=Paper=
{{Paper
|id=Vol-2674/paper05
|storemode=property
|title=Using Knowledge Graph for Explainable Recommendation of External Content in Electronic Textbooks
|pdfUrl=https://ceur-ws.org/Vol-2674/paper05.pdf
|volume=Vol-2674
|authors=Behnam Rahdari,Peter Brusilovsky,Khushboo Thaker,Jordan Barria-Pineda
|dblpUrl=https://dblp.org/rec/conf/aied/RahdariBTB20
}}
==Using Knowledge Graph for Explainable Recommendation of External Content in Electronic Textbooks==
https://ceur-ws.org/Vol-2674/paper05.pdf
Using Knowledge Graph for Explainable
Recommendation of External Content in
Electronic Textbooks
Behnam Rahdari[0000−0001−6514−912X] , Peter Brusilovsky[0000−0002−1902−1464] ,
Khushboo Thaker[0000−0003−3619−9376] , and Jordan
Barria-Pineda[0000−0002−4961−4818]
University of Pittsburgh, Pittsburgh PA 15260, USA
ber58,peterb,k.thaker and jab464 @pitt.edu
Abstract. Over the last 10 years, the world experienced a rapid in-
crease in volume and diversity of digital learning resources. The abun-
dance of digital resources could support a range of powerful educational
scenarios, which were not available before. In this paper, we introduce a
novel approach that combines fully automatic knowledge modeling, stu-
dent modeling, and content recommendation approaches to recommend
relevant Wikipedia articles for students working with online electronic
textbooks. An assessment of our approach with real classroom data in-
dicated several benefits of our approach over the baseline and revealed
interesting patterns of students’ behavior while using the system.
Keywords: Recommender Systems · Personalization · Knowledge Graph
· Student Model · Electronic Textbooks · Concept Extraction.
1 Introduction
Over the last 10 years, the world experienced a rapid increase in volume and
diversity of digital learning resources. On the one hand, a variety of tutorials,
online textbooks, educational videos, and other open educational resources were
posted online to complement traditional textbooks. On the other hand, almost
all traditional textbooks have migrated to digital format and become available
online [6]. The abundance of digital resources could support a range of powerful
educational scenarios, which were not available before. For example, if a textbook
section is challenging for a learner to comprehend, she could be recommended
some useful external materials, which explains the same topics in a way that
is more adapted to her knowledge and interests. If the student fails to solve
problems or answer questions due to the lack of prerequisite knowledge, she
could be guided to the readings that introduce or review the missing knowledge.
Copyright c 2020 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0).
2 B. Rahdari et al.
The ideas of this “smart” learning have been explored in early projects fo-
cused on adaptive textbooks [4], which demonstrated both the feasibility and the
value of knowledge-driven personalized reading support. However, these early at-
tempts focused mostly on so-called closed corpus personalization, i.e., guiding
readers to most relevant parts of the textbook itself. A few attempts to offer open
corpus personalization [3], i.e., recommending most relevant external resources,
failed to scale up because it required expensive expert-driven knowledge analysis
of every external resource [7]. The goal of the project presented in this paper
was to move the idea of open corpus personalization in user-adaptive textbooks
closer to reality using fully automatic knowledge modeling, student modeling,
and content recommendation approaches. As a test-bed for exploring this idea,
we selected the case of recommending relevant Wikipedia pages for a textbook
user - both proactively, when she starts reading a new section and remedially
following an attempt to answer textbook questions.
Following a brief review of related work, this paper introduces the interface of
our digital textbook reading system with embedded recommendations. The next
four sections introduce the underlying mechanisms of our intelligent textbook:
the domain and student modeling approaches, the knowledge graph, and the
recommendation approach based on this infrastructure. The following section
presents the evaluation of the recommendation approach based on real classroom
data. We conclude with a discussion and future work plans.
2 Related Works
Research on recommendation of related reading sources has deep roots in re-
search on educational hypertext and hypermedia. Historically, it has been per-
formed under the name of “intelligent hypertext”, since this approach recom-
mended resources that were not connected by a human-authored link. Research
on intelligent hypertext started in the early days of the educational hypertext
field and originally focused on linking resources using term-based resource sim-
ilarity [10]. Simple keyword-based approaches have been gradually replaced by
semantic-level similarity based on the Semantic Web ideas and domain ontol-
ogy [5, 11] and, later, by modern text-processing approaches such as topic mod-
eling and concept extraction [1, 12].
The emergence of MOOCs and the accumulation of large volume of educa-
tional content online encouraged a new wave of research on “intelligent” linking
focused on connecting primary learning content such as textbooks and MOOCs
with several kinds of external learning resources such as videos, Wikipedia pages,
or research papers [1, 9].
3 Explainable Wikipedia Recommendations in a Digital
Textbook
We implemented Wikipedia recommendation interface in the context of a dig-
ital textbook system Reading Mirror [2]. Reading Mirror is an online reading
Title Suppressed Due to Excessive Length 3
system specifically focused on supporting student learning from modern digital
textbooks (Figure 1). The system supports textbooks in PDF and HTML for-
mats augmenting the reading process with a range of advanced features such as
self-assessment, student knowledge modeling [13], and reading progress tracking
with social comparison (Figure 1D).
Fig. 1. Reading Mirror interface. (A) Table of contents (B) Progress bar with mirrored
social comparison interface; (C) Navigation bar; (D) Reading area; (E) Recommended
Wikipedia articles
Automatic knowledge-driven linking (recommendation) of Wikipedia articles
is one of the newest features of the system. A new set of five most relevant
Wikipedia articles are generated for the target user in two cases. First, when
a user starts reading of a new textbook unit (section or subsection), a set of
best supportive articles is generated. These articles attempt to provide some
alternative reading sources for the knowledge which the unit aims to present as
well as prerequisite knowledge which are required to understand the content of
this subsection, but not yet mastered by the target user (as evidenced by her
knowledge model). Second, when the student answers a question incorrectly a
set of best remedial articles is generated. Remedial articles focus on alternative
presentation of knowledge that the student failed to master (as evidenced by the
wrong answer).
(a) Recommended Item Dialog box (b) Explanations Dialog box
Fig. 2. The preview of recommended item and explanations dialog boxes.
4 B. Rahdari et al.
As shown in Figure 1E, the recommended articles are presented on the left
side of the interface along with internal table-of-contents links (Figure 1A). The
links are ranked by the expected value of each article (importance) in the current
context. A colored heat-bar visualizes this importance: here “Green” means more
relevant and “Red” means less relevant. When student clicks on a recommended
item, the summary of the Wikipedia article will appear first (Figure 2a). After
clicking on “Read the Full Article” button, the complete version of the Wikipedia
article is being presented to the user.
To make the recommendation more transparent, we offered a brief explana-
tion for each recommended item, which could be obtained by clicking on “(Why)”
link at the right-hand side of the item. The goal of explanation is helping stu-
dents to understand the reason for recommending the article. The explanation
dialog (Figure 2b) consists of two parts. The first part lists top three domain
model concepts that user learns by reading this article. These concepts are top
three items (with highest value) in the list of “useful Knowledge” (see section
5) when this recommendation is generated. The second part explains the reason
why the presented concepts are specifically important for the target user. These
reasons are presented as a bullet list and are generated using the current state
of the user knowledge reflected in the student model.
4 Building the Knowledge Graph
We built a graph structure to represent the underlying knowledge layer of our
system. The entities and relationships in this graph demonstrate the connection
between the textbook content, Wikipedia and the student model. The knowledge
graph is hosted on a native graph database (Neo4j) and used for both storing the
data and generating the recommendations.The overall schema of our knowledge
graph is presented in Figure 3.
In the following, we will describe the process of building the knowledge graph.
Question Related_to Article Has_Page Category
Has_Child
Belongs_to Related_to
Section
Related_to
Includes
Includes
Concept Knows User
Fig. 3. Graph Schema representing the entities of the knowledge graph and the rela-
tionship between them
Title Suppressed Due to Excessive Length 5
4.1 Wikipedia Entities Representation
Wikipedia contains a large number of articles. Only a small number of them are
related to the context of any given textbook. To ensure the level of relatedness
and to increase the overall performance of our system we generated a subset
of Wikipedia articles to be recommended to the students. In order to find the
most relevant articles to the context of a textbook in the domain of computer
and information science, we used Wikipedia API and started from a high-level
Wikipedia category, namely “Category:Subfields of computer science” and re-
cursively extracted the subcategories and all the articles associated with them.
Since the Wikipedia category structure is not loop-free, we manually stopped
the recursion after three steps. For each Wikipedia article, we extracted the
following information using the Wikipedia API:
– Title: title of the Wikipedia page
– Summary: a brief description of the article that appear at the top of the
page.
– Full Text: the complete textual content of the page
The total number of 1141 categories and 47772 articles are extracted and added
to the graph during this step. We then connect these entities in the graph using
“Has Page” (when an article belongs to a category) and “has Child” (when a
sub-category belongs to a category) relationships.
4.2 Textbook Entities Representation
The content of the textbook is represented using three main entities: sections,
questions, and concepts. For simplicity, we consider all the variation of the sec-
tion (i.e, sub-sections and sub-sub-sections) as one entity (Section). Each section
or question is associated with a set of concepts that it presents or assesses using
“Includes” relationship. Each question is connected to a section with the “Be-
longs to” relationship. During our calculations, we represent a union of concepts
associated with a question and its corresponding section as relevant concepts to
the question. Sections and questions are connected to their matched concepts
via “Includes” relationship.
4.3 Linking Concepts and Wikipedia Articles
In order to create a relationship between the content of the textbook and ex-
tracted Wikipedia articles, we perform a full-text search on the textual represen-
tation of the articles using each concept as a query. The graph database (Neo4J)
provided us with the full-text indexing functionality which we used to create
the index for the combination of article title, summary, and full-text. To find
the most relevant articles for each concept we used the Neo4J internal full-text
search algorithm (Lucene). This algorithm provides us with a ranked list of rel-
evant articles as well as a relative score that shows the relevance of each result
to the input query. We used this information to connect each concept with the
top 100 relevant articles alongside with their relevance score. The “Related to”
relationship is representing this connection in the graph schema.
6 B. Rahdari et al.
4.4 Student Model Representation
Student models utilize a log of student actions as the input, and predict student
performance with practice activities. To generate and maintain students’ knowl-
edge state for each domain model concept, we used a Comprehension Factor
Analysis framework (CFM ) [13]. CFM incorporates student reading behaviour
along with activity performance which has proved to be beneficial in case of learn-
ing systems based on online textbooks [14]. At each student practice opportunity
CFM provides the probability of student’s success at that point. For our case
we require probability on each domain concept associated with that opportunity
(reading as well as questions). To generate this opportunity we generate prob-
ability of success for each concept at that opportunity ( details in [15]).In the
graph representation, student model maintains the level of knowledge of student
with the concepts at each interaction. This information is represented by a link
(called “knows”) between the “user” node and “concept” node and contains the
following properties:
– Interaction ID: specifies the interaction which the user gained some level of
knowledge with respect to the target concept.
– Type: determines the type of activity (reading a section or answering a ques-
tion) that lead to learning the concept.
– Name: stores the name of the section/question.
– Results: if the type is question, represents whether student answer that ques-
tion correctly or not.
– Level : shows the normalized value of student’s knowledge (between 0 and 1)
on a given concept for a specific section or question
This implementation of the student model allows us to retrieve the students’
level of familiarity with the concepts represented in a section/question after each
interaction of the user with the system.
5 Recommendation Approach
Our system distinguishes two instances for recommending Wikipedia articles:
(1)when student moves to a new section of the book and starts reading and
(2) when students fails to answer a question at the end of the section. These
instances could appear in any order: the students can move to any given section
in the book at any time. Similarly, students can jump right to question section
and start answering the questions before reading previous sections. Students are
also able to return to a section that they previously read or a question that
they already tried. To generate meaningful recommendation that could support
this level of freedom (which is natural for reading a paper or electronic book)
the use of student knowledge level is essential. In the following, we describe the
recommendation approach for both reading and question answering instances.
In order to find the most relevant Wikipedia articles for a given reading
instance, we define two overlapping sets of KCs: (1) the knowledge required
Title Suppressed Due to Excessive Length 7
to fully understand the content of the section (Required Knowledge) and (2)
the current level of student mastery that has been predicted by the student
model (Obtained Knowledge). The “Required Knowledge” for a given section
can be defined by combining all of the concepts associated with the current and
previous sections of the book. This assumption has been made based on the
linear organization used in most textbooks (i.e., all the previous sections in the
textbook are perquisites of the current section).
Having the set of “Required knowledge” for a given section of the textbook
and the set of “Obtained Knowledge” by student while reading that section, we
use set difference to form the “Useful Knowledge” set. The concepts presented
in this set are the ones that are required to understand the section but has not,
or only partially mastered by the student.
Since the student model predicts the level of student knowledge for each
concept as a number between 0 and 1, we consider two conditions for calculating
the importance of each concept in the “Useful Knowledge” set.
– M issingKnowledge: If a concept exists in “Required Knowledge” set but
not in “Obtained Knowledge” set, then its important is equal to 1
– P artialKnowledge: If a concept in the “Required Knowledge” set also exist
in “Obtained Knowledge” set with the predicted value of s, then its impor-
tance is equal to 1-s
As mentioned in section 4.3, we calculated the relevance of each concept
to top 100 Wikipedia articles in our graph. In order to find the most relevant
articles for a reading instance, we multiplied the importance of each concept in
“Useful Knowledge” set by its relevance score to all Wikipedia articles connected
to that concept. Then by aggregating the list for all the concepts presented in
“Useful Knowledge” set over the sum of the final score, we build a ranked list
of Wikipedia articles that are both relevant to a given section and take the
current level of student knowledge into the account. Finally, we select the top 5
ranked articles in the list and present them as recommendation for that reading
instance.
We follow the above approach with a small modifications in recommendations
for question answering instances, which are generated only when the student
failed to answer the question correctly. Main difference is, as mentioned in section
4.2, that the “Required Knowledge” set for a question includes not just concepts
directly associated with the question, but also all “Required Knowledge” in its
corresponding section.
6 The Assessment Process
To assess the potential value of our personalized recommendation approach,
investigated the impact of considering the current level of knowledge represented
in the student-model to generate knowledge-adaptive recommendations. This
sections reviews the details of our evaluation design.
8 B. Rahdari et al.
6.1 Data Source
To assess our recommendation approach in a realistic context, we used log data
collected from the interaction of students with the reading system in a real
semester-long course on Information Retrieval. In this course, the students were
required to read 43 sections of the book and answer questions at the end of each
section (75 questions in total). The log includes data of 22 students who used
the reading system during this course. The students made 9494 interactions with
the system (Average: 431.5, Median: 411.5, SD: 108.2). We followed these inter-
actions reconstruct the state of their student models at every recommendation
opportunity as described in Section 4.4.
6.2 Baseline
To highlight the value of using student knowledge in the recommendation pro-
cess, we compare our knowledge-adaptive recommendation with a baseline that
only considers the content of a given section/question to generate the recom-
mendations. This baseline represents the current state of the art for generat-
ing recommendations of external content [1, 9]. In parallel with adaptive recom-
mendations, we created a set of baseline recommendations for every reading or
question-answering instance.
In order to find the most relevant article with respect to a given section or
question, we first created a list of all articles that are related to the concepts
which represent that section or question. We then aggregated that list over the
sum of the scores for each concept in the list. Finally we re-rank the list based
on the aggregated-sum of scores and selected the top 5 relevant article to each
section or question. This connection is illustrated as ”Related to” relationship
in the graph schema (Figure 3). The “relevance” property of this relationship
represents the relatedness of the section/question to the target Wikipedia article.
7 Results
To determine the effectiveness of our experimental system, we investigated the
following key factors: (1) To what extent the recommended items are affected
by involving the student model into the calculations, (2) Whether including the
student model improved the quality and coverage of the recommended items,
(3) Are there any conspicuous patterns in changes caused by including student
model in the process of recommendations and (4) In what ways the proposed
approach can facilitate the reading process for the students.
7.1 Measure of Ranking Quality - Expected Knowledge Value
In order to compare the results of recommendations between our proposed
method (combination of section/question context and the student model) and
the baseline (only the context of section/question) we calculated the Discounted
Title Suppressed Due to Excessive Length 9
Cumulative Gain (DCG) [8] of every set of recommended items for each instances
of user interaction with the reading system.
i=1
X relevancei
DCG = (1)
n
log2 (i + 1)
The DCG equation, as it shown in 1, takes into account both the relevance
score and the order of items in the recommendations list. The relevance score for
each item is calculated by averaging the similarity score (using Lucene search)
of all the linked concepts to a given section/question and their corresponding
Wikipedia article. This relevance score is being discounted by dividing it with
the log of the corresponding position.
7.2 Overall Expected Knowledge Value of the Recommendations
Figure 4 illustrate the overall quality of recommended items for sections and
questions in the textbook. The x axis shows the normalized average value of
Discounted Cumulative Gain for every given section (Figure 4a) and question
(Figure 4b). As the data shows, the average DCG value is always higher when the
student model is being involved in the process of recommendation. The proposed
approach produced recommendations with in average 23.29% higher DCG value
among all sections and 30.27% among the questions. The higher DCG values
suggests that more concepts were engaged in the process of recommendation
and the recommended items using the proposed approach have higher expected
knowledge values.
δ Baseline Proposed δ Baseline Proposed
100 100
75 75
50 50
25 25
0 0
1.1
1.4
4.2
4.4
4.6
5.2
6.1
6.3
8.1
8.3
8.5
8.7
9.2
12.4
13.2
13.4
13.6
14.2
14.4
14.6
16.2
16.4
261
263
265
267
269
271
273
275
277
279
281
307
309
311
313
315
325
327
329
331
333
335
337
339
341
387
389
391
393
396
398
400
402
404
408
410
412
414
(a) Average DCG values in Sections (b) Average DCG values in Questions
Fig. 4. The Comparison Between Average DCG values of the Sections/Questions of
the Textbook
It would be expected that adaptive recommendations will suggest different
sections to different students at the start of the same section since their knowl-
edge are likely to be different due to differences in reading paths. To investigate
the effect of including the “student model” in the process of recommendations,
10 B. Rahdari et al.
(a) User-Section Heat-Map. (b) User-Question Heat-Map.
Fig. 5. The Comparison Between Average DCG values of the Sections/Questions of
the Textbook
we visualized the difference between expected knowledge values (DCG) amongst
all the students for every section (Figure 5a) and question (Figure 5b) in the
textbook. The apparent fluctuations in expected knowledge value suggests that
every student received a different set of recommendations for a given instance
of interaction with the reading system. This can be considered as the evidence
for the necessity of student-level personalization considering each student has a
different level of mastery for each concept and required to learn divergent set of
concepts at each instance of interaction.
7.3 Predicting User’s Knowledge Requirements
To examine how well the personalized recommendation could help the users, we
examined the “jumping-back” behavior in their reading log. The Reading Mir-
ror system provides students with the ability to jump between sections using the
textbook’s table of content (Figure 1A). Frequently, this functionality is used by
the students to jump back to a certain section of the textbook in order to learn
or refresh their memory on a specific concepts that they need to understand the
current section. Data analysis of student navigation behavior in our class shows,
this jumping-back behavior was quite frequent taking at average 17.27% of all
student navigation steps in the textbook. We believe that in many cases, the
adaptive knowledge-based recommendation of Wikipedia articles could prevent
this unproductive behavior. Unlike non-adaptive Wikipedia article recommenda-
tion (which focuses on the concepts presented in the current page), the adaptive
recommendation attempts to proactively offers information about concepts that
might be necessary to understand the current page or question, but are not
yet known sufficiently by the target user. These recommendation might present
the missing information right in place – eliminating the need of jumping-back
behavior and helping students to integrate the past and the current knowledge.
To assess to what extent the proactively generated adaptive recommendations
could help in this context, we examined each jumping-back case and recorded
all concepts presented on the page that the student selected to jump back. We
then compared this set of concepts with the concepts covered by adaptive rec-
ommendation of Wikipedia articles presented for the given student on the last
Title Suppressed Due to Excessive Length 11
page visited before jumping back. As the Figure 6 shows, the proactive rec-
ommendations cover a remarkable fractions of concepts that were the target of
these back-jumps, 86.63% at average. This result indicates that our adaptive
recommendation approach can accurately predicts the missing knowledge and
considerably reduce the need of jumping through sections to acquire or refresh
these knowledge. In contrast, non-adaptive baseline recommendation focused on
the current page would cover less than a quarter (24.13%) of student background
knowledge needs (Figure 6). This data stresses the importance of considering po-
tentially missing background knowledge in the real classroom context and shows
the value of adaptive knowledge-based recommendation.
Proposed Baseline Average Proposed (86.63) Average Baseline (24.13)
100
75
50
25
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
User ID
Fig. 6. Percentage of potentially missing previous concepts targeted by jumping-back
behavior that are covered by the adaptive and baseline recommendations
8 Summary and Discussion
In this paper we present a novel approach to generate personalized recommenda-
tions of external content for online electronic textbooks. We construct a knowl-
edge graph that represents all three components of “relevant Wikipedia articles”,
“textbook content” and the “student model”. We used this knowledge graph to
generate personalized recommendations based on the relevance to a specific sec-
tion/question but also taking into account the state of the student model in
every instance of interaction with the reading system. The experimental evi-
dence of our data-driven analysis shows that including the student model in the
process of generating the recommendation results in higher expected knowledge
value in the recommendations. Furthermore, we demonstrate that our proposed
approach can accurately predicts the missing knowledge components and poten-
tially reduce the need of jumping-back behaviour amongst students and provides
a smoother reading experience. We are aware of possible limitations of our pro-
posed approach and findings. First, the total number of students in our experi-
mental dataset was limited; including more student data could results in more
accurate conclusion. Second, despite of popularity of data-driven studies in the
12 B. Rahdari et al.
field of recommender system, it has been argued that these studies should be
complemented by controlled user studies where students could observe and rate
generated recommendations. And finally, there are other important components
such as the difficulty of learning concepts, forgetting factor, etc. that we can
potentially incorporate in our recommendation algorithm to produce better and
more accurate results. We hope to address these issues in our future work.
References
1. Agrawal, R., Gollapudi, S., Kannan, A., Kenthapadi, K.: Study Navigator: An
algorithmically generated aid for learning from electronic textbooks. Journal of
Educational Data Mining 6(1), 53–75 (2014)
2. Barria-Pineda, J., Brusilovsky, P., He, D.: Reading mirror: Social navigation and
social comparison for electronic textbooks. In: First Workshop on Intelligent Text-
books at 20th International Conference on Artificial Intelligence in Education. pp.
30–37 (2019), http://ceur-ws.org/Vol-2384/paper03.pdf
3. Brusilovsky, P., Henze, N.: Open corpus adaptive educational hypermedia. In: The
Adaptive Web, pp. 671–696. Springer, Berlin (2007)
4. Bursilovsky, P., Eklund, J., Schwarz, E.: Web-based education for all: A tool for
developing adaptive courseware. In: 7th International World Wide Web Conference.
pp. 291–300 (1998)
5. Carr, L., Hall, W., Bechhofer, S., Goble, C.: Conceptual linking: Ontology-based
open hypermedia. In: 10th International World Wide Web conference. pp. 334–342
(2001)
6. Crestani, F., Landoni, M., Melucci, M.: Appearance and functionality of electronic
books. International Journal on Digital Libraries 6(2), 192–209 (2006)
7. Henze, N., Nejdl, W.: Adaptation in open corpus hypermedia. International Journal
of Artificial Intelligence in Education 12(4), 325–350 (2001)
8. Järvelin, K., Kekäläinen, J.: Cumulated gain-based evaluation of ir techniques.
ACM Transactions on Information Systems 20(4), 422–446 (2002)
9. Kokkodis, M., Kannan, A., Kenthapadi, K.: Assigning educational videos at appro-
priate locations in textbooks. In: Seventh International Conference on Educational
Data Mining. pp. 201–204 (2014)
10. Mayes, J.T., et al.: Strathtutor: The development and evaluation of a learning-by-
browsing on the macintosh. Computers and Education 12, 221–229 (1988)
11. Milne, D., Witten, I.H.: Learning to link with wikipedia. In: 17th ACM Conference
on Information and Knowledge Management. pp. 509–518 (2008)
12. Rihák, J., Pelánek, R.: Measuring similarity of educational items using data on
learners’ performance. In: 10th International Conference on Educational Data Min-
ing (2017)
13. Thaker, K., Carvalho, P., Koedinger, K.: Comprehension factor analysis: Modeling
student’s reading behaviour: Accounting for reading practice in predicting students’
learning in moocs. In: 9th International Conference on Learning Analytics and
Knowledge (2019)
14. Thaker, K., Huang, Y., Brusilovsky, P., He, D.: Dynamic knowledge modeling
with heterogeneous activities for adaptive textbooks. In: the 11th International
Conference on Educational Data Mining. pp. 592–595 (2018)
15. Thaker, K., Zhang, L., He, D., Brusilovsky, P.: Recommending remedial readings
using student’s knowledge state. In: 13th International Conference on Educational
Data Mining. pp. 233–244 (2020)