Axa Assurance Maroc - Insurer Innovation Award 2024
Attention Profiling Algorithm for Video-based Lectures
1. S C I E N C E P A S S I O N T E C H N O L O G Y
www.tugraz.at
Attention Profiling Algorithm
for Video-based Lectures
Josef Wachtler, Martin Ebner and Behnam Taraghi
ZID - Social Learning - TU Graz
HCII 2014
2. 2
Attention Profiling Algorithm for Video-based Lectures
Content
1. Motivation
2. Implementation of the Algorithm
– Operating-Context
– Recording Joined Timespans
– Calculating the Attention-Level
3. Evaluation
4. Conclusion
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
3. 3
Attention Profiling Algorithm for Video-based Lectures
Graz, University of Technology
Europe, Austria, Graz
http://www.tugraz.at
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
4. 4
Motivation
Students’ Attention
students are confronted with a growing quantity of
information
they can handle and process only a limited number of
these information at the same time
selective attention is the most crucial resource for
human learning
so it is from high importance to control and analyze it
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
5. 5
Motivation
Interaction and Communication
should be used in many different forms as well as in
all possible directions
avoid that learners become tired or annoyed
increase the attention and the contribution
feedback for teachers:
Is it possible for the learners to follow the content?
Is the speed appropriate?
...
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
6. 6
Implementation of the Algorithm
Overview
the attention profiling algorithm is divided in two parts:
a detailed recording of the joined timespans of
each single user
the calculation of an attention-level based on the
reaction-times to the interactions
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
7. 7
Operating-Context
Web-Application
on-demand video or live-broadcasting
implements the attention profiling algorithm
different methods of interaction:
automatically asked questions and captchas
asking questions to the lecturer
asking text-based questions to the attendees
multiple-choice questions at pre-defined positions
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
9. 9
Recording Joined Timespans
Functionalities
for each attendee it is possible to say
at which time he/she
watched which part of the video
calculating statistical values
the shortest or the longest joined timespan
the average length of the joined timespans
...
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
10. 10
Recording Joined Timespans
Models
the JoinedUser-model connects an user to an event
the History-model represents a joined timespan with
both, absolute and relative timestamps
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
11. 11
Calculating the Attention-Level
Functionalities
calculation of an attention-level which is based on the
reaction-times of the attendees to the interactions
1. logging the reaction-times
2. calculating the attention-level
maxim: if the attendee reacts slower the
attention-level decreases
result ranges from 0% to 100%
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
12. 12
Calculating the Attention-Level
Logging the Reaction-Times
the Interaction-model
with its concrete sub-class models for each
possible receiver of an interaction
connects an interaction to an user
the CallHistory-model
logs every occurrence of an interaction
in absolute and relative timestamps
the difference between the real start time and the
response time is equal to the reaction-time
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
14. 14
Calculating the Attention-Level
Overview
the calculation is split in three rounds:
1. calculation of an attention-level based on the
reaction-times for every call of an interaction (I)
2. grouping them to attention-levels (AL) of each
interaction-methods (IM)
3. generalizing to an attention-level of a joined timespan
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
15. 15
Calculating the Attention-Level
Round 1
the calculation has two parameters
1. SUCCESS UNTIL states the time until an
attention-level of 100% could be reached
2. FAILED AFTER indicates after which
reaction-time an attention-level of 0% will be
assumed
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
16. 16
Calculating the Attention-Level
Round 1
f(tij) represents the attention-level of the j-th
interaction of the i-th interaction-method
tij is the corresponding reaction-time
f(tij ) =
100 if tij ≤ SUCCESS UNTIL
0 if tij > FAILED AFTER
g(tij ) else
(1)
Where g(tij) is
g(tij ) = 100 −
tij − SUCCESS UNTIL
FAILED AFTER − SUCCESS UNTIL
∗ 100 (2)
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
17. 17
Calculating the Attention-Level
Round 2
ai calculates the attention-level of the i-th
interaction-method by forming the mean
mi is the number of its interactions
ai =
mi
j=0
f(tij)
mi
(3)
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
18. 18
Calculating the Attention-Level
Round 3
takes the attention-level of each interaction-method
(ai) and again forms the mean over them
n is the number of interaction-methods
attention =
n
i=0
ai
n
(4)
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
19. 19
Evaluation
Overview
three goals
1. gaining suitable parameters to force the algorithm
to deliver realistic values
2. comparing the results of the algorithm with the
feedback of the attendees to implement adoptions
3. evaluating the effects of the adoptions
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
20. 20
Evaluation
Gaining Suitable Parameters
live-broadcasting of the lecture Societal Aspects of
Information Technology
analyzing recorded reaction-times of the interactions
the average reaction-time is calculated to place the
parameters around this point
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
21. 21
Evaluation
Compare Results with Feedback
live-broadcasting of the lecture Introduction to
Structured Programming
complete number of attendees vs. active ones
active: watched ≥ 75% and attention-level ≥ 50%
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
23. 23
Evaluation
Feedback
attendees felt uncomfortable with their attention-level
- they assumed a much higher one
impossible to answer faster because the live-stream
does not stop if an interaction occurs
the number of interactions should not be very high
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
24. 24
Evaluation
Adoptions
the video pauses if an interaction occurs
lecturers asked to pause his/her presentation at the
occurrence of an interaction at a live broadcasting
the number of interactions is lowered to a maximum
of three interactions in a period of ten minutes
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
25. 25
Evaluation
Testing Adoptions
8 videos of the lecture Learning in the Net: From
possible and feasible things
complete number of attendees vs. active ones to test
the adoptions
active: watched ≥ 75% and attention-level ≥ 50%
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
27. 27
Evaluation
Discussion
the parameters for the calculation of the
attention-level are highly sensitive
the accuracy depends on many different factors (e.g.
difficulty of the questions, the content of the video, ...)
the timespan between the interactions should not be
to small
the two parts of the attention profiling algorithm are
only powerful in combination
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
28. 28
Conclusion
Conclusion
attention is the most crucial resource in human
learning
attention profiling algorithm with to parts
recording of the joined timespans
calculation of an attention-level
delivers realistic values after some adoptions
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014
29. 29
Thank you ...
... for your attention!
Questions?
Josef Wachtler, josef.wachtler@tugraz.at
Martin Ebner, martin.ebner@tugraz.at
ZID – “Social Learning”
Graz, University of Technology
M¨unzgrabenstraße 35A, A-8010 Graz
http://elearningblog.tugraz.at
Josef Wachtler, Martin Ebner and Behnam Taraghi, ZID - Social Learning - TU Graz
HCII 2014