Symbols Definition:
Q: is the query vector whose dimension is the same as the feature vector of content node (CN) D: is the number of the stage in an LCCG.
S0~SD-1: denotes the stage of an LCCG from the top stage to the lowest stage.
ResultSet, DataSet, and NearSimilaritySet: denote the sets of LCC-Nodes.
Input: The query vector Q, search threshold T and the destination stage SDES where S0≤SDES≤SD-1.
Output: the ResultSet contains the set of similar clusters stored in LCC-Nodes.
Step 1: Initiate the DataSet=φ and NearSimilaritySet =φ.
Step 2: For each stage Si∈LCCG, repeatedly execute the following steps until Si≧SDES
2.1: DataSet = DataSet ∪ LCC-Nodes in stage Si, and ResultSet=φ. 2.2: For each Nj ∈ DataSet,
{ If Nj is near similar with Q
Then insert Nj into NearSimilaritySet.
Else If (the similarity between Nj and Q) ≥ T Then insert Nj into ResultSet. }
2.3: DataSet = ResultSet. //for searching more precise LCC-Nodes in next stage in LCCG
Step 3: Output the ResultSet = ResultSet ∪ NearSimilaritySet.
Chapter 7 Knowledge Controller (KC)
When the learners login to start learning, the Knowledge Controller Module of ILCMS is responsible for initiating a learning activity from the LAR and delivering the suitable learning contents and activities to learners according to their learning results.
Therefore, a Learning Activity Controller (LAC) module implemented in KC module, including System Coordinator (SC) and Inference Engine (IE) [78], will retrieve the appropriate learning objects in LOR, testing sheets in Testing Item Bank (TIB), or application programs (AP) in APR according the personalized learning activity in LAR.
Then, it delivers them to learners for adaptive learning with teaching strategy.
7.1 The Rule Inference Process in KC Module
Due to the pedagogical needs, the teachers can use OOLA model to edit the learning activity as directed graph that includes learning objects, applications and assessment quizzes. The OOLA is a rule-based model and the rule representation of OOLA model is based on the New Object oriented Rule Model (NORM) architecture [78] [123] [124] [125] [139] which modularize the rules as rule classes [11] [92]. While the rule set of specific domain is acquired, the rule classes are instantiated as rule objects and can be executed on the NORM inference engine, DRAMA [78].
Figure 7.1 illustrates a leaning process and associated rule inference process. While an OOLA based leaning activity starts, System Coordinator (CO) in LAC will load a suitable OOLA model and then Inference Engine (IE) will infer a suitable learning node as next node for learning according to the rule definitions within OOLA and learners’
learning results after learners finished each node or time of node ran out. The SC will give learners a learning content or test sheet, or run an application program as learning
service according to different node types which is selected by IE.
Figure 7.1: The Diagram of Rule Inference Process in KC Module
7.2 The Learning Process of OOLA based Learning Activity
As stated previously, in KA, System Coordinator (SC) is responsible for communicating with IE and controlling all related system modules to execute their jobs, such as displaying learning content or running application program. Accordingly, in this dissertation, we propose a learning process algorithm of running OOLA, as shown in Figure 7.2.
Figure 7.2: The Learning Process of Running OOLA Model
Chapter 8 Knowledge Miner (KMin)
Knowledge Miner Module includes a Learning Portfolio Analyzer (LPA), which consists of Learning Portfolio Mining (LPM) [118] and Two-Phase Concept Map Construction (TP-CMC) [110] algorithm. According to the learners’ characteristics, the former applies the clustering and decision tree approach to analyze the learning behaviors of learners with high learning performance. The latter applies Fuzzy Set Theory and Data Mining approach to automatically construct the concept map by learners’ historical testing records. Therefore, after the learners finished the learning activities, teachers can use LPA to analyze the learning portfolios of learners for refining their teaching strategies and contents.
8.1 Learning Portfolio Analysis Using Data Mining Approach
Several articles [9] [36] [65] [112] [140] have proposed that a new learner will get the similar learning performance if providing the learning guidance extracted from previous similar learners. The concept is the same as the adage of Chinese, “Good companions have good influence while bad ones have bad influence.” Therefore, we conclude that a new learner could get the high learning performance if s/he follows the effective learning experience of similar learners. However, this conclusion results in the following three issues should be solved: (1) how to acquire the learning characteristics of learners, (2) how to group learners into several groups according to her/his individual learning characteristics, and (3) how to assign a new learner to a suitable group for offering her/him personalized learning materials.
8.1.1 The Process of Learning Portfolio
During learning activity, learning behaviors of learners can be recorded in the database, called learning portfolio, including the learning path, preferred learning course, grade of course, and learning time, etc., in the e-learning environment. Articles [4] [15] [29] [95] [104] [108] have proved that the information of learning portfolio can help teacher analyze the learning behaviors of learners and discover the learning rules for understanding the reason why a learner got high or low grade.
Therefore, based upon the learning portfolio with the predefined data format, we can apply sequential pattern mining approach to extract frequent learning patterns of learners. Then, according to these mined learning patterns, these learners can be grouped into several groups with the similar learning behaviors using clustering approach. By using the questionnaires including the Learning Style Indicator [77], Group Embedded Figures Test (GEFT) [137], etc. to acquire the learning characteristics of learners, we can acquire the learning characteristics of learners as learner profile which can be used to create a decision tree to predict which group a new learner belongs to.
Thus, in this dissertation, we propose a four phase Learning Portfolio Mining (LPM) Approach using sequential pattern mining, clustering approach, and decision tree creation sequentially. Then, in the last Phase, we also propose an algorithm to create personalized activity tree which can be used in SCORM compliant learning environment.
The Framework of Learning Portfolio Mining (LPM):
As mentioned above, we propose a Learning Portfolio Mining (LPM) approach to extract learning features from learning portfolio and then adaptively construct personalized activity tree with associated sequencing rules for learners.
Figure 8.1: The Flowchart of LPM
As shown in Figure 8.1, LPM includes four phases described as follows:
1. User Model Definition Phase: we define firstly the learner profile including gender, learning style, and learning experience, etc. based upon existing articles and pedagogical theory, and the definitions of what we are going to discover in database.
2. Learning Pattern Extraction Phase: we apply sequential pattern mining technique to extract the maximal frequent learning patterns from the learning sequence within learning portfolio. Thus, original learning sequence of a learner
can be mapped into a bit vector where the value of each bit is set as 1 if the corresponding learning pattern is contained, and distance based clustering approach can be used to group learners with good learning performance into several clusters.
3. Decision Tree Construction Phase: after extraction phase, every created cluster will be tagged with a cluster labels. Thus, two third of the learner profiles with corresponding cluster label are used as training data to create a decision tree, and the remainings are the testing data which can be used to evaluate the created decision tree.
4. Activity Tree Generation Phase: finally, each created cluster including several learning patterns as sequencing rules can be used to generate personalized activity tree with associated sequencing rules of Sequencing and Navigation (SN).
8.1.2 The Clustering Process of Learner
In this section, we will describe the User Model Definition Phase and Learning Pattern Extraction Phase in LPM.
User Model Definition Phase:
Before extracting the learning features, we have to define a user model as learner profile, which will be recorded in database, to represent every learner. The definition is described as follows:
Learner L= (ID, LC, LS), where
z ID: denotes the unique identification of a learner.
z LC = <c1c2…cm>: denotes the sequence of learning characteristics of a learner.
z LS = <s1s2…sn>: denotes the learning sequence of a learner during learning activity, where si is an item of learning content.
In this dissertation, how to efficiently apply the existing pedagogical theories and
how to further propose an efficient approach to solve personalized learning problem are our main concerns. Therefore, we only survey several related articles [9] [25] [29] [43]
[66] [65] [83] [93] [104] [108] [135] [137], which investigated about 1) Learner Model, 2) Learning Style and Motivation, 3) course module category, 4) Learning Style, 5) Cognitive Styles, 6) Gender Difference, and 7) Student Characteristics, and then define the frequent learning characteristics for representing a learner by integrate their proposed leaning characteristics. The defined user model can also be extended if necessary. As shown in Table 8.1, the values of Gender, Age, Education Status, Computer Experience, and Media Preference can be inputted by learners directly and the values of Learning Motivation, Cognitive Style, Learning Style, and Social Status can be acquired by questionnaire, where we use the Learning Style Indicator [77] and Group Embedded Figures Test (GEFT) [137] to acquire the Kolb's Learning Style [66]
and the information about field dependence/independence in Cognitive style, respectively. Here, the numeric value of Age can be transformed into symbolic with {L, M, H}. The transformation principle is described as follows:
In all learners, and l μare the minimal and maximal values of age, respectively.
Let Δ=( -l μ)/3, and then a numeric value of age can be mapped into symbolic value with L in [ , + ), M in [ +l l Δ l Δ , +2l Δ ), and H in [l+2Δ , +3 ]. l Δ
For example, LC = <F, M, S Y, H, FD, D, T, H> denotes that a learner is a Female, Age is Medium among all learners, Education Status is Senior, and etc. Nevertheless, the learning characteristics in user model can be modified for the real needs. In addition, LS denotes a learning sequence of a learner. For example, in Figure 2.2a, LS = <A, AA, AAA, AAB, AB> denotes that a learner studies the learning content A first and then studies the learning content AB. Therefore, based upon the user model, the learner can
be represented as L=(35, <F, M, S Y, H, FD, D, T, H>, < A, AA, AAA, AAB, AB>).
Table 8.1: The Learning Characteristics of Learners
Attribute Value
Gender F: Female, M: Men
Age L: [ , +l l Δ), M: [l+Δ,l+2Δ), H: [l+2Δ, +3 ] l Δ
Education Status E: Elementary, J: Junior, S: Senior, U: Undergraduate, G: Graduate Computer Experience Y: Yes, N: No
Learning Motivation L: Low, M: Medium, H: High
Cognitive Style FD: Field Dependence, FI: Field Independence
Learning Style D: Doer (Concrete Experience & Active Experimentation) W: Watcher (Reflective Observation & Concrete Experience) T: Thinker (Abstract Conceptualization & Reflective Observation) F: Feeler (Active Experience & Abstract Conceptualization) Media Preference A: Audio, V: Video, T: Text, P: Picture, M: Picture & Text
Social Status L: Low, M: Medium, H: High
Learning Pattern Extraction Phase:
After defining the user model, we can apply sequential pattern mining technique to extract the maximal frequent learning patterns from the learning sequence within learning portfolio. Because we want to provide the new learner with effective learning guidance, we collect the learning sequences of learners with high learning performance, e.g., testing grade, from database, as shown in Table 8.2. For extracting the frequent learning pattern, the Learning Pattern Extraction Phase includes three processes shown in Figure 8.2: (1) Sequential Pattern Mining Process, (2) Feature Transforming Process, and (3) Learner Clustering Process.
Table 8.2: The Learning Sequences of 10 Learners
ID Learning Sequence (LS) 1 <B, C, A, D, E, F, G, H, I, J>
2 <A, B, H, D, E, F, C, G, I, J>
3 <A, D, F, G, H, B, C, I, J>
4 <A, B, D, E, C, F, G, H>
5 <A, C, J, F, B, H, D, E, G>
6 <B, H, F, D, E, A, G, C, I>
7 <A, J, E, H, B, C, I, D, G>
8 <B, C, G, E, A, H, D, J, F>
9 <C, E, G, F, J, B, H, A, D>
10 <B, C, A, J, D, E, G, H, F>
Figure 8.2: Learning Pattern Extraction Phase
Sequential Pattern Mining Process:
In this dissertation, we modify a sequential pattern mining approach, called GSP algorithm [4] [97], to extract the frequent learning patterns from learning portfolio because we use the maximal frequent learning pattern to represent the learning features of learners, shown in Figure 8.3.
Algorithm 8.1: Modified GSP Algorithm