ArticlePDF Available

Abstract and Figures

Based on the currently developing trend of so called Massive Open Online Courses it is obvious that learning videos are more in use nowadays. This is some kind of comeback because due to the maxim "TV is easy, book is hard" [1][2] videos were used rarely for teaching. A further reason for this rare usage is that it is widely known that a key factor for human learning is a mechanism called selective attention [3][4]. This suggests that managing this attention is from high importance. Such a management could be achieved by providing different forms of interaction and communication in all directions. It has been shown that interaction and communication is crucial for the learning process [6]. Because of these remarks this research study introduces an algorithm which schedules interactions in learning videos and live broadcastings. The algorithm is implemented by a web application and it is based on the concept of a state machine. Finally, the evaluation of the algorithm points out that it is generally working after the improvement of some drawbacks regarding the distribution of interactions in the video.
Content may be subject to copyright.
PaperScheduling Interactions in Learning Videos
Scheduling Interactions in Learning Videos
A State Machine Based Algorithm
https://doi.org/10.3991/ijai.v1i1.10995
Josef Wachtler (*), Martin Ebner
Graz University of Technology, Graz, Austria
josef.wachtler@tugraz.at
AbstractBased on the currently developing trend of so called Massive
Open Online Courses it is obvious that learning videos are more in use nowa-
days. This is some kind of comeback because due to the maxim “TV is easy,
book is hard [1][2] videos were used rarely for teaching. A further reason for
this rare usage is that it is widely known that a key factor for human learning is
a mechanism called selective attention [3][4]. This suggests that managing this
attention is from high importance. Such a management could be achieved by
providing different forms of interaction and communication in all directions. It
has been shown that interaction and communication is crucial for the learning
process [6]. Because of these remarks this research study introduces an algo-
rithm which schedules interactions in learning videos and live broadcastings.
The algorithm is implemented by a web application and it is based on the con-
cept of a state machine. Finally, the evaluation of the algorithm points out that it
is generally working after the improvement of some drawbacks regarding the
distribution of interactions in the video.
KeywordsPlanning, interactivity, learning videos, algorithm, states
1 Introduction
Due to the lately evolved trend of so called MOOCs (short for Massive Open
Online Courses) learning videos are currently giving some kind of comeback [1]. This
means that videos were considered ineffective for learning purposes. Such a statement
was motivated by the commonly accepted maxim “TV is easy and book is hard” [2].
Based on this maxim it is possible to conclude that it is crucial for the learning suc-
cess that a managed attention of the students is achieved.
Furthermore, students are confronted with a growing number of texts, colors,
shapes and figures. Most of them is filtered out centrally by a mechanism known as
selective attention [3][4]. This system is considered as a major influencing factor of
human learning because it enhances both, behavioral and neuronal performance [5].
As a consequence it seems to be obvious that managing this attention is from high
importance.
58
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
In addition, this importance is pointed out by Carr-Chellman and Duchastel [6]
with the observation that interaction and communication are key factors for students
attention. The interaction and the communication should happen in many different
forms and in any directions. On the one hand this means that in addition to face-to-
face communication there should be other means like e-mail, forums, chatrooms or
something similar. On the other hand the communication should not only happen from
the teacher to the students and vice versa, it should also be possible for the students to
interact with the content itself.
The application of interactive components to learning videos also provides differ-
ent possibilities of learning analytics to both, the teachers and researchers [7]. This
means that interactions could be used to evaluate the understanding of the students. In
addition the behavior of the students could be analyzed by evaluating general proper-
ties like the position of the interaction in the video or the reaction times to the interac-
tions.
Based on the presented statements and studies regarding the benefits of interactivi-
ty for the attention of the students and therefore their increased learning success, it
seems to be obvious to apply some kind of interactivity to learning videos. This is
supported by the possibilities of learning analytics with benefits for teachers and re-
searchers. For that a web application is developed which was first introduced by Eb-
ner et al. [8]. Based on this platform some studies were performed. These studies
pointed out that the timing of the interactions in the video is important (see Section 2).
Because of that this research study introduces a scheduling algorithm for interactions
in videos and evaluates its functionalities. Shortly the investigated research question
is: “How to place different forms of interactivity in videos in automatic and planned
ways?”
As mentioned above some related work is shown at first by presenting several stud-
ies about the developed web application. Furthermore, some studies evaluating other
web applications for interactive videos are shown. After that the methodologies are
explained by Section 3 and this is followed by a detailed description of the developed
algorithm (see Section 4). The section named Evaluation shows the performance of
the algorithm at several usages at lectures at Graz University of Technology and after
that its capabilities are discussed (see Sections 5 and 6). Finally, an outlook is given
and the outcomes of the research work are summed up.
2 Related Work
As mentioned above a web application is developed which implements the algo-
rithm presented by this work. This web application was evaluated by several studies.
One study tries to find a basic recommendation of how to place interactions in
learning videos so that the success rates of the results to the questions asked by the
interactions are maximized [9]. For that the web application was used at a large lec-
ture at Graz University of Technology to present the recordings of the classroom ses-
sions to the students. The evaluation was done by asking subjective questions to stu-
iJAI Vol. 1, No. 1, 2019
59
PaperScheduling Interactions in Learning Videos
dents and by analyzing the reaction times to the interactions as well as the results of
the questions. Based on that some recommendations were formed [9]:
“The maximum number of interactions per hour is ten.”
“Content related questions are important for the satisfaction of the students.”
“General questions are useful to support the attention between content related ques-
tions.”
“The space between content related questions should not be too small (at least three
minutes).”
In addition to these recommendations the mentioned study states some hypotheses
about the placement of the interactions in the video. Some of these hypotheses were
further evaluated by another study [10]. The first hypothesis claims that at the begin-
ning of a video, questions are more likely to be answered wrong. This claim was con-
firmed by this follow up study. In contrast to this a second hypothesis was not proven
to be correct. It claims that the interval lengths between the interactions have a corre-
lation to the correctness. It was observed by the first study [9] using longer videos that
the correctness rate is decreasing if the questions are placed too tight. As mentioned
this was not confirmed by the later study [10] using shorter videos.
Because this study took place at a school it was possible to compare the results of
the class which used the interactive videos with a class taught in conventional manner
[10]. By the comparison of the results of the classes it was possible to evaluate the
long-term learning success. It was measured that the first class achieved remarkably
better results than the second.
The web application also implements an attention-profiling algorithm first intro-
duced by Wachtler and Ebner [11]. This algorithm provides the possibility to fully
track the watched parts of the video for each student and furthermore it computes an
attention-level which indicates how attentive each watching student was by evaluating
the reaction times to the interactions. The mentioned interactions are scheduled by the
algorithm presented by this present work.
A further study used this attention-profiling algorithm to monitor the attendance of
students at videos and compared this approach with methods used in a classroom [12].
For a better accuracy of the attendance monitoring the results of some questions pre-
sented by interactions during the video were part of the monitoring. It was pointed out
that this possibility of monitoring the attendance at videos delivers the same accuracy
as different methods used in standard classroom situations.
Another possible usage for interactive components in videos is assessment. This
scenario was addressed by a study performed at a large lecture at Graz University of
Technology [13]. For that exam questions were presented to students during the vid-
eo. It was communicated to the students that these questions will be part of the final
grade and because of that students worked more concentrated and received a better
grade point average than before using the interactive videos. Furthermore, it was ob-
served that the required time for the teacher to evaluate the performance of students
has been optimized by using an approach with interactive videos.
In contrast to these studies using the developed web application which implements
the algorithm introduced by this work, the benefits of interactive videos were also
60
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
pointed out by several other research studies. One of these studies [14] compared the
usage of interactive videos with non-interactive videos for the purpose of learning.
For that a group of 36 students from the University of Offenburg were required to
learn how to tie four nautical knots by watching videos. None of the students had any
previous knowledge on the subject. The students were divided in two groups, one with
interactive videos and one without interactivity. They were permitted to watch the
videos as often as they needed to learn to tie the knots. To measure the benefits of the
interactivity the required time to learn was recorded. It was pointed out that the group
with the interactive videos required significantly less time than the group with the
non-interactive ones. The differences between the required learning time of the two
groups ranges from 66% to 95% of additional time for videos without interactivity.
Furthermore, it was pointed out that the offered methods of interactivity were used
heavily by all students.
A different study [15] evaluated the usage of in-video quizzes at a Machine-
Learning course on Coursera. This course consists of 113 videos with a total length of
19.5 hours. In summary the videos were enriched with 109 in-video quizzes. Out of
the 96,195 registered users 41,643 watched at least one video till its end. The study
pointed out that 74% of the users who got to see a quiz tried to answer it. A common
behavior of the watchers is that they seek forward to the quizzes and after seeing the
question they tend to seek backward before answering it. A likely explanation is that
the students try to find the content which is required to answer the questions. In the
most extreme circumstances students simply jump from quiz to quiz. The correctness
rate of the quizzes is quite high because 76% of the students provided a correct an-
swer at the first attempt. Furthermore, it was observed that the drop-out rate is lower
at the videos with quizzes in comparison to the ones without.
To address the mentioned forward and backward seeking in the video around the
quizzes a further study [16] presents an application which implements a video view-
ing approach driven by questions. The application divides the video in segments and
assigns a question to each segment. Users should navigate through the video by an-
swering the questions. This means that the users are able to skip segments if they are
able to answer the question before watching it. The answered questions and watched
segments are forming a timeline of the history of the students’ progress. This history
enables the students to review the watched content. The application additionally en-
courages the students to rewatch segments which they have not mastered completely.
To evaluate the application 18 participants were recruited. They were required to
watch two videos in two different settings. The first video was presented to them
using the in-video quizzes of Coursera and the second video used the developed ap-
plication. It was concluded that with the developed application students are more
often answering questions and are also more often reviewing the related video seg-
ments in comparison to the standard in-video quizzes. Furthermore, it was pointed out
that the students remembered the correct answers to the questions significantly better
at the question driven video application.
iJAI Vol. 1, No. 1, 2019
61
PaperScheduling Interactions in Learning Videos
3 Methodology
To develop the algorithm two different methods of (software) development are
used. The first one is named “Test Driven Development” (TDD) [17] and the second
one is called “Rapid Prototyping” [19]. These two methods are interwoven because
the new features created with TDD are brought to the test users very fast according to
the principles of Rapid Prototyping.
In general test driven development is based on the strategy that tests for the differ-
ent units of the software are written before the actual source code of it [17]. These
tests are defining the requirements of the software to develop. So, one might assume
that this approach is simply a testing strategy. However a more detailed examination
of the different parts of TDD shows a different picture [18]:
Test: This phase consists of writing tests for each unit or part of the software. To
enable this early writing of tests a clear interface for testing has to be defined. It is
obvious that the tests are run after the source code of the software is finished. It is
not necessary that the tests as well as the source code are written by the same per-
son. Furthermore, the execution of the tests is completely independent from writing
them.
Driven: With TDD the whole software development process is driven forward
because the creation of tests and its implementation in the software leads to the fact
that only the required parts to pass the tests are implemented. After that a refining
of both, tests and source code has to happen. It can be seen that all four general
phases of software development (analysis, design, implementation and testing) are
covered and supported.
Development: TDD alone is no full software development process. It produces
tests and code. Based on this it helps to loop through the mentioned general phases
of software development. However it has to be part of a larger process like Rapid
Prototyping.
Fig. 1. Three steps of the rapid prototyping loop [19][20]
62
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
Rapid Prototyping means that newly developed features are brought to the testing
users or even to productive usage very fast and the feedback is used to improve the
prototype [19][20]. Now the loop can start again. This basic principle can be seen in
Figure 1. In conjunction with TDD this means that the prototype is created by writing
tests that cover its general functionalities and based on this the prototype is imple-
mented. After that the users are working with the prototype and with the feedback.
The tests of TDD are refined and a new iteration of the prototype will keep the loop
running.
As mentioned above TDD and Rapid Prototyping are used to develop the algo-
rithm. This means that the requirements of the algorithm are initially defined by the
unit tests and a prototype is developed according to these tests. This prototype is used
at lectures at Graz University of Technology and the interactions scheduled by the
algorithm at these lectures are analyzed. For that the distribution of interactions for
each student is searched for unwanted effects (see Section 5).
4 Implementation
This section presents the developed scheduling algorithm in a very detailed way.
For a better understanding of its purpose the web application which implements the
algorithm is presented at first (see Section 4.1). After that an overview of the different
components (see Section 4.2) is given and this is followed by the explanation of the
underlying models in Section 4.3. Now the state-machines of the interactions are
shown (see Section 4.4). This is followed by a description of some helper modules by
Section 4.5 and finally all parts of the algorithm are combined together (see Section
4.6).
4.1 The web application
As mentioned above the developed algorithm is implemented by a web application
named LIVE (short for Live Interaction in Virtual learning Environments) [8][21].
This web application is implemented in the programming language Python using the
Django-Web-Framework. It enables the enrichment of videos and live broadcastings
with different types of interactivity. This means that the videos could be interrupted
with questions and watching students are also able to invoke some interactions (e.g.
asking a question to the teacher). In Figure 2 it can be seen how a video is interrupted
by an interaction presenting a multiple choice question. The purpose of these interac-
tions is the support of the students’ attention. So the interactivity is used to help the
students to stay attentive during the whole video [9].
Furthermore, there are a lot of different analysis features for the teacher. For exam-
ple a detailed recording of the joined and watched timespans is available [11][21].
With this it is possible to state for each student when she/he watched which part of the
video (see Figure 3). In addition, a timeline of the watching history (see Figure 4) of
all students is computed.
iJAI Vol. 1, No. 1, 2019
63
PaperScheduling Interactions in Learning Videos
Fig. 2. A video is interrupted by an interaction.
Fig. 3. The watching history of a student shows a timeline of the video and marks the watched
parts with red bars. On mouse-over additional details are printed.
Fig. 4. The diagram prints the timeline of the video on the x-axis and the number of watching
students on the y-axis. The red line represents the number of views along the timeline
and the green line states the number of different users. The vertical red crosshair could
be moved along the timeline to show exact values in the “Details” box.
In summary the main functionalities of LIVE are given by the following list
[8][10][11][21]:
Only available for registered and authenticated users
Different methods of interactivity
Automatically asked questions and CAPTCHAs (short for Completely Auto-
mated Public Turing test to tell Computers and Humans Apart)
64
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
Asking questions to the teacher
Asking text-based and multiple-choice questions to the students
Students can set their level of attention
Error reporting to the teacher
Analysis features
Evaluations based on an attention-profiling-algorithm
Evaluation of the answers of the students to the different questions
All of the mentioned forms of interactivity are encapsulated in so called "interac-
tion methods" [22]. Each interaction method is completely independent from the other
interaction methods and provides all functionalities of a single form of interactivity
(e.g. asking text-based questions to the students). This includes the creation, the show-
ing and the analysis of this single interaction method. From a technical point of view
such an interaction method is a plugin which implements an API provided by the web
application.
Each interaction method is of a given type. This interaction type defines the meth-
od used for scheduling the interactions as well as the primary targeted audience (stu-
dents or teacher). Currently there are four types [22]:
Automatic: Interaction methods of this type are called automatically and in a ran-
dom way a given number of times per hour. This means that neither the teacher nor
the students are required to trigger an interaction.
Student Triggered: At this interaction type the student has to start the interaction
by himself. For that there are control elements at the right hand side of the video.
Teacher Triggered: While a live broadcasting is running the teacher could start
interactions of this type. This is done by using similar control elements.
Planned: The interactions of this type are shown at planned positions in the video.
It seems to be obvious that in this case the teacher has to create and plan the inter-
actions before the video is released.
4.2 Overview
The algorithm to schedule interactions in videos consists of several components. In
Figure 5 these components are shown.
Fig. 5. The components of the algorithm
iJAI Vol. 1, No. 1, 2019
65
PaperScheduling Interactions in Learning Videos
The underlaying part is the data storage which is a relational database represented
by models. It is responsible for saving all the data required to handle the scheduling of
the interactions. The models are accessed by the state machines of the interactions
during the video. This indicates that the state machines control the main functionali-
ties of the algorithm because depending on the state interactions are shown or discard-
ed. To support the state machines there is the interaction planner and the interaction
loader. The first one creates a plan of interactions for each student at her/his first join-
ing to the video. For that the planner has to access the models. The second supporting
part is the loader and as indicated by its name it is responsible for showing the interac-
tions to the students when they are scheduled.
4.3 The models
All the data of the algorithm is stored by the models in a database. Figure 6 shows
the used models of the algorithm in a simplified version. As mentioned above the
interactions are grouped in interaction methods of a given type. This grouping is rep-
resented by the models InteractionMethod and InteractionType. The first one holds
some metadata of the interaction method and defines whether the video should be
paused if an interaction occurs by the attribute pause_on. The second is responsible
for indicating the type of the interaction method.
Fig. 6. The models of the algorithm in a simplified version
An interaction is represented by the base model Interaction. Because an interaction
could be shown to the students or the teacher there are two sub-models. The first is
66
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
named Student Interaction and it is used for interactions shown to students and the
second (Teacher Interaction) represents interactions displayed to the teacher at live
broadcastings only. For simplifying reasons the following description of the algorithm
is focused on students at videos. This is done because the basic working principle is
the same at live broadcastings and for teachers.
The base model saves the name of the view which actually displays the interaction
represented by the model. Furthermore, it holds the scheduled starting point of the
interaction in the form of a relative position in the video. In addition to these values
the sub-model for interactions shown to students holds the user to which the interac-
tion is shown. In a similar way the sub-model for the teacher interactions saves the
event of which a teacher is available to receive the interaction.
If the interaction method is of type 4 (planned interactions) there has to be a mech-
anism to plan the interactions. For that there is the model called Planned Student In-
teraction which is responsible for saving these planned interactions. It saves the name
of the view which is called to display the interaction and the scheduled time of start in
a very similar way like the normal interactions. Furthermore, it saves the event to
which it belongs and the mode (video or live broadcasting) of the event. Because the
PlannedStudentInteraction is translated to a standard Student Interaction for each
watching student (see below) there is also a list of these created interactions.
Every time an interaction is shown to a student or to a teacher a new instance of a
CallHistory is created. It saves different values of date and time indicating when the
interaction is shown, answered or simply closed. In detail the different fields are used
in the following way:
Opened_at: Saves the date and time of the creation of the Call History. This is done
by the interaction planner, by the students or the teacher (see below).
Closed_at: Holds the date and time of closing the Call History. This happens when
the interaction is never displayed.
Real_start_relative: position in the video of the displaying of the interaction
Real_start_absolute: date and time of the displaying of the interaction
Response_relative: position in the video of the user’s response
Response_absolute: date and time of the user’s response
To manipulate these values the offered methods of the Interaction model are used
and they are mostly delegating to the corresponding methods of the Call History after
some security checks. The state machine presented by the following section operates
mainly on these values.
4.4 The state machine
As indicated above the state machine operates on the interactions. This means that
each interaction has a state and based on that state certain actions are performed. Fig-
ure 7 visualizes the states and the transitions. When a student joins a video the first
time all interactions are created for her/him in the initial state named “open” with a
new CallHistory. This is done by the interaction planner (see next section). During the
iJAI Vol. 1, No. 1, 2019
67
PaperScheduling Interactions in Learning Videos
video runs the current position is monitored. If the scheduled starting position of the
interaction is reached the state switches to “pending”. Now there are two possibilities:
The first one is that the interaction loader (see section below) brings the interaction to
its displaying queue but the interaction is never displayed and because of that the state
switches to “closed”. This is the case if there is another interaction currently on dis-
play in its waiting state. The second possibility is that the loader displays the interac-
tion and this leads to a state switch to “waiting”. Now the interaction is waiting for a
response from the student. In the case of a response the interaction switches to the
state “done” and in the other case if there is no response it switches to “not_done”.
Finally, it could happen that the video is watched more often than once and this leads
again to the interaction planner which is creating a new Call History. This causes a
reset of the state of all interactions to their initial state “open”.
Fig. 7. The different states of an interaction and its transitions
As indicated above the state is defined by the values of the members of the current
Call History. Table 1 shows this relation of each state to these values. It can be seen
that each state has a different distribution of set (x) fields.
Table 1. The model fields are defining the states.
opened_at
closed_at
real_start_
relative
real_start_
absolute
response_
relative
response_
absolute
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
68
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
4.5 Planning and loading interactions
As mentioned above there are two helper modules for the state machine, namely
the interaction planner and the interaction loader. The first one is responsible for
planning the interactions for each student at their first joining to the video. The second
one displays the interaction to the students and handles the response.
The interaction planner is responsible for creating a plan of interactions for each
student on her/his first joining to the video. In general this planning happens by creat-
ing instances of the Student Interaction model. Depending on the type of the interac-
tion method they are created in a different way.
If the type of the interaction method is “1” (automatic) the interactions of this
method are created in a random way. For that each interaction method of this type has
to provide a parameter stating how often its interactions should be displayed per hour.
With this value a slot length in seconds is computed by dividing the seconds of an
hour (3600) by the given value of the parameter. Now an interaction is created for
every slot which fits into the length of the video. In each slot the interaction is placed
randomly however there is the following restriction. To avoid that interactions are
placed too fast after each other at the borders of the slots the first as well as the last
ten percent are not used for the interaction of the slot.
The interactions of methods of the types “2” (student triggered) and “3” (teacher
triggered) are not created by the interaction planner. This is the case because these
interactions are created on an action triggered by the students or by the teacher on the
fly during the video runs. This creation has to be done by the interaction methods
themselves.
Finally, the type “4” (planned) is covered by the interaction planner. As shown
above there is an instance of Planned Student Interaction for each planned interaction.
Because these instances are not related to a student the interaction planner creates a
Student Interaction for each of them and links them to their corresponding Planned
Student Interaction. These interactions are created by the interaction planner before
the automatic ones. This is done because the interaction planner tries to leave a space
(+/- 10% of the slot length) of no other interactivity around each planned interaction.
A further duty of the interaction planner is to reset the state machine of each inter-
action on re-joining of a student to the video. This means that a new Call History is
created for the interactions of the re-joining student.
Below the complete interaction planner is shown in pseudo code. It can be seen
that the creation of interactions happens only at the first joining of a user. The first
for-block is responsible for creating the planned interactions and the second one han-
dles the creation of the automatic interactions.
if users_first_join() {
for pi in set(PlannedInteraction) {
interaction = create StudentInteraction from pi
pi.interactions.add(interaction)
}
for each interaction_method of type automatic {
iJAI Vol. 1, No. 1, 2019
69
PaperScheduling Interactions in Learning Videos
slot_length = 3600 / interac-
tion_method.interactions_per_hour
slots_per_video = video_length / slot_length
for each slot in slots_per_video {
do {
pos = random(slot.begin + 10% of slot_length,
slot.end - 10% of slot_length)
}
while pos is not around a pi
automatic_interaction = StudentInteraction at pos
}
}
}
for each StudentInteraction {
create a new CallHistory
}
During the video runs the interaction loader monitors the position in the video. It
reports this position to the state machines of the interactions so that the state machines
can switch the state. If an interaction switches to the state “pending” the interaction
loader tries to show the interaction to the student by displaying its callback view.
Furthermore, the video is paused if the interaction is part of an interaction method
requesting this pausing. Once the interaction is shown the state switches to “waiting”.
In the case that another interaction is on display the pending interaction is not shown
and remains pending until the other interaction is finished. If the interaction is current-
ly displayed and therefore in its waiting state the interaction loader waits for a re-
sponse from the student. If this response finally comes the state machine is informed
and because of that, it switches to one of its end-states.
4.6 The algorithm
Now all parts of the algorithm are combined. In Figure 8 it is shown how the algo-
rithm operates with the different components. At first it can be seen that the teacher
has to create the interactions which are planned. Strictly this is not part of the algo-
rithm but it is required for better understanding.
When a student joins a video the algorithm starts with the interaction planner. As
explained above it creates the interactions with the state machines. For that it inde-
pendently plans the automatic interactions and uses the templates (Planned Student
Interaction) created by the teacher for the planned interactions. Now the interaction
loader takes over and feeds the state machines with the current position so that the
state switches can happen.
70
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
Fig. 8. The flow of events of the algorithm
From a users point of view the teacher has to create the planned interactions
through forms offered by the interaction methods. After the teacher has released the
video to students, they are able to watch it. When they are joining the video for the
first time, the interaction planner builds the plan of interactions for each student.
While the students are watching, interactions will be displayed to them. In the case of
a live broadcasting interactions are also displayed to the teacher. This is handled by
the interaction loader in both cases.
5 Evaluation
According to the described methodology tests are written for the different units or
parts of the algorithm. If all of the tests are passed successfully the prototype is used
in a productive environment. With the gathered results and observations from the
productive usage the tests are refined to match new or adapted requirements. This
indicates that the correct functionality from a technical point of view is validated by
the unit tests and the practical aspects are evaluated by analyzing the results of test
usages as well as by the feedback from the users.
The algorithm presented by the previous section is explained in its current version.
The following steps of evaluation are performed on earlier versions and the revisions
are leading to this current one.
5.1 Setting
The created prototypes were used at different lectures and projects at Graz Univer-
sity of Technology. This included the presentation of recordings of lectures held in
iJAI Vol. 1, No. 1, 2019
71
PaperScheduling Interactions in Learning Videos
classrooms to the students as well as common learning videos. In summary 950
watching students and 93 videos were part of the evaluation. In combination this led
to a total number of 7531 different distributions of interactions. To evaluate the distri-
bution of the interactions the following values of the interactions were recorded and
analyzed for each student:
Date and time
Position in the video
Timespan since last interaction
5.2 Automatic and random interactions
With the recorded values for each distribution of interactions it is possible to count
the number of distributions with unwanted effects. These unwanted effects are mostly
defined by feedback of students and could be summarized to the following constrains:
The interval between interactions should be 10% of the video length or higher.
However, the same interval should not be longer than 35% of the video length.
To fulfill these constrains it took three iterations of the Rapid Prototyping loop. Af-
ter each productive usage of the algorithm the distributions of interactions were
searched for these unwanted effects. The results could be seen in Table 2.
Table 2. The results of the analysis of the distributions of interactions.
Number of
distributions
Interval length < 10%
of video length
Interval length > 35%
of video length
2167
520 (23.99%)
628 (28.98%)
3514
457 (13.01%)
492 (14.00%)
1850
0
0
In the first version of the interaction planner, the scheduling of the automatic and
random interactions (type 1) was implemented without the explained slots. The inter-
actions were randomly distributed through the length of the video without any re-
strictions. This requirement was fully covered by the unit tests and after the passed
successfully the algorithm was used in a productive environment.
It was observed that with the completely random approach an uneven distribution
of interactions could occur [9]. This means that in 29% of the distributions longer
phases of no interactivity happened. Furthermore, it is clear that the same effect lead
to an agglomeration of interactions in 24% of the cases. These results gathered by
analyzing the recorded data were also reported by the watching students. Most of
them claimed that an even distribution of interactions through the video is favored.
These claims led to the creation of the constrains mentioned above.
Because of that the requirements were refined. With these new requirements the
slot mechanism presented above was introduced. Again the prototype was used at a
lecture and the distribution of the interactions was analyzed. Now a much better dis-
tribution was observed but in 13% of the cases the interactions were placed too nar-
72
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
row. It seems to be obvious that this could happen at the borders of the slots. If in one
slot the interaction was placed very near to the end of the slot and at the proceeding
slot at the beginning. The same effect led to longer phases of no interactivity in 14%
of the cases. This led again to a refinement of the requirements and the tests which
were used to develop the current version of the algorithm. Now the constrains are
fulfilled by all distributions. Figure 9 visualizes the differences of the three versions
by showing three example distributions of interactions. The points are representing
the interactions. It can be seen that with each version the distribution becomes more
even.
Fig. 9. The different versions of the algorithm schedule the interactions differently.
5.3 Planned nteractions
The first version of the interaction planner created the random interactions before
the planned ones. However the usage at a lecture pointed out that it could happen that
random interactions are occurring at the same time or very close to planned interac-
tions. Because of that the unit tests were refined. This lead to a change of order and
now the planned interactions are created at first. This enables the interaction planner
not to place a random interaction too close to a planned interaction.
6 Discussion
As stated above the main drawbacks of the earlier versions of the algorithm are re-
lated to the distribution of the interactions along the timeline of the video. It has been
shown that with an uncontrolled random distribution it could happen for some stu-
dents that the interactions are placed to dense. As a consequence this leads to longer
phases of no interactivity. This behavior is not appreciated because of two reasons. At
first the goal of interactive components is the support of the attention of the students
while watching a video. As shown by other studies [9][10] a regular occurrence of
interactions is important. The second reason is the feedback of the students [9]. They
are also more satisfied with an even and regular distribution of interactions. These
iJAI Vol. 1, No. 1, 2019
73
PaperScheduling Interactions in Learning Videos
mentioned reasons were confirmed by using the algorithm at lectures to show the
recordings of the lecture held in class. This means that these recordings are of a con-
siderable length (one hour or more). In comparison to that the usage of the algorithm
for shorter videos (10 to 20 minutes) points out that the distribution is not directly
related to the success rate of the students [10]. A reason for the discovered differences
regarding the distribution of interactions in videos of different lengths might be that in
shorter videos the frequency of interactions is typically higher and because of that a
more dense setting of interactive components is more accepted. It can be seen that the
remarks about the connection between the distribution of the interactive components
and the success rate of the students were gathered in the context of different studies.
For that the possibilities of learning analytics provided by the application of the inter-
actions to videos were used [7].
7 Outlook
The discovered discrepancy between the longer and shorter videos regarding the
distribution of interactions implies that further research is required. It is planned that
recordings of lectures should be presented to students in different ways. Each record-
ing should be presented to one group of students in its complete and therefore long
version. A second group should watch the recording in smaller parts. The distribution
of interactions in the videos of the two groups should be the same. With that it should
be possible to gain further insight in the relevance of the distribution of interactions at
videos of different lengths.
8 Conclusion
With this work an algorithm to schedule interactive components in learning videos
as well as live broadcastings is introduced. The application of interactivity to videos
seems to be important because videos were not well accepted for learning purposes
[2]. However, learning videos are more in use today [1] due to the increased offering
of MOOCs. To support the attention of the watching students providing interactive
components in different forms is a widely accepted tool [6].
To address this issue a web application first introduced by Ebner et al. [8] is devel-
oped. This web application adds different methods of interactions to videos and live
broadcastings. For that the scheduling algorithm presented by this work is used. This
algorithm consists of several parts which are explained by Section 4. The basic prin-
ciple of the algorithm is a state machine for each individual interaction. Depending on
its state an interaction is handled differently. This means that such an interaction
could be waiting to be displayed, it could be currently visible or it could be already
finished.
The algorithm is developed and evaluated using the mechanisms of “Test Driven
Development” and “Rapid Prototyping”. This indicates that TDD is used to define the
requirements and to validate the correct implementation of these requirements. To
evaluate the requirements a prototype which is passing the unit tests of TDD is used in
74
http://www.i-jai.org
PaperScheduling Interactions in Learning Videos
a productive environment. With the experiences and gathered results of the productive
usage the requirements are refined and the circle starts again. The evaluation of the
algorithm according to this explained principle pointed out that after some iterations
the scheduling produces satisfying distributions of the interactions. With a working
algorithm the initial research question (see Section 1) is finally answered.
9 References
[1] H. Khalil and M. Ebner, “Interaction possibilities in moocshow do they actually happen,”
in International Conference on Higher Education Development, 2013, pp. 124.
[2] G. Salomon, “Television is easy and print is tough: The differential investment of mental
effort in learning as a function of perceptions and attributions.” Journal of educational psy-
chology, vol. 76, no. 4, p. 647, 1984. https://doi.org/10.1037//0022-0663.76.4.647
[3] J. Moran and R. Desimone, “Selective attention gates visual processing in the extrastriate
cortex,” Science, vol. 229, pp. 782784, Aug. 1985. https://doi.org/10.1126/science.40237
13
[4] R. M. Shiffrin and G. T. Gardner, “Visual processing capacity and attentional control,”
Journal of Experimental Psychology, vol. 93, no. 1, pp. 7282, 4 1972. https://doi.org/10.
1037/h0032453
[5] H. Spitzer, R. Desimone, and J. Moran, “Increased attention enhances both behavioral and
neuronal performance,” Science, vol. 240, pp. 338340, Apr. 1988. https://doi.org/10.11
26/science.3353728
[6] A. Carr-Chellman and P. Duchastel, “The ideal online course,” British Journal of Educa-
tional Technology, vol. 31, no. 3, pp. 229241, 2000. [Online]. Available: https://doi.org/
10.1111/1467-8535.00154
[7] J. Wachtler, M. Khalil, B. Taraghi, and M. Ebner, “On using learning analytics to track the
activity of interactive mooc videos,” 2016.
[8] M. Ebner, J. Wachtler, and A. Holzinger, “Introducing an information system for success-
ful support of selective attention in online courses,” in Universal Access in Human-
Computer Interaction. Applications and Services for Quality of Life. Springer, 2013, pp.
153162. https://doi.org/10.1007/978-3-642-39194-1_18
[9] J. Wachtler and M. Ebner, “Impacts of interactions in learning-videos: A subjective and
objective analysis,” in EdMedia: World Conference on Educational Media and Technolo-
gy, vol. 2015, no. 1, 2015, pp. 16421650.
[10] J. Wachtler, M. Hubmann, H. Z¨ohrer, and M. Ebner, “An analysis of the use and effect of
questions in interactive learning-videos,” Smart Learning Environments, vol. 3, no. 1, p.
13, 2016. https://doi.org/10.1186/s40561-016-0033-3
[11] J. Wachtler and M. Ebner, “Attention profiling algorithm for video-based lectures,” in
Learning and Collaboration Technologies. Designing and Developing Novel Learning Ex-
periences. Springer, 2014, pp. 358367. https://doi.org/10.1007/978-3-319-07482-5_34
[12] J. Wachtler and M. Ebner, “On using interactivity to monitor the attendance of students at
learning-videos,” in EdMedia: World Conference on Educational Media and Technology.
Association for the Advancement of Computing in Education (AACE), 2017, pp. 356366.
[13] J. Wachtler, M. Scherz, and M. Ebner, “Increasing learning efficiency and quality of stu-
dents homework by attendance monitoring and polls at interactive learning videos,” in
EdMedia+ Innovate Learning. Association for the Advancement of Computing in Educa-
tion (AACE), 2018, pp. 13571367.
iJAI Vol. 1, No. 1, 2019
75
PaperScheduling Interactions in Learning Videos
[14] S. Schwan and R. Riempp, “The cognitive benefits of interactive videos: learning to tie
nautical knots,” Learning and instruction, vol. 14, no. 3, pp. 293305, 2004. https://doi.org/
10.1016/j.learninstruc.2004.06.005
[15] G. Kovacs, “Effects of in-video quizzes on mooc lecture viewing,” in Proceedings of the
third (2016) ACM conference on Learning@ Scale. ACM, 2016, pp. 3140. https://doi.
org/10.1145/2876034.2876041
[16] G. Kovacs, “Quizcram: A question-driven video studying interface,” in Proceedings of the
33rd Annual ACM Conference Extended Abstracts on Human Factors in Computing Sys-
tems. ACM, 2015, pp. 133138. https://doi.org/10.1145/2702613.2726966
[17] K. Beck, “Test-driven development: by example”. Addison-Wesley Professional, 2003.
[18] D. Janzen and H. Saiedian, “Test-driven development concepts, taxonomy, and future di-
rection,” Computer, vol. 38, no. 9, pp. 4350, 2005. https://doi.org/10.1109/mc.2005.314
[19] X. Yan and P. Gu, “A review of rapid prototyping technologies and systems,” Computer-
Aided Design, vol. 28, no. 4, pp. 307318, 1996. https://doi.org/10.1016/0010-
4485(95)00035-6
[20] P. F. Jacobs, “Rapid prototyping & manufacturing: fundamentals of stereolithography”.
Society of Manufacturing Engineers, 1992.
[21] J. Wachtler and M. Ebner, “Support of video-based lectures with interactions-
implementation of a first prototype,” in World Conference on Educational Multimedia,
Hypermedia and Telecommunications, vol. 2014, no. 1, 2014, pp. 582591.
[22] J. Wachtler, “LIVE documentation,” online, 2018, https://josefwachtler.wordpress.com.
10 Authors
Josef Wachtler is currently doing his PhD at Graz University of Technology in the
area of interactive learning videos. This means that his research is focused on the
development and the evaluation of interactive video tools for different purposes. He is
working at the Department Educational Technology at Graz University of Technology
and is responsible for developing different kinds of learning applications and tools.
For more information like further publications, please visit his blog:
https://josefwachtler.wordpress.com/
Martin Ebner is currently head of the Department Educational Technology at
Graz University of Technology and therefore responsible for all university wide e-
learning activities. He holds an Adjunct Prof. on media informatics (research area:
educational technology) and works also at the Institute for Interactive Systems and
Data Science as senior researcher. His research focuses strongly on seamless learning,
learning analytics, open educational resources, maker education and computer science
for children. Martin has given a number of lectures in this area as well as workshops
and keynotes at international conferences. For publications as well as further research
activities, please visit his website: http://martinebner.at
Article submitted 2019-06-06. Resubmitted 2019-07-15. Final acceptance 2019-07-20. Final version
published as submitted by the authors.
76
http://www.i-jai.org
... In the fifth article by Josef Wachtler and Martin Ebner [5], the authors present an algorithm which schedules interactions in learning videos and live broadcastings based on the concept of a state machine. More specifically, they propose an adaptive and interactive video platform that also support the integration of learning analytics which can help us better understand how students interact with videos. ...
Article
Full-text available
In this editorial, the first issue of the International Journal of Learning Analytics and Artificial Intelligence for Education is presented. The Journal of Learning Analytics and Artificial Intelligence for Education is a peer-reviewed, open-access journal that aims to disseminate the highest quality research in the field. The journal aims to increase knowledge and understanding of ways in which learning analytics and artificial intelligence can support and enhance education. The editorial presents the scope and fields of interest for the journal and an overview of the articles published in the first issue.
Conference Paper
Full-text available
Due to the fact that students are confronted with a growing amount of texts, colours, figures and shapes and due to their ability to process only a limited number of such information simultaneously, it seems to be obvious that efforts should be made to increase the students' attention-levels. This is important because research results have indicated that selective attention is considered as the most valuable resource in the process of human learning. The application of interaction and communication to the process of learning is a useful strategy to direct the students' attention. It seems to be obvious that this is also true for learning videos. Therefore, this work contains a description of how a video platform with interactive components can be used to support the students and teacher. The video platform is explained and evaluated by analysing its usage in a large teaching course at an institution of higher education. The application of this strategy improved the students' performance and optimized the teacher's workload.
Conference Paper
Full-text available
Many studies are claiming that compulsory attendance has its benefits for students because it is positively influencing their performance in consequence. This means that fully attending students are receiving a better grade than those who skip parts of the course. Whereas there are existing technologies to monitor the attendance of students in standard classroom situations the possibilities to provide such functionalities at online videos are limited. Because of that this work introduces an approach to assess the performance of students at videos. For that a web-platform which offers a detailed recording of the watched parts of videos is presented. The evaluation shows the performance of the students at a course consisting of videos with compulsory attendance and furthermore it points out that the approach for monitoring the attendance at online videos is basically working.
Article
Full-text available
This study focuses on the positioning of interactive questions within learning videos. It is attempted to show that the position of a question’s occurrence has an impact on the correctness rate of its answer and the learning success. As part of the study, the interactive learning videos in which the questions are placed are used as teaching materials with a class. The pupils have been working with the videos for around one month and some interesting results could be obtained. It is shown that questions which are asked too early in the videos are answered incorrectly more often than other questions. This manuscript also recommends an adequate positioning of the first question in learning videos. The new hypothesis that the length of intervals between popping up questions plays a minor role at rather short learning videos is constructed in this publication. Moreover, the positive impact on the long-term learning success of the participants of learning videos is determined.
Conference Paper
Full-text available
It is widely known that interaction, as well as communication, are very important parts of successful online courses. These features are considered crucial because they help to improve students’ attention in a very significant way. In this publication, the authors present an innovative application, which adds different forms of interactivity to learning videos within MOOCs such as multiple-choice questions or the possibility to communicate with the teacher. Furthermore, Learning Analytics using exploratory examination and visualizations have been applied to unveil learners’ patterns and behaviors as well as investigate the effectiveness of the application. Based upon the quantitative and qualitative observations, our study determined common practices behind dropping out using videos indicator and suggested enhancements to increase the performance of the application as well as learners’ attention.
Conference Paper
Full-text available
QuizCram is a question-focused format for navigating and reviewing lecture videos. QuizCram shows users a question to answer, with an associated video segment. Users navigate through the video segments by answering questions. We also allow users to review using a timeline of previously answered questions and videos. To encourage users to review questions, QuizCram keeps track of their question-answering and video-watching history and recommends users to review questions they have not fully mastered. QuizCram-format courses can be generated automatically from lectures with in-video quizzes, though the format is flexible enough to accommodate multiple questions per video segment. Our user study comparing QuizCram to in-video quizzes found that users practice answering and reviewing questions more when using QuizCram, and are better able to remember answers to questions they encountered.
Conference Paper
Full-text available
It is clear that a system known as selective attention is the most crucial resource for human learning (Heinze et al. 1994). Due to this fact a web based information system is developed to support the attention of the watchers of a learning video. This is done by enriching the video with different forms of interactions. Among others there are interactions presenting multiple-choice questions at predefined positions in the video as well as randomly occurring interactions displaying general questions, which are not content related. To gain a basic plan which points out how to distribute the interactions over the video the usage of the information system at a lecture is analyzed. For that on the one hand some feedback of the users is provided and on the other hand an objective analysis of the results of the multiple-choice questions is done by evaluating their time of occurrence in the video.
Conference Paper
Full-text available
In general videos have a more or less consuming character without any interaction possibilities. Due to this fact a web-based application is developed which offers different methods of communication and interaction to a certain learning-video. This should help attendees to avoid that they become tired and annoyed. The lecturer is able to use interactions to omit whether learners are able to understand the content of the video. For instance the developed web-application offers the possibility to add multiple-choice-questions at predefined positions. Furthermore there are many different kind of analysis as for example a detailed attention-profile. In this publication the implementation of a first prototype is described as well as a first field study which points out that learners? interactions and engagements increased arbitrarily.
Conference Paper
Full-text available
Due to the fact that students' attention is the most crucial resource in a high-quality course it is from high importance to control and analyze it. This could be done by using the interaction and the communication because they are known as valuable influencing factors of the attention. In this publication we introduce a web-based information system which implements an attention-profiling algorithm for learningvideos as well as live-broadcastings of lectures. For that different methods of interaction are offered and analyzed. The evaluation points out that the attention profiling algorithm delivers realistic values.
Conference Paper
Full-text available
Human learning processes are strongly depending on attention of each single learner. Due to this fact any measurement helping to increase students' attention is from high importance. Till now there are some developments called Audience-Response-Systems only available for face-to-face education even for masses. In this publication we introduce a web-based information system which is also usable for online-systems. Students' attention will be conserved based on different interaction forms during the live stream of a lecture. The evaluation pointed out that the system helps to enlarge the attention of each single participant.
Conference Paper
Online courses on sites such as Coursera use quizzes embedded inside lecture videos (in-video quizzes) to help learners test their understanding of the video. This paper analyzes how users interact with in-video quizzes, and how in-video quizzes influence users' lecture viewing behavior. We analyze the viewing logs of users who took the Machine Learning course on Coursera. Users engage heavily with in-video quizzes -- 74% of viewers who start watching a video will attempt its corresponding in-video quiz. We observe spikes in seek activity surrounding in-video quizzes, particularly seeks from the in-video quiz to the preceding section. We show that this is likely due to users reviewing the preceding section to help them answer the quiz, as the majority of users who seek backwards from in-video quizzes have not yet submitted a correct answer, but will later attempt the quiz. Some users appear to use quiz-oriented navigation strategies, such as seeking directly from the start of the video to in-video quizzes, or skipping from one quiz to the next. We discuss implications of our findings on the design of lecture-viewing platforms.