How Engaging Is Your Lecture?
Aku • Loukas • Thaha • Vlad
UBISS 2019 • Oulu, Finland
Video
Problem
Lecture Boredom
23.000 universities Webometrics
50% time boring for students Götz
60% of students bored Guardian
2019 end year course evaluations
Context
Lecture Atmosphere
Many signals: lecturer speaks, slides
unfold, students roll eyes, ...
Context sensing: possible with
available, unobtrusive technologies
Insufficient research to quantify
classroom audience engagement
Input
Sound + Ratings
Lecture sound recordings for
classroom atmosphere
Ratings given by us as ground truth
of audience engagement
Method
Study Setup
16 lecture videos from YouTube ⨉
5 audio clips per video ⨉
4 workshop students ⨉
1 rating per clip =
320 data points
Output
Lecture Properties
Independently assess 4 types of
lecture property ranges Maggiani
1. Dreadful - Cheerful
2. Timid - Assertive
3. Monotonous - Dynamic
4. Vague - Clear
Visualization
Monotonous
Dynamic
Results
Model Accuracy
Machine Learning Classification
65%
Dreadful - Cheerful
72%
Timid - Assertive
75%
Vague - Clear
55%
Monoton - Dynamic
Prototype
Functionality
1. User clicks Start Recording
2. App records sound
3. App analyzes sound
4. User sees engagement
ratings over time
5. User clicks Stop Recording
Technology
● Client Javascript
● Server Python
● Model Python
Lessons
Learned
Models gave good results for the
time and resources we had
Android sound mobile sensing was
particularly challenging
Opportunity for holistic sensing of
classrooms at scale
Thank You
Loukas Konstantinou • Cyprus University of Technology
Vlad Manea • University of of Copenhagen
Thaha Mohammed • Aalto University
Aku Visuri • University of Oulu
UBISS 2019 • Oulu, Finland

How engaging is your lecture?