-learning system to learn the corresponding courseware with incorrect testing responses. The genetic-based personalized e-learning system will temporally disable the courseware that their ranking priorities are less than the priority of the current learning courseware until the current learning courseware has been acquired.Fig. 13displays the learning courseware with the first learning priority that has not been acquired by the learner in the generated learning path. Currently, course material organized on web pages with flash animation and synchronous voice comments is the course element in the proposed system. Moreover, one randomly selected testing item related to the current learning Table 1
The specifications of server and client sides Server side
1 Host Middle-level server HP ML-370
2 Operating system Windows server 2000
3 Web server Apache 1.3.29
4 Database MySQL 4.0.16
5 Language PHP 4.3.4
Client side
1 Host Multimedia personal computer
2 Operating system Windows XP
3 Browser IE 6.0
4 Other peripheral equipment Speaker or earphone
courseware is arranged in the bottom-right window to help system get the learner’s comprehension degree for the learning courseware. When a learner can pass the corresponding test question of some learning course-ware, this indicates that the learner has acquired the learning courseware. If a learner cannot pass two ran-domly selected testing questions for some learning courseware, the genetic-based personalized e-learning will guide the learner to conduct the remedy learning. In this work, the courseware database contains course materials with easier difficulty level than the current learning courseware used for supporting remedy learning.
Fig. 11. A pre-test for a beginner.
The remedial course materials convey similar learning concepts with the current learning courseware, but they contain different learning content. The remedy learning mechanism aims at improving the learning perfor-mance of individual learners for the courseware that they cannot acquire well through the standard course-ware. Fig. 14 reveals the learner interface of the freely browsing learning mode. In the freely browsing learning mode, no any learning path is planned for an individual learner and no any course material is dis-abled to forbid browsing, thus learners can freely click any course material for learning.
Fig. 13. The courseware with first learning priority in the generated learning path.
6.3. An example for personalized learning path generation
This section gives an example to show how to plan a learning path for an individual learner according to the incorrect testing responses in a pre-test. First, the course modeling process mentioned in Section3 is used to determine the difficulty parameter of each courseware. Restated, courseware organized on a single web page is the smallest course element in the proposed personalized courseware generation approach. In our experiments, the course unit ‘‘Fraction’’ of elementary school mathematics in Taiwan is used to gen-erate personalized learning path, which includes many courseware with various levels of difficulty to con-vey the concept of the ‘‘Fraction’’. Assume that a pre-test in the course unit ‘‘Fraction’’ is performed by a learner, and totally occurs 17 incorrect testing items. Assume that Table 2 illustrates the concept relation degrees of corresponding courseware that learner gives incorrect testing item responses. Table 3 lists the titles of corresponding courseware and their difficulty parameters that the learner gives incorrect testing item responses.
Based on the corresponding concept relation degrees and difficulty parameters listed inTables 2 and 3, the genetic algorithm was employed to construct a personalized learning path with high learning quality according to the proposed fitness function.Table 4illustrates the generated learning path by the genetic algorithm. This study found that the generated learning path recommends learning path with smooth learning concept to a learner under simultaneously considering the difficulty parameters of courseware and concept continuity. Restated, the learning concepts with high concept relation degree will be successively recommended during a learning process under simultaneously considering the difficulty parameters of courseware. This is very ben-eficial to a learner because it can guide the learner to achieve more effective and efficient learning. Additionally, Fig. 15 shows the convergence curve of the proposed fitness function using the genetic algorithm with the adjustable weight 0.7. This result demonstrates that the proposed genetic-based personalized learning path generation scheme can indeed generate a learning path with high quality for an individual learner to support personalized learning service.
6.4. Experiments
This section explains how to employ statistics method to assess the learning performance for the proposed genetic-based curriculum sequencing scheme.
6.4.1. Experimental design
To evaluate whether the proposed learning mode of curriculum sequencing recommendation is superior to the freely browsing learning mode, 220 three-grade elementary school students who were majoring in the ‘‘Fraction’’ unit in a mathematics course were invited to participate in the experiment.Table 5displays the statistics information for assessing learning performance in the experiment. Among 220 elementary school stu-dents, there are 92 students who were served as the control group to perform the freely browsing learning mode, and the remaining students were served as the treatment group to perform the proposed learning mode of curriculum sequencing recommendation only for the courseware with wrong answer responses in a pre-test. Both the learning modes simultaneously perform a pre-test and post-test for comparing the difference of learn-ing performance before and after learnlearn-ing.
In the experiment, teacher first detailed the system operation procedures for all participators in the first hour, and then all participators logged in the system to perform the planned learning process according to two different experimental groups from the following two to four hours. Each participator must follow three learning stages to complete the entire learning process, i.e. pre-test process, learning process, and post-test pro-cess, no matter what learning modes were used.Fig. 16exhibits the actual teaching scene at Hualien County Jiamin Elementary School in the experiment.
6.4.2. Experimental analysis
Since a part of participators had not completed the entire learning processes, this study thus fil-tered out these learning records. Table 6 lists the number of the participators who finished the entire learning processes or filled out the satisfaction investigation of questionnaire. Table 7 displays the
Table 2
The concept relation degrees for the incorrect testing items
rij C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C1 1 0.0118 0.1685 0.0173 0.0173 0.0760 0.0339 0.0509 0.0567 0.1847 0.0138 0.0329 0.0329 0.0114 0.0114 0.0114 0.0114 C2 0.0118 1 0.2062 0.6572 0.1988 0.0205 0.0091 0.0137 0.0026 0.0026 0 0.0050 0.0050 0 0 0 0 C3 0.1685 0.2062 1 0.1571 0.1571 0.0162 0.0072 0.0108 0.0020 0.0021 0 0.0039 0.0039 0 0 0 0 C4 0.0173 0.6572 0.1571 1 0.5517 0.0301 0.0134 0.0201 0.0038 0.0038 0 0.0073 0.0073 0 0 0 0 C5 0.0173 0.1988 0.1571 0.5517 1 0.0301 0.0134 0.0201 0.0038 0.0038 0 0.0073 0.0073 0 0 0 0 C6 0.0760 0.0205 0.0162 0.0301 0.0301 1 0.5561 0.7492 0.0325 0.0329 0.0286 0.0504 0.0504 0.0237 0.0237 0.0237 0.0237 C7 0.0339 0.0091 0.0072 0.0134 0.0134 0.5561 1 0.4837 0.0207 0.0210 0.0257 0.0413 0.0413 0.0213 0.0213 0.0213 0.0213 C8 0.0509 0.0137 0.0108 0.0201 0.0201 0.7492 0.4837 1 0.0124 0.0125 0.0261 0.0438 0.0438 0.0216 0.0216 0.0216 0.0216 C9 0.0567 0.0026 0.0020 0.0038 0.0038 0.0325 0.0207 0.0124 1 0.7833 0.0940 0.1394 0.1394 0.0780 0.0780 0.0780 0.0780 C10 0.1847 0.0026 0.0021 0.0038 0.0038 0.0329 0.0210 0.0125 0.7833 1 0.1309 0.1929 0.1929 0.1085 0.1085 0.1085 0.1085 C11 0.0138 0 0 0 0 0.0286 0.0257 0.0261 0.0940 0.1309 1 0.4809 0.4809 0.2285 0.2285 0.2285 0.2285 C12 0.0329 0.0050 0.0039 0.0073 0.0073 0.0504 0.0413 0.0438 0.1394 0.1929 0.4809 1 0.4158 0.4342 0.2342 0.4342 0.2342 C13 0.0329 0.0050 0.0039 0.0073 0.0073 0.0504 0.0413 0.0438 0.1394 0.1929 0.4809 0.4158 1 0.2342 0.4342 0.2342 0.4342 C14 0.0114 0 0 0 0 0.0237 0.0213 0.0216 0.0780 0.1085 0.2285 0.4342 0.2342 1 0.3619 0.4380 0.3619 C15 0.0114 0 0 0 0 0.0237 0.0213 0.0216 0.0780 0.1085 0.2285 0.2342 0.4342 0.3619 1 0.3619 0.4380 C16 0.0114 0 0 0 0 0.0237 0.0213 0.0216 0.0780 0.1085 0.2285 0.4342 0.2342 0.4380 0.3619 1 0.3619 C17 0.0114 0 0 0 0 0.0237 0.0213 0.0216 0.0780 0.1085 0.2285 0.2342 0.4342 0.3619 0.4380 0.3619 1 C.-M. Chen / Computers & Educati on 51 (2008) 787–814
comparison of learning performance for both the learning modes. The results reveal that the learning scores of 48.15% learners who learnt by the proposed learning mode of curriculum sequencing recom-mendation are progressive, but only 32% learners have progressive learning scores by the freely brows-ing learnbrows-ing mode. Next, the statistical information was utilized to further analyze the learnbrows-ing performance by statistical method.
In the work, the Matched-Pairs T-Tests was employed to analyze whether the freely browsing learning mode or the proposed learning mode of curriculum sequencing recommendation provides benefits in terms of learning performance promotion based on pre-test and post-test scores. Three cases are respectively dis-cussed as follows:
Table 3
The corresponding difficulty parameter for each courseware in the ‘‘Fraction’’ unit
Courseware Title of courseware Difficulty parameter
C1 Equal parts 1.8
C2 Division as sharing 1.5
C3 Division as separating 1
C4 Sharing with a remainder 0.1
C5 Separating with a remainder 0
C6 Parts of a whole 0.1
C7 Improper fractions 0.2
C8 Sequence of fractions 0.4
C9 Compare proper fractions with the same denominator 0.5
C10 Compare proper fractions with different denominators 0.7
C11 Add and subtract fractions 1.2
C12 Adding fractions 0.8 C13 Subtracting fractions 1 C14 Missing addend 1.3 C15 Missing subtrahend 1.5 C16 Missing summand 1.6 C17 Missing minuend 1.8 Table 4
The generated learning path by genetic algorithm with the adjustable weight 0.7
Learning path Difficulty
parameter
Concept relation degree between two successive courseware
C1 Equal parts 1.8 –
C11 Add and subtract fractions 1.2 0.0138
C8 Sequence of fractions 0.4 0.0261 C6 Parts of a whole 0.1 0.7492 C7 Improper fractions 0.2 0.5561 C12 Adding fractions 0.8 0.0413 C2 Division as sharing 1.5 0.0050 C3 Division as separating 1 0.2062
C9 Compare proper fractions with the same denominator
0.5 0.0020
C5 Separating with a remainder 0 0.0038
C14 Missing addend 1.3 0
C10 Compare proper fractions with different denominators
0.7 0.1085
C15 Missing subtrahend 1.5 0.1085
C17 Missing minuend 1.8 0.4380
C16 Missing summand 1.6 0.3619
C4 Sharing with a remainder 0.1 0
Case 1: The Matched-Pairs T-Tests for assessing the learning performance promotion of the freely browsing learning mode.
Table 8 lists the statistics information for the Matched-Pairs T-Tests of the control group to evaluate whether the freely browsing learning mode provides benefits in terms of learning per-formance promotion. Based on the goal, this study gave two research hypotheses for this case, and described as follows:
H0-Case 1: Suppose the learners who participated in the freely browsing learning mode have the same mean score in both the pre-test and post-test.
101.5 102 102.5 103 103.5 104 104.5 1 10 19 28 37 46 55 64 73 82 91 100 109 118 127 136 145 Iteration
Fitness function value
Fig. 15. Convergence curve of fitness function of genetic algorithm with the adjustable weight 0.7.
Table 5
The statistics information for assessing learning performance
Experimental group
Pre-test
Learning process
Post-test
Number of learners Control group (learning by the freely browsing learning
mode)
X Freely browsing a half of courseware in the designed course unit at least
X 92
Treatment group (learning by the proposed learning mode of curriculum sequencing recommendation)
X Only learning the courseware with wrong answer responses in a pre-test
X 128
H1-Case 1: Suppose the learners who participated in the freely browsing learning mode do not have the same mean score in both the pre-test and post-test.
Table 9shows the results of the Matched-Pairs T-Tests of the control group. The results indi-cate that the hypothesis H0 is satisfied under the significant level a = 0.05 and P = 0.273 > 0.05. That is, no any reasons can conclude that the freely browsing learning mode provides benefits in terms of learning performance promotion. Actually, our experiment shows that the mean score of the post-test of learners lowers 2.27 points.
Case 2: The Matched-Pairs T-Tests for assessing the learning performance promotion of the proposed learning mode of curriculum sequencing recommendation.
Table 10lists the statistics information for the Matched-Pairs T-Tests of the treatment group to evaluate whether the learning mode of curriculum sequencing recommendation provides bene-fits in terms of learning performance promotion. Therefore, this study also gave two research hypotheses for this case, and described as follows:
H0-Case 2: Suppose the learners who participated in the learning mode of curriculum sequencing recom-mendation have the same mean score in both the pre-test and post-test.
H1-Case 2: Suppose the learners who participated in the learning mode of curriculum sequencing recom-mendation do not have the same mean score in both the pre-test and post-test.
Table 11shows the results of the Matched-Pairs T-Tests of the treatment group. The results indicate that the hypothesis H0 is satisfied under the significant level a = 0.05 and P = 0.602 > 0.05. In other words, no any reasons can conclude that the learning mode of cur-riculum sequencing recommendation provides benefits in terms of learning performance pro-motion. However, our experiment shows that the mean score of the post-test of learners promotes 0.93 points.
Table 6
The statistical information for the participators who finished the planned learning processes and the satisfaction investigation
Analysis item Control group Treatment group
Finishing the entire learning process 75 108
Finishing the satisfaction investigation of questionnaire 61 103
Finishing the satisfaction investigation of learning mode 23
Table 7
Comparison of learning performance for both the learning modes Comparison item Learning mode
The freely browsing learning mode for the control group
The learning mode of curriculum sequencing recommendation for the treatment group
Number of learners 75 108
Number of learners with progressive score
24 (32.00%) 52 (48.15%)
Number of learners with retrogressive score
36 (48.00%) 36 (33.33%)
Number of learners with constant score
15 (20.00%) 20 (18.52%)
Table 8
The statistics information for the Matched-Pairs T-Tests of the control group
Learning mode Comparison item
Mean Number of learners Std. deviation Std. error mean
Pre-test for the freely browsing learning mode 73.80 75 23.20 2.68
Case 3: The Matched-Pairs T-Tests for assessing the learning performance promotion of the learners who have the same pre-test score in both the learning modes.
Table 12lists the statistics information for the Matched-Pairs T-Tests of both learning groups to evaluate whether the learning mode of curriculum sequencing recommendation provides Table 9
The results of the Matched-Pairs T-Tests of the control group
Paired differences t df Sig.
(2-tailed) Mean Std. deviation Std. error mean 95% Confidence interval of the difference Lower Upper Pair 1: Pre-test of freely browsing learning mode–post-test of
freely browsing learning mode
2.27 17.77 2.05 1.82 6.35 1.105 74 .273
Table 10
The statistics information for the Matched-Pairs T-Tests of the treatment group
Learning mode Comparison item
Mean Number of
learners
Std. deviation Std. error mean Pre-test for the learning mode of curriculum
sequencing recommendation
76.67 108 21.34 2.05
Post-test for the learning mode of curriculum sequencing recommendation
77.59 108 22.98 2.21
Table 11
The results of the Matched-Pairs T-Tests of the treatment group
Paired differences t df Sig. (2-tailed)
Mean Std. deviation Std. error mean 95% Confidence interval of the difference
Lower Upper Pair 1: Pre-test of the learning mode
of curriculum sequencing recommendation–post-test of the learning mode of curriculum sequencing recommendation
.93 18.41 1.77 4.44 2.59 .523 107 .602
Table 12
The statistics information for the Matched-Pairs T-Tests of both the control and treatment groups under the learners with the same pre-test score
Learning mode Comparison item
Mean Number of learners Std. deviation Std. error mean
Post-test for the freely browsing learning mode 72.32 69 25.96 3.13
Post-test for the learning mode of curriculum sequencing recommendation
benefits in terms of learning performance promotion for the learners with the same pre-test score in both the learning modes. Thus, this study gave the following two research hypotheses for the case, and described as follows:
H0-Case 3: Suppose the learners with the same pre-test score in both the learning modes have the same mean score of post-test.
H1-Case 3: Suppose the learners with the same pre-test score in both the learning modes do not have the same mean score of post-test.
Table 13shows the results of the Matched-Pairs T-Tests of both the control and treatment groups. The results indicate that the hypothesis H1 is satisfied under the significant level a = 0.05 and P = 0.001 < 0.05. That is, we can logically conclude that the learning mode of curriculum sequencing recommendation is supe-rior to the freely browsing learning mode in terms of learning performance promotion in this case. Actually, our experiment shows that the mean score of the post-test of learners with the same pre-test score who per-formed the learning mode of curriculum sequencing recommendation is obviously 7.17 points higher than the learners who performed the freely browsing learning mode.
In conclusion, the proposed learning mode of curriculum sequencing recommendation indeed surpasses the freely browsing learning mode because it can guide learners to conduct efficient and appropriate learning paths as well as avoid cognitive overload or disorientation during learning processes.
6.5. Questionnaire analysis
To evaluate learners’ satisfaction degree for the proposed genetic-based personalized e-learning system, referring to Chen’s et al. research (Chen, Hsieh, & Hsu, 2007), a questionnaire which involves 26 questions distinguished six various question types was designed to measure whether the proposed genetic-based person-alized e-learning system satisfied the real requirements of most learners. The six question types contain the
Table 14
The descriptions of question types
Question type Number of
questions
Description The services of software and
hardware
8 To investigate whether learners satisfy the provided user interface, course materials, and remedy learning mechanism
Learning interests 2 To investigate whether learners are interested in using the proposed genetic-based personalized e-learning system for mathematical learning
Learning mode 4 To investigate whether learners can accept the proposed learning mode with personalized learning path guidance
Learning interaction between teachers and learners
3 To investigate whether the proposed genetic-based personalized e-learning system affects learning interaction between teachers and learners
Learning attitude 5 To investigate whether learners with computer use the proposed genetic-based personalized e-learning system for mathematical learning at home
Learning performance 4 To investigate whether the proposed genetic-based personalized e-learning system can promote learners’ learning performances and confidence
Table 13
The results of the Matched-Pairs T-Tests of both the control and treatment groups
Paired differences t df Sig. (2-tailed)
Mean Std. deviation Std. error mean 95% Confidence interval of the difference Lower Upper Pair 1: Post-test of freely browsing learning
mode–post-test of the learning mode of curriculum sequencing recommendation
The satisfaction evaluation results of questionnaire Question
type
Question Satisfaction degree (%)
The freely browsing learning mode (there are totally 61 valid questionnaires)
The learning mode of curriculum sequencing recommendation (there are totally 103 valid questionnaires)
Very approved Approved No opinion Disapproved Very disapproved Very approved Approved No opinion Disapproved Very disapproved The services of software and hardware
1. I agree that the proposed system has provided a friendly user interface to support learning mathematics via web-based learning
52 33 13 0 2 61 26 13 0 0
2. I agree that the learning contents with flash animation provided by the proposed system can deepen my impression of learning mathematics
54 28 15 2 2 63 23 13 0 1
3. I agree that the learning
mathematics by the proposed system with interactively learning interface is a very interesting learning mode
51 31 18 0 0 59 25 12 3 1
4. I can completely understand the meaning of all designed course materials in the proposed system
44 31 15 2 8 53 29 15 2 1
5. I can completely understand the meaning of the corresponding test question of the learned courseware in the proposed system
59 23 13 5 0 57 28 11 3 1
6. I agree that using the proposed system for mathematical learning is very interesting because I can operate all system functions as well as control self-learning time
74 18 5 3 0 65 26 7 1 1
7. I agree that the immediate test function provided by proposed the system can help me understand whether I have acquired the learned courseware or not. Meanwhile, I like the learning mode very much
56 26 8 2 8 57 26 9 4 4
8. I agree that the remedy learning mechanism provided by the proposed system for unfamiliar learning concepts is very helpful to my learning 46 31 16 3 3 57 19 15 6 3 Average 82 13 5 85 12 4 C.-M. Chen / Computers & Educati on 51 (2008) 787–814
interests system for mathematical learning is very interesting
2. I agree that the learning contents provided by the proposed system can promote my interests of learning mathematics
61 23 10 5 2 60 21 17 2 0
Average 88 7 6 88 10 2
Learning mode
1. I agree that using the freely browsing learning mode for mathematical learning is a very good learning mode because I can freely click any course materials by myself for learning
64 26 8 0 2 – – – – –
2. I agree that directing learners to enter the post-test process after they learnt one half of course materials by the freely browsing learning mode is a good learning mode
54 26 13 5 2 – – – – –
3. I agree that using the learning mode of curriculum sequencing recommendation is a very good learning mode because I only need to learn unfamiliar courseware based on a planning learning path under skipping acquired courseware
– – – – – 51 27 13 2 7
4. I agree that the learning mode of curriculum sequencing
recommendation provides a reasonable learning path because it simultaneously consider courseware difficulty level and the concept continuity of learning path to plan the learning order of courseware
– – – – – 64 19 13 1 3
Average 85 11 4 81 13 6
(continued on next page)
C.-M. Chen / Computers & Education 51 (2008) 787–814 809
Question type
Question Satisfaction degree (%)
The freely browsing learning mode (there are totally 61 valid questionnaires)
The learning mode of curriculum sequencing recommendation (there are totally 103 valid questionnaires)
Very approved Approved No opinion Disapproved Very disapproved Very approved Approved No opinion Disapproved Very disapproved Learning interaction between teachers and learners
1. I agree that I do not need to rely on teacher instruction if I can use the proposed system for mathematical learning
49 28 13 7 3 49 26 12 4 10
2. I agree that I can learn better if I can get assistance from teacher except learning mathematics by the proposed system
64 26 8 0 2 68 18 8 3 3
3. I feel that using the proposed system for mathematical learning will reduce interactive chance with teacher
41 20 28 8 3 46 22 22 7 3
Average 76 16 8 76 14 10
Learning attitude
1. I would like to reply once again for the test question that I cannot give correct answer when I learnt mathematics by the proposed system
56 23 16 5 0 60 20 11 4 5
2. I feel that time always passes very quickly when I use the proposed system for mathematical learning
48 16 23 7 7 52 28 15 3 2
3. I feel so happy when I think of using proposed system for mathematical learning
56 31 10 2 2 63 23 10 2 2
4. I feel that using the proposed system for mathematical learning is very convenient because I can learn at any time and any place by Internet. Therefore, I have a strong willing to learn mathematics by the proposed system once again
49 28 11 8 3 50 25 15 5 5
Average 77 15 11 76 12 10
5. Do you have any computers that can be used at home?
Yes 87 No 13 – Yes 87 No 13 –
5.1 Can you surf Internet at home? Yes 69 No 31 – Yes 77 No 20 –
C.-M. Chen / Computers & Educati on 51 (2008) 787–814
proposed system for mathematical learning by Internet at home?
Learning performance
1. The remedy learning mechanism provided by the proposed system can help me acquire the courseware that I could not understand in the past, thus promoting my learning confidence
52 21 18 7 2 57 27 12 1 3
2. I feel that learning mathematics by the proposed system is superior to the conventional classroom learning
44 23 26 5 2 51 19 22 6 1
3. I agree that the learning process followed by the pre-test, learning and post-test is a very efficient learning mode
54 26 8 7 5 54 20 17 5 4
4. I agree that the proposed system is an effectively assisted learning tool for mathematical learning
56 31 8 2 3 61 21 12 0 6 Average 77 15 8 78 16 6 C.-M. Chen / Computers & Education 51 (2008) 787–814 811
services of software and hardware, learning interests, learning mode, learning interaction between teachers and learners, learning attitude, and learning performance.Table 14gives a summarization of question types with brief descriptions. There are totally 164 effective questionnaires filled out by learners who participated in this experiment. Among 164 effective questionnaires, 61 learners adopted the freely browsing learning mode and 103 learners used the proposed learning mode of curriculum sequencing recommendation for mathematics learning. The evaluation results of satisfaction degree are detailed inTable 15. To conveniently observe the evaluating results, the investigation results of ‘‘strongly agreed’’ and ‘‘agreed’’ are merged as ‘‘approved’’, and the investigation results of ‘‘strongly disagreed’’ and ‘‘disagreed’’ are merged as ‘‘disapproved’’.
The evaluating results listed inTable 15indicate that over 76% learners are satisfactory in terms of the ser-vices of software and hardware, the promotions of learning interests, learning attitude and learning perfor-mance for both the evaluated learning modes. Most learners agreed that both the evaluated learning modes provide satisfied software and hardware environments and are very helpful to their mathematical learning. Moreover, both the evaluated learning modes can attract learners to learn mathematics using leisure time by Internet at home. Most learners also agreed that both the evaluated learning modes help them conduct effi-cient learning. Particularly, the remedy learning mechanism provided by both the evaluated learning modes can help learners acquire the courseware that they could not understand well in the past, thus promoting their learning effectiveness and confidences. Additionally, 81% learners agreed that the learning mode of curriculum sequencing recommendation provide an appropriate learning path for aiding mathematical learning because they only need to learn unfamiliar courseware, but skipping acquired courseware. Similarly, 80% learners also agreed that using the freely browsing learning mode for mathematical learning is a good learning mode because they can freely select any course materials by themselves for learning. Encouragingly, 77% learners agreed that they do not need to rely on teacher instruction if they can use any one of both the learning modes for mathematical learning, but up to 90% learners agreed that they can learn better if they can also get assis-tance from teacher except learning by any one of both the evaluated learning modes.
The satisfaction evaluation results of questionnaire mentioned-above show that both the evaluated learning modes obtain very high satisfaction degree for mathematical learning. To further compare the difference of satisfaction degree between both the learning modes, this study invited 23 learners who had adopted any one of both the learning modes for mathematical learning to conduct another learning mode, then they were invited to fill out another questionnaire for assessing satisfaction degrees of learners who simultaneously conducted both the learning modes.Table 16presents the evaluation results of questionnaire containing 23 effective samples. The results indicated that 70% learners felt much more prefer the learning mode of curriculum sequencing recommendation than the freely browsing learning mode because it can help them plan Table 16
The questionnaire for assessing satisfaction degrees of learners who conducted both the learning modes
Question Number of learners who
selected the first item
Number of learners who selected the second item 1. The personalized e-learning system provides two learning modes. I
prefer (1) the learning mode of curriculum sequencing recommendation. (2) The freely browsing learning mode
16 (70%) 7 (30%)
2. Based on the aspect of curriculum sequencing, I prefer (1) the learning system help me plan a logical learning path only for the courseware that I have not acquired. (2) To learn freely for all courseware by myself
16 (70%) 7 (30%)
3. In two evaluated learning modes, I feel that (1) only learning the courseware that I have not acquired is enough. (2) Learning all courseware can let me learn much more
7 (30%) 16 (70%)
4. In two evaluated learning modes, I feel that (1) only learning the courseware that I have not acquired can learn more efficiently than learning all courseware. (2) Learning all courseware can let me learn more efficiently than only learning the courseware that I have not acquired
13 (57%) 10 (43%)
5. In two evaluated learning modes, I think that (1) only learning the courseware that I have not acquired can improve my learning performance. (2) Learning all courseware can enhance my learning performance
a logical learning path only for the courseware that they have not acquired. In the meanwhile, 57% learners felt that only learning the courseware that they have not acquired yet can learn more efficiently than learning all courseware. However, there are up to 70% learners to think that learning all courseware can enhance their learning performances. The phenomenon shows that most students of Taiwan’s elementary schools are facing a very competitive learning environment among schoolmates, such that they cannot adequately trust to con-duct more efficiently learning like the proposed learning mode of curriculum sequencing recommendation even they prefer it than the freely browsing learning mode.
7. Conclusion
This study proposes a genetic-based personalized learning path generation scheme for individual learners to support personalized web-based learning. The proposed personalized learning path generation scheme can simultaneously consider courseware difficulty level and the concept continuity of successive courseware according to the incorrect testing responses in a pre-test while implementing personalized curriculum sequenc-ing dursequenc-ing learnsequenc-ing processes. Compared to the freely browssequenc-ing learnsequenc-ing mode used in most web-based learn-ing systems, experimental results indicated that the proposed learnlearn-ing mode of curriculum sequenclearn-ing recommendation can precisely plan a personalized learning path for the courseware that a learner has not acquired yet based on a difficulty parameter and concept continuity of successive courseware, and moreover can promote learner’s learning effectiveness during learning processes. In the meanwhile, the investigation results of questionnaire revealed that most learners agreed the learning mode of curriculum sequencing recom-mendation is superior to the freely browsing learning mode in terms of learning efficiency. An important advantage is that the learning mode of curriculum sequencing recommendation customizes learning for those who have very specific needs and not much time or patience to complete topics they have learned.
Acknowledgment
The authors would like to thank the National Science Council of the Republic of China, Taiwan, for finan-cially supporting this research under Contract No. NSC95-2520-S-004-001.
References
An Extension Chinese Lexicon Scanner (2006). Available fromhttp://dlll.nccu.edu.tw/rank/ecscanner/. AppServ Open Project (2007). Available fromhttp://www.appservnetwork.com/.
Baker Frank, B. (1992). Item response theory: Parameter estimation techniques. New York: Marcel Dekker.
Berghel, H. (1997). Cyberspace 2000: Dealing with information overload. Communications of the ACM, 40(2), 19–24.
Borchers, A., Herlocker, J., Konstanand, J., & Riedl, J. (1998). Ganging up on information overload. Computer, 31(4), 106–108. Brusilovsky, P., Eklund, J., & Schwarz, E. (1998). Web-based education for all: A tool for development adaptive courseware. Computer
Networks and ISDN Systems, 30, 291–300.
Brusilovsky, P., & Vassileva, J. (2003). Course sequencing techniques for large-scale web-based education. International Journal of Continuing Engineering Education and Lifelong Learning, 13(1–2), 75–94.
Chen, C.-M., Hsieh, Y.-L., & Hsu, S.-H. (2007). Mining learner profile based on association rule for web-based learning diagnosis. Expert Systems with Applications, 33(1), 6–22.
Chen, C.-M., Lee, H.-M., & Chen, Y.-H. (2005). Personalized e-learning system using item response theory. Computers and Education, 44(3), 237–255.
Chen, C.-M., Liu, C.-Y., & Chang, M.-H. (2006). Personalized curriculum sequencing using modified item response theory for web-based instruction. Expert Systems with Applications, 30(2), 378–396.
Frakes, W. B., & Baeza-Yates, R. (1992). Information retrieval data structures and algorithms. Englewood Cliffs: N. Prentice Hall. Genesereth, M. (1997). An agent-based framework for interoperability. In J. M. Bradshaw (Ed.), Software agents (pp. 317–345). Menlo
Park, Calif.: AAAI Press.
Horton, W. (2000). Designing web-based training: How to teach anyone anything anywhere anytime? John Wiley & Sons, Inc.
Hsu, T. C., & Sadock, S. F. (1985). Computer-assisted test construction: A state of art. TME report 88, Princeton, New Jersey, Eric on Test, Measurement, and Evaluation, Educational Testing Service.
Jih, H. J. (1996). The impact of learners’ pathway on learning performance in multimedia computer aided learning. Journal of Network and Computer Applications, 19, 367–380.
Lee, M.-G. (2001). Profiling students’ adaptation styles in web-based learning. Computers and Education, 36, 121–132.
