I. Introduction
For lectures at education institutions and training at businesses,
educational methods using e-learning have become widespread, and
self-completing learning systems are sometimes seen. Many of the systems
now in service are merely for browsing learning texts or for online
testing and do not exhibit the advantages of e-learning. With the spread
of SCORM [1], which means a unified standard for learning systems,
learning history databases, and learning texts, the number of learning
texts will increase and make it difficult for learners to select
appropriate ones.
Considering these issues, we developed in this research a
bidirectional recommendation system optimum for learning by seeking the
relationship among learning texts from the historical log information of
learning texts used by the learner, or the learning information
indicating how the learner learned. We also developed An Individual
Reviewing System (AIRS) [2]. This paper gives an outline of AIRS and
describes the implementation and evaluation of the bidirectional
recommendation system.
We surveyed the state-of-the-art about the recommendation
technologies such as collaborative filtering, as follows. In [3], the
design and implementation of a recommender system using social networks
was described. In [4], a web content recommendation system based on the
similarities is proposed. In [5], collaborative filtering based on
C-SVM(Support Vector Machine) was proposed examined. In [6], data mining
technologies, such as clustering and sequential pattern mining, for
online collaborative learning data are studied. In [7], monitoring
online tests, such as learner behavior and test quality, through data
visualization are discussed. In [8], an automated learning and skills
training system for a database programming environment is presented. On
the other hand, we developed the new bidirectional recommendation system
[9, 10]. In addition, we are currently studying a method for a
collaborative learning recommendation system that mines the data of
similar users sharing non-favorite subjects using historical logs and
user attribute data [11], too.
2. Outline of AIRS
AIRS is a web-based system that distributes review-only learning
texts for efficient reviewing by learners. Here, reviewing is defined as
"rechecking contents once learned from a lecture, etc., and
strengthening the understanding" [12]. Figure 1 shows a schematic
of AIRS.
Figure 2 displays the AIRS Japanese top page after a learner, who
is going to review the database contents especially selection function,
logs in AIRS. This page is comprised of the book marks (located at the
upper side) of the contents available with AIRS, the contents menu
(located at the left side) corresponding to the selected book mark, and
the help messages for beginners (located at the right side). In figure
2, the contents menu displays database, data model, RDB, design
methodology, and SQL. The learner selects the book mark such as
selection and sorting before the learner can select the corresponding
database contents menu. Then, the learner can proceed to revie the
database contents.
[FIGURE 1 OMITTED]
[FIGURE 2 OMITTED]
AIRS uses Java Servlets and JavaServer Pages (JSP). Tomcat5.5 is
used for the Servlet container and Apache2 for the web server. MySQL is
used for the learning history database of MySQL learners and for the
recommendation database as shown in Figure 3.
This AIRS comprises a database server that runs databases, a
content server that makes teaching content available, and a system
server that runs AIRS. Figure 4 show the flow of texts contents
recommendation with AIRS. First, a learner selects the content. Second,
the optimum form of expression is requested. Finally, the content is
provided from content server to the learner.
2.1 Learning texts
For AIRS, extracts from the courses, "Database System"
and "Basics of Information Processing," at our institution are
distributed as learning texts to meet the flow of lessons. Each learning
text is composed hierarchically by chapter and section and has three
different expression methods for each class. In other words, the
composition of the training text may be different even for the same
contents.
[FIGURE 3 OMITTED]
[FIGURE 4 OMITTED]
Expression method 1 may use minimum necessary explanations such as
formulas and definitions with figures only, Expression method 2 may use
static explanations such as statements and figures, and Expression
method 3 may use dynamic explanations using FLASH. For example, Section
8 of Chapter 3 in the database system describes the normal form only in
Expression method 1 and attempts the dynamically form in Expression
method 3. Table 1 defines the expression methods and Figure 5 shows the
hierarchical structure of the learning texts. We call each expression
method contents.
2.2 Recommendation of learning texts
AIRS has a learning text recommendation function. This section
describes the recommendation function.
2.2.1 Recommendation of expression method
A recommendation technologu now in use is collaborative filtering
[13]. Collaborative filtering is a technology to analyze user
characteristics from historical log data and to recommend information
from the historical logs of another user having similar characteristics.
For the automatic recommendation function of each expression method,
AIRS uses this technology [14]. Suppose that a database has browsing
history information indicating which section was browsed by Learner A
and which expression method was used for this browsing. Then, Learner B,
who has the similar browsing history information, is found. Based on
this information, an expression method is determined for the
recommendation to Learner A. Figure 6 shows a schematic of a
recommendation by collaborative filtering. The expression method
browsing historical logs 1 to 3 indicates that each expression method
was browsed and 0 indicates that the learning text was not browsed. When
Learner A browses Learning texts 3, Expression method 2 is recommended
by considering that Learner A may very possibly prefer Expression method
2 because Learner B, who has similar characteristics, browsed Expression
2 of Learning text 3.
[FIGURE 5 OMITTED]
[FIGURE 6 OMITTED]
2.2.2 Recommendation of section
This research examines the function of recommending and displaying
a section of strong relationship by calculating the relationship among
sections. By naming this as a "bidirectional recommendation
system," we proposed and developed its algorithm.
3. Bidirectional recommendation system
AIRS is not a self-completing system because it supports efficient
reviewing. This system is developed based on a concept that it grasps a
general idea quickly and provides efficient reviewing by selecting exact
points that need to be clarified and checked. However, learners strongly
tend to browse learning texts sequentially from the first one and follow
along a learning flow. For example, learning texts with as many as 210
sections may also be a factor that makes learners follow a sequence.
Therefore, we think efficient reviewing is possible by presenting
sections highly related to the current reviewing section if the sections
related to the current browsing section can be recommended.
3.1 Recommendation types
As typical recommendation systems, reference [15] introduces the
following two types of recommendations:
1. Recommendation by user relationship.
2. Recommendation by item relationship.
Recommendation by user relationship is making a recommendation from
the viewpoint of the user relationship by analyzing the action history
of each user. This is called user correlation and is used for the
aforementioned collaborative filtering.
Recommendation by item relationship is making a recommendation from
the viewpoint of the item relationship. This is called item correlation.
These two recommendations can be classified into the following two
analysis techniques:
1. Correlation based on the user's action history.
2. Correlation based on the item's information.
For bidirectional recommendation, the relationship among sections
is analyzed by using the former correlation based on the user's
action history.
3.2 Necessity of bidirection
The recommendation engines now in use for collaborative filtering
are mainly used to recommend commodities. Therefore, there is no need to
recommend having already purchased ones. In learning, however, not only
learners advance from the current learning text to the next, but also
learners return to the previous learning text. Therefore, we think it
necessary to make not only a unidirectional recommendation that
recommends new contents only but also a bidirectional recommendation
that also recommends the previously browsed contents.
3.3 Outline of bidirectional recommendation
When a learner is browsing the learning text "Basics of
substitution," it is natural to advance to the next step
"Basics of 'while' statement" or "Basics of
'if' statement." However, browsing the basic contents
"Variable types" again is also natural in learning. In other
words, the learning efficiency is expected to improve by recommending
not only learning texts frequently shifted from but also frequently
shifted to "Basics of substitution." Figure 7 shows a
schematic of a bidirectional recommendation.
3.4 Bidirectional recommendation algorithm
To realize bidirectional recommendation, a binary tree is created
with each section as the root. Sections highly related to a section are
linked as subnodes to form a small-scale tree up to Level 2. With each
section as the root, trees for as many as all sections are created under
the tree. The tree, different from the aforementioned hierarchical
structures of learning texts, indicates the relationship among sections.
By referencing this tree, highly related sections before and after are
found and recommended.
3.4.1 Tree configuration
This tree for bidirectional recommendation is created from
historical log data in which the shift of learning texts by each learner
is recorded. The log data integrated from learners is called the
learning text shift history. Figure 8 shows this integration.
To extract sections highly related to a section from this learning
text shift screen, we apply the user count algorithm used in the web
recommendation system PHOAKS by Terveen [16] et al. This is used to find
an item used by many users. Sections where a learner shifted from a
section can be considered as highly related. The following Equation (1)
is used to calculate the degree of the relationship with a learning text
from the learning text shi history:
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.] (1)
where
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.] (2)
If the frequency of shift from learning text [U.sub.i] to [U.sub.j]
is represented as qi,j, the degree of relationship [p.sub.a,j] of
[U.sub.j] with [U.sub.a] can be calculated. The right side of Equation
1. becomes equal to the ratio of learners who shifted to learning
text [U.sub.a].
[FIGURE 7 OMITTED]
[FIGURE 8 OMITTED]
A binary tree is created with the top two sections shifted from
Section 1 on Level 1 and the top two sections shifted from there on
Level 2. Repeating this on every section produces as many trees of
relationship as the sections. Figure 9 shows a tree indicating the
relationship of the section "Function."
3.4.2 Recommendation method using tree
A bidirectional recommendation is made by using the tree mentioned
in the preceding section. When a learner is browsing the section
"Function," the system refers to the tree of the relationship
shown in Figure 9. The system determines the sections of "Example
of function" and "Value exchange" most possibly related
to this tree as recommendation items.
[FIGURE 9 OMITTED]
[FIGURE 10 OMITTED]
Next, all tress are searched to find the trees containing the
section "Function." Here, the section "'if'
statement" shown in Figure 10 is found. From this tree, the section
"Basics of 'while' statement" is determined as a
recommendation item because it is found highly related to the section
"Function." Figure 10 shows this recommendation and gives a
sample tree of the section "'if" statement."
In addition, when a tree containing the section
"Function" is found, "Function" is also determined
by similar processing as a recommendation item.
The system presents the learner with the recommendation items
determined by the above processing. This enables the recommendation of a
highly probable shift to sections before and after the current learning
section. In other words, the contents to be learned next and reviewed
can be recommended in two directions. The system of recommending related
items before and after was named "bidirectional recommendation
system."
Figure 11 shows the general flowchart of the bidirectional
recommendation. Through the above flow, sections to be presented to a
learner browsing the section "Function" as related pages are
determined.
Up to four pages can be presented. Figure 12 shows the presentation
of related Japanese homepages in our developed recommendation system.
Precisely looking at figure 12, four recommended items such as example
of function, value exchange, basic of 'while' statement, and
inds of variabes are displayed.
[FIGURE 11 OMITTED]
[FIGURE 12 OMITTED]
4. Evaluation
This section compares the recommendation results available from the
log data of learning texts arranged along the lesson flow with that from
the log data of learning texts arranged at random. Thus, we verified the
validity of the related pages presented by the bidirectional
recommendation system and obtained the following results.
4.1 History log data for verification
In our university department, the course "Database
system" was held for sophomores in September 2007. A total of 125
students who took the course used AIRS. The acquisition of learning text
historical log data began on September 19, 2007. A total of 7355
historical logs acquired by the end of the course on December 27, 2007
were used for verification. Since the system is in actual operation, the
links of learning texts are arranged along the course flow. The tree of
relationship created from the historical log data are used as
Experimental data 1.
To acquire a learning text shift history for comparison, a total of
24 people used AIRS (one postgraduate, one research student, 22
undergraduates (13 seniors and 9 juniors)). In the period of use from
January 8, 2008 until January 18, 2008, AIRS was used for approximately
10 minutes a day for learning. For comparison, the links of learning
texts were arranged at random, totally irrespective of the course flow.
The total number of historical logs reached 768. The group of trees
created from the historical log data is called Experimental data 2.
Table 2 summarizes the experimental data.
Table 3 shows part of the tree stored in a database. The content_id
represents a section, layer1_1 and layer1_2 represent the first layer of
the tree, and layer2_1 to layer2_4 represent the second layer.
4.2 Precision of recommendation
By using the two tree groups mentioned in the previous section,
bidirectional recommendations were made and the recommendation results
were compared. The object learning texts here are composed of 58
sections from Chapters 1 and 3 of "Database system." The
recommendation using Experimental data 1 presented 152 related pages and
that using Experimental data 2 presented 212 related pages. The
recommendation using Experimental data 2 proposed more related pages
because various shift historical logs could be acquired due to the
random arrangement of learning texts.
As shown in Table 4, when the recommendation results using
Experimental data 2 were compared with those using Experimental data 1,
the number of the same recommendation items were 32 and the hit ratio of
the items against the related pages using the Experimental data 2 was
15.1% (=32/212). The number of sections containing at least one hit was
25 and the average ratio of the hit against the entire sections was
43.1% (=25/58).
Then, recommendations of not complete but approximate hits were
also compared as shown in Table 4. These can be considered to be
identical recommendations, such as "key insertion into B+
tree," identical to "key deletion from B+ tree," and
"union operation," identical to "difference
operation." The number of hits under this condition was 74 and the
ratio was 34.9% (=74/212). The number of recommendations containing at
least one hit was 45 and the average ratio was 77.6% (=45/58).
We measured the number of the recommended contents with the
bidirectional recommendation system on the base of the flow of lessons
previously mentioned. Table 5 shows the number of the recommended after
pages and the recommended before pages using Experimental data 1 and
Experimental data 2. In addition, Table 5 indicates the ratio of the
recommended after pages and the recommended before pages, too.
In case of Experimental data 1, the ratio of the recommended before
pages is 47.4 %, which is less than that of the after ones. On the other
hand, in case Experimental data 2 the ratio of the recommended before
pages is 60.0 %, which is greater than that of the after ones.
4.3 Discussion
The comparison this time may not be very accurate because the
counts and periods of experimental data were different. By comparison,
however, we could at least confirm the ratio of almost equally
recommended sections to be as high as 77.6%. In addition to that, we
found that the bidirectional recommendation system realizes the number
of the recommended before pages is almost equal to that of the
recommended after pages, such as 50 to 60%.
Collaborative filtering or any other recommended engine using
personal historical log data cannot make an optimum recommendation to
the first-time user. A recommend engine that calculates the item
relationship from the historical log data of all users makes a
recommendation that can present the same result even for the first-time
user. The bidirectional recommendation system using this method can
recommend learning texts appropriately to all learners if there is
historical log data about the learning texts. We show the comparison
results of the bidirectional recommendation system and collaborative
filtering in the Table 6.
5. Conclusion
By experiments, we could prove the bidirectional recommendation
system capable of making an appropriate recommendation, regardless of
the arrangement of learning texts. The bidirectional recommendation
system not requiring personal historical log data was also proved
effective for recommending learning texts.
We currently consider the following three future research subjects.
The first subject is to verify that learning texts can be recommended
appropriately to learners by the recommendation of sections by combining
the bidirectional recommendation and the recommendation of expression
using collaborative filtering. The second subject is to develop the
analysis method of the recommendation evaluation results. The third
subject is to compare our approach with existing data mining techniques
such as the correlation between users (resp. items) could be extracted
by association rules or sequential pattern approaches.
6. Acknowledgments
We would like to express our deepest gratitude to the subjects of
the experiment and to the many students who used AIRS. This research was
partially supported by a Grant-in-Aid for Scientific Research C (Subject
No. 18500731: Research on Learning Texts Recommendation Technology for
ELearning) and is partially supported by a Grant-in-Aid for Scientific
Research C (Subject No. 21500908: Research on Adaptive Recommendation
Technology based on Bidirectional Recommendation Technology for
E-Learning Texts).
Received: 11 July 2009; Revised: 28 August 2009; Accepted: 12
November 2009
References
[I] R. Wisher, "Sharable Content Object Reference Model
(SCORM) 2004 4th Edition Overview Version1.0", @ Advanced
Distributed Learning, pp.1-18 (31 March 2009).
[2] M. Yamaguchi, Y. Takahashi, T. Matsuzawa, T. Takahashi, and
S.Dohi, "Development and Evaluation of e-Learning System AIRS with
Multi-Expression Contents", IEICE-ET2006118 , vol.106, no.583, pp.
69-74 (2007).
[3] H. Hotta, and M. Hagiwara, "Design and Implementation of
Social-network-based Recommender System", IPSJ Transactions on
Databases, vol. 2, no.1, pp. 46-56 (2009).
[4] A. Sasaki, T. Miyata, Y. Inazumi, A. Kobayashi, and Y. Sakai,
"Web Content Recommendation System Based on Similarities among
Contents Cluster of Social Bookmark", IPSJ Transactions on
Databases, vol. 48, no.SIG20, pp. 14-27 (2007).
[5] K. Oku, S. Nakajima, J. Miyazaki, and S. Uemura,
"Context-Aware SVM for Context-Dependent Information
Recommendation", DBSJ Letters, vol. 5, no.1, pp. 5-8, ( 2006).
[6] D. Perea, J. Kay, I. Koprinska, K. Yacef, and O.R. Zaiane,
"Clustering and Sequential Pattern Mining of Online Collaborative
Learning Data", IEEE Transactions on Knowledge and Data
Engineering, vol. 21, no.6, pp. 759-772 (2009).
[7] G. Costagliola, V. Fuccela, M. Giordano, and G. Polese,
"Monitoring Online Tests through Data Visualization", IEEE
Transactions on Knowledge and Data Engineering, vol. 21, no.6, pp.
773-784 (2009).
[8] C. Pahl and C. Kenny, Interactive Correction and
"Recommendation for Computer Language and Training", IEEE
Transactions on Knowledge and Data Engineering, vol. 21, no.6, pp.
854-865 (2009).
[9] Y. Wada, T. Matsuzawa, M. Yamaguchi, and S. Dohi,
"Proposal and Verification of Bidirectional Recommendation System
for Learning Web Digital Texts", Proceedings of the Second
International Conference on the Applications of Digital Information and
Web Technologies (ICADIWT2009), pp. 210215 (August 2009).
[10] T. Matsuzawa, M. Yamaguchi, Y. Wada, and S.Dohi,
"Development of Bidirectional Recommendation System for
Texts", 2007 PC Conference, pp. 197-198 (2007-8).
[II] Y. Wada, Y. Hamadume, S. Dohi, and J. Sawamoto,
"Technology for Recommending Optimum Learning Texts Based on Data
Mining of Learning Historical Data", Proceedings of International
Workshop on Informatics (IWIN 2009) (September 2009) (to be appeared).
[12] Y. Takahashi, T. Matsuzawa, M. Yamaguchi, S. Dohi, and Y.
Wada, "Recommendation and Delivery of the Optimum Texts for
e-Learners", IPSJ SIG Technical Reports 2007CE-88(22), vol. 207,
no.12, pp.157-162 (2007).
[13] N. Ohsugi, A. Monden, S. Morisaki, and K. Matsumoto,
"Software Function Recommender System Based on Collaborative
Filtering", IPSJ Journal, vol. 45 Cno.1, pp. 267278 (2004).
[14] T. Matsuzawa, M. Yamaguchi, Y. Wada, and S. Dohi,
"Implementation of recommend system for learning texts", IPSJ
SIG Technical Reports 2008-CE-93(18), vol. 208, no.13, pp.127-132
(2008).
[15] N. Yamazaki, Software Design ISSN 0916-629, Vol. 268
CGijutsu-Hyohron Co., Ltd CTokyo Japan, 2007-8.
[16] L. Terveen, W. Hill, B. Amento, D. Mcdonald, and J. Creter,
"PHOAKS: a system for Sharing recommendations", Commnucations
of the ACM, vol. 40, no3, pp. 59-62(1997).
Yuji Wada (1), Syunsuke Matsuzawa (2), Mikuru Yamaguchi (3),
Shinichi Dohi (1)
(1) Information Management Tokyo Denki University Chiba, Japan
{yujiwada, dohi}@sie.dendai.ac.jp
(2) Mitsubishi Electric Information Systems Kanagawa, Japan
(3) NTT Data Systems Corporation Tokyo, Japan
Table 1. Definition of expression methods
Expression method 1 Formulas, definitions, and minimum
necessary figures only
Expression method 2 Static explanation
Expression method 3 Dynamic explanation
Table 2. Summary of experimental data
Experimental data 1 Experimental data 2
Period of use Sept. to Dec. 2007 Jan. 2008
User 125 students 24 students
Log count 7355 768
Arrangement Along course At random
Table 3. Tree stored in a database
content_id layer1_1 layer1_2 layer2_1
14 20 23 11
15 173 10 4
16 19 12 1
Table 4. Number and ratio of hits
Complete Hit Approximate Hit
Count (page) 32 74
Ratio (page) 15.1% 34.9%
Count (section) 25 45
Ratio (section) 43.1% 77.6%
Table 5. Number and ratio of the bidirectional recommendations
Experimental data1 Experimental data2
Count (after) 80 85
Ratio (after) 52.6% 40.0%
Count (before) 72 127
Ratio (before) 47.4% 60.0%
Table 6. Comparison of bidirectional recommendation and
collaborative filtering
Bidirectional Collaborative filt
recommendation
Recommended items Sections Expression methods
Recommendation Associations Associations
criterion between contents between user log
Recommendation Common to all Different to each
results learners learner
Required log data All learners' log Individual log
data