Towards Automatic Building of Learning Pathways
Patrick Siehndel
1
, Ricardo Kawase
1
, Bernardo Pereira Nunes
2
and Eelco Herder
1
1
L3S Research Center, Leibniz University Hannover, Hannover, Germany
2
Department of Informatics, PUC-Rio, Rio de Janeiro, RJ, Brazil
Keywords:
Learning Support, Learning Pathways, Digital Libraries.
Abstract:
Learning material usually has a logical structure, with a beginning and an end, and lectures or sections that
build upon one another. However, in informal Web-based learning this may not be the case. In this paper,
we present a method for automatically calculating a tentative order in which objects should be learned based
on the estimated complexity of their contents. Thus, the proposed method is based on a process that enriches
textual objects with links to Wikipedia articles, which are used to calculate a complexity score for each object.
We evaluated our method with two different datasets: Wikipedia articles and online learning courses. For
Wikipedia data we achieved correlations between the ground truth and the predicted order of up to 0.57 while
for subtopics inside the online learning courses we achieved correlations of 0.793.
1 INTRODUCTION
When learning about a new topic, especially in a do-
main that is new to the learner, it is not always di-
rectly clear in which order relevant resources can best
be read or learned, ensuring that the basic concepts
are introduced first, followed by more advanced ma-
terial that elaborates on these concepts. This is com-
monly known as Learning pathway. In fact, a learning
pathway is described as the chosen route, taken by a
learner through a range of learning activities, which
allows them to build knowledge progressively(Jih,
1996).
Our approach exploits latent concepts inside
learning objects and, according to the estimated com-
plexity of these concepts, provides a tentative order-
ing for a set of learning objects. The results provide
learners with an ordered learning script to follow, sim-
ilar to a course in which lectures are arranged in a
specific order.
For our method, we exploit information from
Wikipedia, which we use as an external knowledge
base. Wikipedia contains over 4 million articles (con-
cepts) that virtually cover all concepts that are rele-
vant for referencing. Further, each Wikipedia article
contains links to reference articles and it is manually
categorized. We exploit a set of features extracted
from Wikipedia and its category graph to estimate the
complexity of a given text. Our methods are based on
the assumptions that:
Wikipedia categories contain a useful link struc-
ture for ordering objects based on their difficulty;
Concepts that are mentioned inside Wikipedia ar-
ticles provide useful background knowledge for
understanding the meaning of an article.
Our method uses the Wikipedia Miner
1
toolkit for
detecting concepts in the analyzed learning objects.
The detected concepts are basically text snippets that
can be related to a Wikipedia article. All Wikipedia
articles belong to one or more categories, and these
categories are organized in a graph structure. We use
this graph structure for identifying categories that are
more general and therefore supposedly known by a
user.
The main aspect of our work is to help learners
to identify a meaningful order of given learning ma-
terial. An example: in mathematics, it is obvious
that learning basic principles like summing or divid-
ing should come before starting with topics such as
‘curve sketching’. Essentially, the problem we aim to
solve can be summarized as follows: given a set of
learning objects, we bring them into a reasonable or-
der, to help learners finding a good starting point as
well as a good way through the provided material.
The rest of the paper is organized as follows: In
Section 2, we discuss related work on the topics of
learning object recommendation and ordering. The
proposed method is explained in detail in Section 3.
1
http://wikipedia-miner.cms.waikato.ac.nz/
270
Siehndel P., Kawase R., Pereira Nunes B. and Herder E..
Towards Automatic Building of Learning Pathways.
DOI: 10.5220/0004837602700277
In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 270-277
ISBN: 978-989-758-024-6
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
The experimental evaluation of the whole process is
presented in Section 4, where we used two different
data sets to analyze the performance of our method:
Wikipedia articles and online learning courses from
Khan Academy. We conclude the paper in Section 5
by summarizing our main contributions.
2 RELATED WORK
A dynamic course generator is presented by Farrell
et al. (Farrell et al., 2004). The course is assembled
based on keyword queries and the metadata of learn-
ing objects contained in a given repository. The se-
quence relies on the relationships that are manually
assigned to each learning object and its classifica-
tion (e.g. introduction, methodology or conclusion).
Hence, the objects are selected and reordered accord-
ing to a user query and its classification. Chen (Chen,
2008) present an evolutionary approach that uses a
genetic algorithm to generate a learning plan. The
genetic approach is based on a pretest performed by
students, where missed concepts help in creating new
learning plans, according to concepts and levels of
difficulty of the learning objects. Our approach fol-
lows the dynamic nature of these approaches, since
we only need an input concept to determine a learn-
ing object sequence.
Ullrich and Melis (Ullrich and Melis, 2009) or-
der learning objects according to the learning goal of
each student. For this, they classify objects into dif-
ferent classes, such as illustrate or discover, where the
course is assembled by a sequence of examples or in
depth.
In the areas of Intelligent Tutoring Systems and
Adaptive Hypermedia the adaptive sequencing is
common technique (Brusilovsky and Mill
´
an, 2007).
In scenarios where metadata for the given learning
objects is available systems like PASER (Kontopou-
los et al., 2008) allow the calculation of a learning
path. In our scenario we address informal learning
situations in which this metadata is not available.
Another perspective on the sequencing of learn-
ing objects is discussed by Kickmeier-Rust et
al. (Kickmeier-Rust et al., 2011), where they use a
combination of a storytelling model and competence
structures to identify the learning state of a student in
games. By identifying the state of the student, they
propose a new sequence of learning objects, while
keeping the story lines. Limongelli et al. (Limongelli
et al., 2009) present a framework to create personal-
ized courses. The sequencing of learning objects is
generated taking into account the cognitive state of
the student and her learning style. Sequences change
according to the results obtained by the students, in
order to cover a concept not understood. Missed
concepts are identified through exercises during the
learning process. Instead of discovering the learning
state of students, we focus on a general applicable ap-
proach. Our approach identifies which topics are nec-
essary to understand a topic independent of the stu-
dent. On the one hand, we do not provide personal-
ized learning paths; on the other hand, we overcome
the cold-start problem where there is no a priori infor-
mation of the students.
Champaign and Cohen (Champaign and Cohen,
2010) introduce a work based on student development
after consuming a given learning object. Each student
is assessed and the most successful sequence of learn-
ing objects is selected and recommended to students
with similar profiles. Similarly, Knauf et al. (Knauf
et al., 2010) focus on similar profiles to recommend
similar learning paths. However, the similarity be-
tween students is based on learning models that de-
scribe the abilities of each student. A path taken by
a successful student is recommended to another one
with similar characteristics. In contrast, the goal of
our approach is to recommend learning objects fol-
lowing the learning goals of a student; as the student
selected a topic to learn, the sequencing is determined
by knowledge and concepts needed to understand a
learning object.
3 METHOD
Our method for ordering learning objects and provid-
ing background links is divided into two main steps.
The first step is the annotation of the content with
links to relevant Wikipedia articles. This step is de-
scribed in more detail in Section 3.1. The second step
exploits detected topics and Wikipedia as a knowl-
edge base to calculate the order of learning objects
in a given set.
3.1 Annotation and Features
For annotating the content of the given learning ob-
jects, we used the Wikipedia-Miner Toolkit (Milne
and Witten, 2008). The tool annotates a text (links
terms to articles) in the same way a human would do
it in Wikipedia. With this information, based on the
detected topics inside a given learning object, we cal-
culate a set of features that indicate the complexity
and relevance of a topic. The features we use for or-
dering the given objects are:
1. Number of inlinks: the number of Wikipedia arti-
cles that link to the detected topics.
TowardsAutomaticBuildingofLearningPathways
271
2. Number of outlinks: the number of links to other
articles contained by detected topics.
3. Text Length of Linked Articles: the length of the
detected articles.
4. Average Word Length of Linked Articles: the av-
erage length of the words in the detected articles.
5. Average Word Length of Learning Object: the av-
erage length of the words in the learning object.
6. Distance to Root of Linked Articles: the average
distance to the root categories of the articles.
7. TF/IDF Score of Words in Linked Articles: the
TF/IDF values of the words inside the detected ar-
ticles.
8. TF/IDF Score of Words in Learning Object: the
TF/IDF values of the words inside the learning ob-
ject.
The first two features are chosen based on the as-
sumption that the number of inlinks and outlinks are
indicators of the generality of a Wikipedia article: if
many articles link to one page, it indicates that this
concept is a basic (popular) concept. As in (Kamps
and Koolen, 2009), inlinks and outlinks are deemed to
be good indicators of an article’s relevance to a given
topic.
We also assume that the text length and the aver-
age words length are good indicators about how com-
plex a topic is. Another important feature is the aver-
age distance of the related categories to the root node
of the category tree. This feature is based on the as-
sumption that more complex topics inside Wikipedia
are deeper down in the category graph and is compa-
rable to the generality feature in (Milne and Witten,
2012). The TF/IDF feature represents the assumption
that words that rarely appear inside our corpus are
related to more complex topics. All of our features
are represented in four ways: we use the minimum,
maximum, mean and standard deviation of each of
these features to represent one learning object, which
gives us a 32-dimensional float vector representing
one learning object.
3.2 Learning to Order Objects
Our ordering approach is based on machine learning
algorithms. The given features are used to generate
a model that calculates a score for every learning ob-
ject. This score indicates the estimated complexity of
the concepts within the learning object. In our ex-
periments, we used four different machine learning
algorithms to produce the models. Two of these algo-
rithms create a tree structure based on the given train-
ing data. In addition, we used a regression model and
a Support Vector Machine for regression to calculate
the order of the given objects.
Note that the score is based on a comparison of
learning objects, and this only makes sense if the
learning objects cover related topics from a single do-
main. For example, answering the question if one
should learn a topic like ‘European History’ before
learning ‘Linear functions’ is out of the scope of this
paper. Due to the different nature of different learning
domains, the quality of the generated order is higher
when only a single domain is considered. Our ap-
proach can help users to decide which object in a
given set might be useful to be learned first, assum-
ing that the objects are related per se.
4 EXPERIMENTAL EVALUATION
In this section we evaluate the performance of the
proposed method by analyzing the quality of the
predicted order of different sets of learning objects.
We performed our evaluations with two different
datasets: Wikipedia articles from different domains
and a large set of online learning courses from the
Khan Academy
2
. We chose these datasets, as they
contain elements that can be used as a ground truth
that indicates how complex a given element is. For
the Wikipedia articles, we chose the distance from the
root node as an indicator for complexity. For the on-
line course dataset, we exploited its hierarchical struc-
ture, which also indicates an order in which the ele-
ments should be learned.
4.1 Ordering Evaluation with
Wikipedia Data
In this section, we describe the outcomes of our ex-
periments with Wikipedia data. We show that there
are useful correlations between the depth of a concept
in the Wikipedia tree and other features that we use to
define the complexity of a topic. For Wikipedia arti-
cles there is no predefined order that defines which ar-
ticle one should read first. We decided to take the dis-
tance to the root node of an article as an indicator for
the complexity of the given topic. Every Wikipedia
article belongs to at least one category, and based on
the conventions how articles are added to categories,
the articles should be added to the most specific cat-
egory. Due to this, articles that belong to lower level
categories cover in most cases more specific topics.
2
https://www.khanacademy.org/
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
272
Figure 1: Correlation between features and distance to root of Wikipedia articles.
Table 1: Average distance to the root of articles from different categories.
Category Arts Culture Geography History
Avg. Distance 4.329 5.468 5.726 4.436
Category Humanities Mathematics Science Average
Avg. Distance 4.869 3.321 4.1 4.06
4.1.1 Dataset
In order to understand how the different features
correlate and to get a first overview, we analyzed
sets containing 500 articles from different Wikipedia
Main Topics. The main topic is defined by following
the category graph up to the first level of categories.
The categories we chose to analyze are: Arts’, ‘Cul-
ture’, ‘Geography’, ‘History’, ‘Humanities’, ‘Math-
ematics’ and ‘Science’. All of the mentioned cate-
gories also have a relation to topics taught in school
and are therefore of special interest.
4.1.2 Ordering Wikipedia Articles
For learning an order inside the Wikipedia articles,
we started by analyzing the distance to the root of arti-
cles belonging to different categories. As results show
in Table 1, different categories have different average
distances to the root node. This is caused by the sin-
gular link and category structures inside the different
categories. In comparison to ‘Mathematics’, which
seems to have a relative flat category graph, we see
that the average distance to the root for ‘Geography’
articles is much higher. Due to the large differences
between the different categories, we decided to also
analyze the correlations between the distance to the
root and the calculated features for each category sep-
arately. The results for 6 of the features we analyzed
is shown in Figure 1.
Overall we analyzed the correlations between 33
different features gathered from an article and its dis-
tance from the root of the category graph. Since our
primary goal is to calculate an optimal order in which
items should be learned, we analyzed how our fea-
TowardsAutomaticBuildingofLearningPathways
273
Table 2: Results of predicted distance to root for Wikipedia articles using Machine Learning Algorithms.
SMOReg M5P Additive Regression Bagging
All Articles 0.4878 0.5004 0.4422 0.5054
Arts 0.3587 0.4019 0.3611 0.3836
Culture 0.5253 0.509 0.5076 0.5213
Geography 0.0502 0.0027 0.3591 0.3835
History 0.2819 0.2516 0.267 0.2056
Humanities 0.0076 0.213 0.1777 0.2373
Mathematics 0.0907 0.4225 0.306 0.4313
Science 0.074 0.5704 0.5309 0.5478
tures correlate with the complexity of the Wikipedia
articles. We divided the articles in two groups,
based on their positions inside the category graph of
Wikipedia. The first group consists of basic articles
(distance<4), while the second group consists of ad-
vanced articles (distance>=4). Figure 1 shows the
correlations between these groups and six features.
The singularities between the different categories in-
dicate that learning objects of each category may re-
quire different strategies.
It is noteworthy to mention that the feature “Dis-
tance to Root (Article)” is the most important feature.
This feature is calculated based on the links to other
articles inside the article that we want to rank. The
positive correlations of maximum, minimum, and av-
erage show that articles that are already deep inside
the Wikipedia category graph tend to have links to ar-
ticles that are also deep inside this graph.
Another noteworthy fact is that the average word
length inside the ranked articles has a negative cor-
relation with the group index. This indicates that
longer words (on average) tend to be in articles that
are higher in the tree; this was not expected, as we ex-
pected to find longer words in articles that are deeper
in the category tree.
The correlations found between the distance to the
root of an article and the several features that we ex-
tracted from the articles indicate that it is possible to
calculate an order for learning objects. To further ana-
lyze how well these features can be used to predict the
complexity of a given text, we used machine learning
algorithms to predict the actual distance to the root
of a given article based on all the extracted features.
The algorithms used are all integrated in Weka3
3
(Hall
et al., 2009). We used SMOreg (Shevade et al., 2000),
which is as implementation of a Support Vector Ma-
chine for regression, M5P (Quinlan, 1992)(Wang and
Witten, 1997), which implements algorithms for cre-
ating M5 Model trees and rules, AdditiveRegres-
sion (Friedman and (y X)-values, 1999), which is an
improved regression-based classifier, and a Bagging-
3
http://www.cs.waikato.ac.nz/ml/weka/
Algorithm (Breiman, 1996) based on a RepTree algo-
rithm.
We predicted the distance for every category on
its own and for all articles of the different categories
together. The results are based on a 10-fold cross val-
idation and shown in Table 2. We see that, for differ-
ent categories, different algorithms produce the best
result. On average, the Bagging-based approach pro-
duced the best results. Additionally, this algorithm
shows a very low standard deviation over the different
categories. In general, we see that the distance to the
root for articles of the topics ‘Arts’, ‘Humanities’ and
‘Geography’ is harder to predict than in the cases of
articles from ‘Science’, ‘Mathematics’ and ‘Culture’.
In summary, our first set of experiments shows
that it is possible to predict a meaningful order for
Wikipedia articles based on features extracted from
Wikipedia’s link structure and the textual features
within these articles.
4.2 Evaluation with Online Learning
Data
In addition to the evaluation on Wikipedia Data, we
performed an analysis of the proposed method on a
real world dataset of learning courses. While the
outcomes of the first experiments proved that the as-
sumption that features gathered from text snippets and
related Wikipedia articles can be used to calculate the
complexity of given texts (by means of the distance
of the article to the root node), we now use the given
order of a set of learning objects as ground truth.
4.2.1 Dataset
The dataset used for this series of experiments was ex-
tracted from the online courses of Kahn Academy
4
.
We analyzed the text of 2508 different lectures related
to the main topics ‘Math’, ‘Science’ and ‘Humani-
ties’. These items are organized in a three-level hier-
archy: the first level is a general category like ‘Sci-
4
https://www.khanacademy.org/
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
274
Figure 2: Correlation between features and learning object locations inside a course.
ence’; inside this category, there are different groups
like ‘Chemistry’ or ‘Biology’. Below this level are
the actual courses, like ‘Cell division’ or ‘Oxidation
reduction’. The learning objects within a course are
manually ordered in a meaningful way, representing
the order in which a student is supposed to learn.
Overall, we analyzed 110 different learning courses.
Statistics on these courses are shown in Table 3
We chose to take the given order of the objects
inside a course as ground truth for evaluating our ap-
proach. Calculating an order for a higher level does
not make sense for all the given objects. For exam-
ple, it is hard to say that ‘Biology’ should be learned
before ‘Chemistry’, or that the ‘Industrial Revolution’
has to be learned before Art History’, but when learn-
ing about matrices it seems to be useful to learn ‘Ma-
trix multiplication’ before ‘Determinant calculation’.
4.2.2 Ordering Learning Objects
We started the analysis of learning objects in the same
way as we did for Wikipedia articles: by analyzing
correlations between the order of objects and the cal-
culated features. The results showed us that the order
of these items follows a more complex structure that
is hard to grasp by just taking into account the lin-
ear relations between the order of the objects and the
values of the calculated features.
Figure 2 displays the correlation values between
learning objects and different features. We can see
that with the shown features we do not obtain the
same correlations for all kind of topics as we got for
the Wikipedia articles. A closer look at the anno-
tated articles revealed that this is most likely caused
by noise inside the transcripts of the online courses.
This noise origins from the fact that the transcripts
only represent the spoken content of the video lec-
tures, which is hard to understand without the whole
content of the video. Combined with a fair number
of non-relevant remarks that were still included in the
transcript, the quality of the extracted articles is not
as good as in the previous experiment. Despite this
drawback, for many of the features there are clear re-
lations between the features and the location inside
the course. We decided to perform the same tests as
before to calculate the actual position of the learning
objects inside the courses.
The results of this series of experiments are dis-
played in Table 4. The results were produced using all
mentioned features using a 10-fold cross validation.
The highest overall achieved correlation between
the actual position and the calculated position was at
0.554, when the algorithm is applied on all available
learning objects. When training and testing on subcat-
egories of the data, we achieve results of up to 0.793.
The results differ strongly between the different do-
mains of the online courses. For the elements of the
domain ‘Humanities’ none of the tested algorithms
achieved good results, while the order of ‘Science’-
related elements was relatively well calculated by all
algorithms. We also see that not all different algo-
rithms are in the same way suitable for predicting the
actual rank of the items. On average, the best results
were achieved using the Bagging approach.
TowardsAutomaticBuildingofLearningPathways
275
Table 3: Statistics on the Learning Object Dataset.
Main Category #Groups #Courses Avg. Items per Course
Humanities 2 18 30.92
Mathematics 5 40 33.38
Science 4 47 10.76
Table 4: Results of predicted positions of learning objects using Machine Learning Algorithms.
SMOReg M5P Additive Regression Bagging
All LOs 0.292 0.338 0.408 0.554
Mathematics 0.094 0.365 0.397 0.416
Science 0.357 0.779 0.71 0.793
Humanities 0.056 0.141 0.135 0.127
Wikipedia 0.488 0.500 0.442 0.505
4.3 Discussion
The series of conducted experiments shows that the
proposed method can be used for calculating the com-
plexity of a given topic, based on text features and fea-
tures extracted from Wikipedia. Additionally, there
are evidences that for some categories it is harder to
predict its complexity than for others. Especially con-
tent from the area of Humanities seems to be harder
to order than content from disciplines like Mathemat-
ics or Science. This might be due to a more com-
plex structure of the underlying content: in Mathe-
matics, the order in which elements need to be learned
is much clearer, due to the fact that concepts build up
on one another. By contrast, in disciplines like His-
tory, this is in most cases not true.
5 CONCLUSION
In this paper, we presented a method for ordering
learning objects based on the complexity of the cov-
ered content. The proposed method is based on fea-
tures that are extracted from the original items, as well
as from the knowledge stored in Wikipedia. By using
Wikipedia, we exploit a knowledge base that is con-
stantly updated and freely available. We analyzed the
performance of the method on two different datasets,
and achieved correlations between the ground truth
and the predicted values of up to 0.793 for special
topics of learning courses. The results show that text-
based learning material can automatically be sorted in
a meaningful order. However, the quality varies, de-
pending on the domain and the textual quality of the
elements. For example, written text from Wikipedia
is easier to order than noisy video transcripts.
The results of the experiments also showed that
the proposed method works better with domains like
Mathematics or Science compared to domains like
Humanities or History. In general it seems to be
useful to train different models for different domains
since the values of some features vary over different
domains.
The proposed order, as provided by our method,
can help learners to find a good starting point for their
learning pathways inside a set of learning resources.
Also, it might help them to choose how to continue
their learning process once a lesson has been learned
or a resource has been visited. In addition, the meth-
ods may help teachers to analyze how the complexity
of their courses evolves over time, which may help
them to find a more suitable order for the elements
they are teaching. A big advantage of the proposed
method is that no metadata is required for calculat-
ing an order. This allows to incorporate every kind of
textual resource into the learning process.
As future work we plan to build a model that
can identify prerequisite knowledge for given learn-
ing courses. This will allow teachers and learners
to better build a background knowledge for teach-
ing/learning activities.
ACKNOWLEDGEMENT
This work has been partially supported by the Euro-
pean Commission under ARCOMEM (ICT 270239)
and QualiMaster (ICT 619525).
REFERENCES
Breiman, L. (1996). Bagging predictors. Machine learning,
24(2):123–140.
Brusilovsky, P. and Mill
´
an, E. (2007). User models for
adaptive hypermedia and adaptive educational sys-
tems. In The adaptive web, pages 3–53. Springer-
Verlag.
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
276
Champaign, J. and Cohen, R. (2010). A model for con-
tent sequencing in intelligent tutoring systems based
on the ecological approach and its validation through
simulated students. In Guesgen, H. W. and Murray,
R. C., editors, FLAIRS Conference. AAAI Press.
Chen, C.-M. (2008). Intelligent web-based learning system
with personalized learning path guidance. Computers
& Education, 51(2):787 – 814.
Farrell, R. G., Liburd, S. D., and Thomas, J. C. (2004). Dy-
namic assembly of learning objects. In Proceedings
of the 13th international World Wide Web conference
on Alternate track papers & posters, WWW Alt. ’04,
pages 162–169, New York, NY, USA. ACM.
Friedman, J. H. and (y X)-values, O. K. (1999). Stochastic
gradient boosting. Computational Statistics and Data
Analysis, 38:367–378.
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,
P., and Witten, I. (2009). The WEKA data mining soft-
ware: an update. Special Interest Group on Knowl-
edge Discovery and Data Mining Explorer Newslet-
ter, 11(1):10–18.
Jih, H. J. (1996). The impact of learners’ pathways on learn-
ing performance in multimedia computer aided learn-
ing. J. Netw. Comput. Appl., 19(4):367–380.
Kamps, J. and Koolen, M. (2009). Is wikipedia link struc-
ture different? In Proceedings of the Second ACM
International Conference on Web Search and Data
Mining, WSDM ’09, pages 232–241, New York, NY,
USA. ACM.
Kickmeier-Rust, M., Augustin, T., and Albert, D. (2011).
Personalized storytelling for educational computer
games. In Ma, M., Fradinho Oliveira, M., and
Madeiras Pereira, J., editors, Serious Games Develop-
ment and Applications, volume 6944 of Lecture Notes
in Computer Science, pages 13–22. Springer Berlin
Heidelberg.
Knauf, R., Sakurai, Y., Takada, K., and Tsuruta, S.
(2010). Personalizing learning processes by data min-
ing. In Advanced Learning Technologies (ICALT),
2010 IEEE 10th International Conference on, pages
488 –492.
Kontopoulos, E., Vrakas, D., Kokkoras, F., Bassiliades, N.,
and Vlahavas, I. (2008). An ontology-based planning
system for e-course generation. Expert Systems with
Applications, 35(1):398–406.
Limongelli, C., Sciarrone, F., Temperini, M., and Vaste, G.
(2009). Adaptive learning with the ls-plan system: A
field evaluation. Learning Technologies, IEEE Trans-
actions on, 2(3):203 –215.
Milne, D. and Witten, I. H. (2008). Learning to link with
wikipedia. In Proceedings of the 17th ACM con-
ference on Information and knowledge management,
CIKM ’08, pages 509–518, New York, NY, USA.
ACM.
Milne, D. and Witten, I. H. (2012). An open-source toolkit
for mining wikipedia. Artificial Intelligence.
Quinlan, J. R. (1992). Learning with continuous classes.
In Proceedings of the 5th Australian joint Conference
on Artificial Intelligence, volume 92, pages 343–348.
Singapore.
Shevade, S. K., Keerthi, S. S., Bhattacharyya, C., and
Murthy, K. R. K. (2000). Improvements to the SMO
algorithm for SVM regression. Neural Networks,
IEEE Transactions on, 11(5):1188–1193.
Ullrich, C. and Melis, E. (2009). Pedagogically founded
courseware generation based on htn-planning. Expert
Systems with Applications, 36(5):9319 – 9332.
Wang, Y. and Witten, I. H. (1997). Inducing model trees
for continuous classes. In Poster Papers of the 9th
European Conference on Machine Learning (ECML
97), pages 128–137. Prague, Czech Republic.
TowardsAutomaticBuildingofLearningPathways
277