Web-Based Intelligent Learning: Using Machine

learning
Sara Vamshi Teja
Research Scholar
University of East London, United Kingdom
[email protected]

Abstract.
This work proposes an intelligent learning diagnosis system that supports a Web-
based thematic learning model, which aims to cultivate learners’ ability of knowledge
integration by giving the learners the opportunities to select the learning topics that
they are interested, and gain knowledge on the specific topics by surfing on the
Internet to search related learning courseware and discussing what they have learned
with their colleagues. Based on the log files that record the learners’ past online
learning behaviour, an intelligent diagnosis system is used to give appropriate
learning guidance to assist the learners in improving their study behaviours and
grade online class participation for the instructor. The achievement of the learners’
final reports can also be predicted by the diagnosis system accurately. Our
experimental results reveal that the proposed learning diagnosis system can
efficiently help learners to expand their knowledge while surfing in cyberspace Web-
based “theme-based learning” model.
Keywords: Web-based learning, Theme-based learning, Fuzzy expert system, K-
nearest neighbour, Naïve Bayesian classifier, Support vector machines, Learning
diagnosis

1 Introduction
The surprising development of information technology has created a new vision for
network learning that its influence has already spread over the world to facilitate
educational innovation. Therefore, many countries have been paying attention to
computer technology and expect it can facilitate the education reform in an effective
and efficient ways. It is well known that the application of computer and Internet
teachings to traditional teaching requires some kind of transformation. Consequently,
the research and development of proper learning model has to seriously consider the
mutual interaction between the users and the computers, the instructor and the
learners, and the interaction among the learners. Embed the related research issues to
the above process, the splendid research results then can be expected.
The theme-based learning is to learn an integrated knowledge by defining a central
“theme” at the very start and compose related knowledge surrounds the central theme
from various aspects. Such a learning model emphasizes the training of the learners
with the competency of knowledge integration. Compared with traditional teaching,
which teaches fragmentary information within the limitation of subjects, units,
chapters, and sections, the intention of theme-based learning is to take a theme as a
starting point and stretch out of it based on the learners’ interests. Accordingly, the
learners can voluntarily construct their own knowledge since the theme is strongly
connected with our daily life and developed from learners’ willingness.
A theme-based learning process can be divided into exterior circulation and interior
circulation as illustrated in Figure 1 [7]. Exterior circulation activities are 1) Identify
a central theme, 2) Identify related subject domains based on learner’s interest, 3)
Collect information for the specific topics, 4) Integrate collected information to build
shared knowledge, and 5) Exhibit learning outcomes and share with others. The
activities of the exterior circulation are explicit learning behaviours. On the other
hand, the interior circulation consists of implicit mental activities, which are Plan,
Action, and Introspection, respectively. When learners engage in the theme based
learning processes on Web, they are experiencing the activities of exterior and interior
circulation synchronously. Since the explicit feature of the learning processes can be
controlled or guided effectively by the careful design and implementation of the Web-
based learning environment, it is expected that the interior circulation, which
represents the invisible mental behaviour of the learners, can make great progress
simultaneously.
The exterior circulation of the theme-based learning model, as Figure 1 illustrates,
can be implemented as a Web based system that helps to manage the learning
processes. The learning activities with regard to student learning can be divided into
five stages as follows:

(1) Identify a central theme

The learners engaged into the theme-based learning can propose their own
interested topics to ask for feedback from other team members. Meanwhile, every
learner can also join other member’s proposed topic. After interaction and brain
storming, the ones who are interested in the same topic are formed as a learning team,
and this topic is the central theme that this team would investigate. The motivation of
such an arrangement is that “a student can learn better if he/she was interested in the
learning topic”. The theme should be closely connected with the learners’ daily life
and an extensive range of survey which is not limited in a specific field is encouraged.

(2) Identify related subject domains based on learner’s interest

At this stage, the theme is defined and the learning team for each theme is formed.
Based on the learner’s own specific interest, each team member tries to find the issues
in the related subject domain derived from the theme. Notably, the interaction of
learners on the learning platform can influence the relatively inactive learners to
trigger their interests effectively on some specific topics through the events and
activities originated by their team members.

(3) Collect information for the specific topics

 Team members will cooperate with each other to collect data and information
related to the interested topic at this stage. With the help pf search engine and other
assistant tools, wealthy knowledge related to the interested topic can be built up. If the
data collected from the cyberspace is not enough, real-world resources such as
libraries, face-to-face interviews et al. also can be utilized. The collected data or
information is then processed to form the knowledge stored in the learner’s long-term
memory.

(4) Integrate collected information to build shared knowledge

Each team member tries to organize the data or information collected at the
previous stage and to generate a thematic report. The reports can be shared with other
teammates through peer review and online discussion.

(5) Exhibit learning outcomes and share with others

The thematic report for the each learner is expected to be refined to certain degree
through consecutive discussion with the teammates and constructive suggestion
offered by the teacher. The elaborated report is finally displayed on the public area to
make it accessible to the teachers and all the learners.
Besides putting the learning activities that correspond to the exterior circulation of
the theme-based learning model into practice, an intelligent diagnosis system is also
incorporated in the proposed Web-based thematic learning platform. Notably, a fuzzy
expert system and a composite classifier are used to give the learning guidance to the
learners, and assist the instructor in grading each learner’s online class participation
and predicting the performance of each learner’s final written report. The remainder
of the paper is organized as follows. Section 2 gives a brief description of architecture
of the proposed Web-based thematic learning platform. In section 3, I will show the
details of the intelligent diagnosis system. Section 4 reviews and discusses the
experimental results. Conclusions and the future work are made in Section 5.

1. Identify a
Central
Theme.

5. Exhibit 2. Identify related

Learning Subject domains
And share with
outcome Plan Based on learner’s
Others Interest.

Introspection Action

4. Integrate collected 3. Collect information

Information to build For the specific-
Shared knowledge. Topics.
 Figure 1. A theme-based learning model.

2 Architecture of the Theme-Based Learning Website

A theme-based learning system is composed of three functional modules as shown in
Figure 2. They are User Interface Agent, Learner Profile Management Agent, and
Learning Diagnosis System, respectively.

2.1 User Interface Agent

The learners can login the theme-based learning system through User Interface
Agent to participate in the learning activities such as data searching, data managing,
discussing with the colleagues and the teacher online, posting and replying the
articles, etc.

Figure 2. Architecture of the theme-based learning website.

2.2 Learning Profile Management Agent

The connections and the interactions for the learners with the system and other
learners are built up in this module. The system can generates the learners’ learning
profiles, including the total time that the learners stayed in the platform, the frequency
of login sessions, the learning materials collected by the learners, the articles posted
or replied by the learners, and the online group discussion time spent by the learners,
etc., which provide the teachers and Learner Diagnosis System to follow the learners’
learning status such that effective scaffolding and constructive suggestion or analysis
for the learners can be given timely.
2.1 Learning Diagnosis System
The learners are expected to make progress based on certain proper learning
advices given by the Learning Diagnosis System. The system also generates the
online participation assessment at the end of each learning activity stage as mentioned
in Section 1, according to the learner profile logged in the system. The teacher can
either take this class participation assessment as a proportion of the learners’ final
grade, or use this assessment to uncover the learners that fall behind at the end of each
learning activity stage. Meanwhile, the system can also predict the performance of the
learners’ final report so that the teacher can use this predicted achievement for further
analysis on the learner’s study behaviour when there is a gap between expected result
and the learner’s actual performance.

3 Learning Diagnosis System

The Learning Diagnosis System employed in this work consists of two major parts.
One is a fuzzy expert system which not only gives appropriate diagnosis messages to
the learners but also delivers each learner’s online participation assessment to the
teacher at the end of each learning activity stage based on the learners’ profile. The
other part is a so called composite classifier which is used to predict the learners’
accomplishment on the final report. The motivation of using a fuzzy expert system to
give diagnosis and class participation assessment whereas using a composite classifier
to predict the learners’ final accomplishment is that the fuzzy expert system can
function more like human experts who explain the reasoning processes behind their
recommendation. On the other hand, it is not too difficult to find some advanced
machine learning techniques combined with wrapper attribute selection method
possess a better prediction capability than a fuzzy expert system.

3.1 Architecture of the fuzzy expert system

An expert system is a program that behaves like an expert in some problem
domain. The principle use of expert systems is to seek information from a variety of
sources including databases and the users to solve finite, well-defined problems [3].
To deal with uncertain and incomplete information, the fuzzy expert system
incorporates fuzzy logic into the reasoning process and knowledge representation
scheme [1].
The input to a fuzzy expert system is a crisp value that is given by the learner
profile database. The fuzzy rule base is composed of a set of fuzzy if/then rules and
the intersection or minimum operation is employed to generate a corresponding fuzzy
subset for each fuzzy rule. The aggregator then combines all of the fuzzy subsets
assigned to the output variable together to form a single fuzzy subset for the output.
Lastly the aggregated linguistic values from the inferred fuzzy control action are fed
into the defuzzifier to generate a non-fuzzy control output. Notably, the generalized
bell-shaped membership function is chosen for three antecedents and the consequent
in the fuzzy expert system. The three antecedents are the number of articles
posted/replied by the learners, the number of the learning materials that each learner
collected, and the frequencies of visiting the platform by the learners, respectively.
Meanwhile, the Mamdani defuzzification method is employed to compute the
centroid of membership function for the aggregated output, where the area under the
graph of membership function for the aggregated output is divided into two equal
subareas.
The inferential rules of the fuzzy expert system are treated differently when they
are used to generate some appropriate suggestion or diagnosis messages for the
teacher and the learners. The system may give a feedback message to the learners
when the membership grades of linguistic variables such as “low” or “high” is the
largest among the three for each input. The learning diagnosis system will not only
give the teacher a summary report of suggestion messages that the learners received,
but also offer the teacher each learner’s online participation assessment based on the
non-fuzzy output of the defuzzifier. The teacher can locate the learners that fall
behind and give them individual guidance by examining the class performance record
given by the system. Notably, although the exact calculation of class participation
assessment involves the 27 inferential rules, only 11 inferential rules are used in
actual computation since the rest are considered unreasonable.

3.2 Architecture of the composite classifier

The motive of using the composite classifier in our work is that the composite
classifier has the advantage of making decisions more reliable and accurate than a
single classifier although the combined model is typically hard to analyse in intuitive
terms what factors are contributing to the improved decisions [3]. The composite
classifier in this work is mainly composed of three independent classifiers, i.e., a K
nearest neighbour classifier, a naïve Bayesian classifier, and a support vector
machines classifier, respectively. Each of three independent classifiers uses wrapper
approach to select the desirable input parameters during training. The training data is
taken from the learners’ profiles database maintained for the learning groups in past
learning programs. A vote is taken if there are conflicts among the prediction results
of the three classifiers. The output of the composite classifier is the predicted grade
for the learners’ final report. The teacher can either cite this predicted grade as portion
of the learners’ final achievements or perform a further investigation if there is a
discrepancy between the learners’ actual achievement and study behaviour.
 Wrapper attribute selection method
It is well known that the performance of most machine learning algorithms can be
deteriorated by some irrelevant or unhelpful attributes. Thus it is common to precede
classification work with an attribute selection stage which strives to eliminate all but
the most relevant attributes. This is also one of the major reasons that the prediction
capability of a fuzzy expert system is worse than some advanced machine learning
algorithms since the inputs to the fuzzy expert system are always chosen by the
human experts and these selected attributes might not be the most promising ones for
the fuzzy expert systems.
The attribute selection methods can be divided into two broad categories in the
literature [4], i.e. filter methods and wrapper methods. Filter methods select
predictive subset of the attributes using heuristics based on characteristics of the data,
whereas wrapper methods make use of the classifier actually used to evaluate the
accuracy of attribute subsets. Wrapper methods generally result in better performance
than filter methods because the latter suffers from the potential drawback that the
attribute selection principle and the classification step do not necessarily optimize the
same objective function. .
In the wrapper approach, the learner is applied to subsets of attributes and tested on
a hold-out set. From the results of these tests, a good subset of attributes is selected.
For example, for forward selection, a classifier is built for each attribute individually;
and the most accurate attribute is “accepted” into the subset of good attributes. That
attribute is removed, and the process is repeated, adding each of the remaining
attributes and evaluating its performance. The “best” set of two attributes is thus
created. This proceeds incrementally until an attribute set with maximal accuracy is
achieved. Similarly, backward selection proceeds by eliminating one attribute at a
time, finding the least beneficial attribute and eliminating it, and repeating the
process, eliminating the least accurate attributes until eliminating further attributes
decreases accuracy.

 K nearest neighbour classifier

To classify an unknown data sample X, the k nearest neighbour classifier simply
examines the k closest training samples to X and assigns it to the most common class
among these k closest samples. In other words, we are seeking those training samples
that are most similar to X, and then classify X into the most heavily represented class
among these most similar objects [2]. Notably, “closeness” is defined in terms of
Euclidean distance.
A k nearest neighbour classifier has several attractive properties. For example, it is
easy to program and no optimization or training is required; Extension to multiple
classes is straightforward. Although a potential drawback of the k nearest neighbour
classifier is that it does not build a model, relying instead on retaining all of the
training data set points. Thus, searching through a large training data set to find the k
nearest can be a time-consuming process. However, this problem can be evaded here
since small data sets in low dimension are used in this work.

 Naïve Bayesian classifier

The naïve Bayesian classifier predicts an unknown data sample, X, belonging to the
class with highest posterior probability, conditioned on X [6]. Bayesian classifiers
have minimum error rate in comparison to all other classifiers theoretically. Even
though this is not the case in practice due to inaccuracies in the simplified
assumptions made for its use, such as class conditional independence and the lack of
available probability data, empirical studies given in the literature shows the
performance of naïve Bayesian classifier is still comparable to other complex
machine learning techniques such as neural networks. I thus adopt this simple but
effective approach as an element of the composite classifier.
 Support vector machines
Support vector machines (SVM) have recently gaining popularity due to its numerous
attractive features and eminent empirical performance [4]. The main difference between the
SVM and conventional regression techniques is that it adopts the structural risk minimization
(SRM) approach, as opposed to the empirical risk minimization (ERM) approach commonly
used in statistical learning. The SRM tries to minimize an upper bound on the generalization
rather than minimize the training error, and is expected to perform better than the traditional
ERM approach. Moreover, the SVM is a convex optimization, which ensures that the local
minimization is the unique minimization.
To solve a nonlinear regression or functional approximation problem, the SVM nonlinearly
map the input space into a high-dimensional feature space via a suitable kernel representation,
such as polynomials and radial basis functions with Gaussian kernels. This approach is
expected to construct a linear regression hyperplane in the feature space, which is nonlinear in
the original input space. Then the parameters can be found by solving a quadratic
programming problem with linear equality and inequality constraints.
As the SVM outperforms other conventional regression methods in the application of time
series and internet traffic predictions in the literature [4], I thus try to replace the fuzzy logic
module in the bandwidth reservation scheme with the SVM as illustrated in the previous
section to estimate the reserved bandwidth in the neighbouring cells.

4 Experimental Results and Analyses

To examine the educational effect of the theme-based learning system, two fifth
grade classes at an elementary school have been chosen to practice theme-based
learning in classroom teaching. One of the two classes was experimented with the
proposed Web-based thematic learning platform in a Natural Science course, wherein
a fuzzy expert system is incorporated into the intelligent diagnosis system to grade
students’ class participation and learning guidance to the learners based on learning
profiles so that the students can receive just-in-time support or suggestion to help
them gain better learning achievement. The second experiment was conducted in
another Natural Science course for the other fifth grade class whereas the diagnosis
system was removed from the Web-based thematic learning platform in order to
demonstrate the performance of the proposed diagnosis system.
Table 1 compares the pupils’ achievement in two classes. The statistical results
were obtained by running t test with the SPSS software package. The average score
received by the 25 pupils whose study behaviour was rectified by the diagnosis
system is apparently better than the average score of 18 pupils in the other
experiment, although the gap between these two mean scores are not quite significant
due to the different grading standard for the two teachers. However, I observe that the
proportion of the pupils that failed in the course substantially lowered down with the
aid of the diagnosis system. Meanwhile, the diagnosis system greatly reduces the
teaching load of the teacher so that the teacher can have more free time to give
individual guidance to the specific pupils that fall behind or behave inactively.
As mentioned in Section 3, I use three different classifiers to predict the pupils’
achievement in the final report based on the pupils’ profiles, and the majority of their
votes are taken as the final prediction result. To verify the performance of each
individual classifier, I ran a series of tests on each of three classifiers by using a so-
called leaving-one-out cross validation technique (LOOCV) [5] with 18 and 25
samples for the two classes, respectively. The inputs to each classifier are identical to
those used in the fuzzy expert system. Notably, the LOOCV method removes a single
sample in each trial, trains on the rest, and then tests the PNN classifier on the
removed single sample.
The performance comparison given in Table 2 reveals that all of the three
classifiers can achieve a much higher prediction rate for the class using the theme
based learning platform in which the diagnosis system is embedded. It can be inferred
from the results that the learning guidance given by the fuzzy expert system
significantly affected the learners’ study behaviour and boosted the quality of the
learners’ final reports further.
Table 1 Comparison of the pupils’ achievement in two experimental classes

Learning Learning
Platform Platform
t test with without
Diagnosis Diagnosis
System System
Mean 86.60 82.22
Standard
9.443 13.225
Deviation
Standard Error
1.889 3.117
Mean
Degree of
41
Freedom
Significance
0.212
(2tailed)

Table 2 The LOOCV prediction results of three classifiers.

Learning Learning
Platform Platform
Classifier with without
Diagnosis Diagnosis
System System
Naïve
92% 67%
Bayesian
K Nearest
88% 50%
Neighbour
Support 84% 67%
 Vector
Machines

5. Conclusions and Future Work

The Web-based thematic learning system implemented in this work mainly practices the
exterior circulation of the Theme-Based Learning model as illustrated in Figure 1. An
intelligent diagnosis system, which is composed of a fuzzy expert system and a composite
classifier, is proposed to support the Web-based thematic learning platform. Experimental
results show that the fuzzy expert system is very effective in offering the learning guidance to
the pupils. Besides, a series of leaving-one-out cross validation tests demonstrate the high
prediction accuracy of three components of the composite classifier on the final report quality
based on the pupils’ learning profiles. In case an inconsistency between the learners’
predicted result and their actual performance takes place, the teacher can also proceed with a
further analysis on the cause of the discrepancy.
In the future work, I plan to build a platform for practicing theme-based learning between
cooperative schools so that cross-classes or cross-schools theme based learning can be
spanned. The system will also have new developed mobile learning tools to support outdoor
teaching by allowing the learner to access the archived learning resources. Meanwhile, the
user interface will be redesigned to be more user friendly, and intelligent detecting
mechanism and text mining techniques will be incorporated into the leaner profile
management agent to locate the distractive learners and questionable plagiaries, respectively.

References

[1] Buckley, J., and Eslami., An introduction to fuzzy logic and fuzzy sets
(advances in soft computing).
Physica Verlag, 2002.
[2] Denoeux, T. “A k-nearest neighbor classification rule based on Dempster-
Shafer theory,” IEEE Transactions on Systems, Man and Cybernetics, vol. 25,
no. 5, pp. 804-813.
[3] El-Shal, S.M., and Morris, A.S., “A fuzzy expert system for fault detection in
statistical process control of industrial processes,” IEEE Transactions on
Systems, Man and Cybernetics, Part C, vol. 30, no. 2 , pp. 281 -289, 2000.
[3] Applied data mining. John Wiley & Sons Ltd., 2003.
[4] Hall, M. and Holmes, G. “Benchmarking attribute selection techniques for
discrete class data mining,” IEEE Transactions on Knowledge and Data
Engineering, vol. 15, no. 3, pp. 1437 – 1447, 2003.
[5] Hand, D., Mannila, H., and Smyth, P., Principles of data mining, MIT press,
2001.
[6] Liu, J.N.K., Li, B.N.L., and Dillon, T.S., “An improved naive Bayesian
classifier technique coupled with a novel input solution method [rainfall
 prediction],” IEEE Transactions on Systems, Man and Cybernetics, Part C,
vol. 31, no. 2, pp. 249-256, 2001.
[7] Liu, M. C., Chen, L. C., and Tung, S. C., “The implementation and study of
the Web-based thematic learning model,” Proceeding of the 6th Global
Chinese Conference on Computers in Education, Beijing, 2002.