Presentation from Learning Analytics and Knowledge 2013.
The relatively low completion rates of learners have been a central critique as MOOCs grow in popularity. This focus on completion rates, however, implies a monolithic view of disengagement that fails to acknowledge alternative forms of participation in MOOCs. We develop a classifier which identifies four prototypical trajectories that learners take through MOOCs: Completing learners, Disengaging learners, Auditing learners, and Sampling learners. These subpopulations are defined by learners’ longitudinal patterns of engagement with assessments and video lectures; the subpopulations can be used as a lens to learn more about other aspects of the learners or the courses. Link to full paper in the final slide.
2. MOOCs (in this paper) are
instructionist + individualised
• 6-10 weeks long
• 2-3 hours of video lectures/week
• autograded assessments with regular
deadlines
• discussion forum
3. Massive Open Online Courses
Heterogeneous population:
Learners join from anywhere in the
world, at any age, for any reason
4. Defining Success for Open-Access Learners
Assessment scores are problematic:
• not comparable across courses
• not available for all learners because
test-taking is not aligned with learner
goals
5. Defining Success for Open-Access Learners
Completion rates are highly problematic:
• numerator = certificate earners, i.e.
learners who take assessments
• denominator =
o total enrolled? overestimate; indicator of
interest and not participation
o total active? how defined?
• ignore plurality of learner intentions
• no nuance about subpopulations to help
us design interventions or customized
course features
6. Defining Success for Open-Access Learners
Process measures hold promise:
• conceptualize learning as an ongoing
set of interactions with learning objects
and other humans
• allow early detection and prediction
• indicate points for intervention
7. Defining Success for Open-Access Learners
Assessment scores
Completion rates
Process measures
9. Classification Criteria
Classification methods for MOOC subpopulations:
Universal – valid across multiple courses
Theory-driven – reflect the processes of learning
Parsimonious – based on small, meaningful feature set
Predictive – suggest likely outcomes
Dynamic – account for new information over time
21. Coarse Engagement Labels
(T) On Track: Did the weekly assignment on
time
(B) Behind: Did the weekly assignment, but
finished after the due date
(A) Auditing: Watched videos but did not do
the assignment
(O) Out: Did not interact with the course,
either through videos or assignments
We were able to predict who would take
the final AUC = 0.96
22. The Aggregate Class A = Auditing
O = Out
T = On Track
B = Behind
In this picture Out Is not to scale!
23. The Aggregate Class A = Auditing
O = Out
T = On Track
B = Behind
5k
In this picture Out Is not to scale!
24. The Aggregate Class A = Auditing
O = Out
T = On Track
B = Behind
7k
In this picture Out Is not to scale!
36. Four Prototypical Trajectories
Consistent across three courses:
Auditing learners watch lectures throughout course, but
attempt very few assessments
Completing learners attempt majority of assessments offered
in course
Disengaging learners attempt assessments at beginning of the
course, but then sparsely watch lectures or disappear entirely
Sampling learners briefly explore course by watching a few
videos
46. Cluster Validation
• Different values of k (split by time)
• Including “assignment pass” (95%
overlap)
• Excluding “behind” (94% overlap)
• Silhouette of 0.8 (that’s pretty good)
• Pass the common sense test
47. High Level
Clustering Four
Engagement in Prototypical
MOOCs Patterns
49. Overall Experience
Au
d it
in g
Co
mp
let
ing
Di
se
HS
ng
ag
ing
Sa
mp
lin
g
Completing (and Auditing) Au
have best experience
d it
in g
Co
mp
let
ing
Di
se
UG
ng
ag
ing
Sa
mp
lin
g
Identify subpopulations early Au
d it
in g
to customize course features Co
mp
let
ing
Di
se
GS
ng
ag
ing
Sa
mp
lin
g
3.0 3.5 4.0 4.5 5.0
Overall Experience
50. Discussion Forum
Au
d it
in g
Co
mp
let
ing
Di
se
HS
ng
ag
ing
Completing learners are Sa
mp
most active on the forum
lin
g
Au
d it
in g
Co
mp
let
ing
Di
se
UG
ng
ag
Causal relationship?
ing
Sa
mp
lin
g
Au
d it
in g
Co
mp
let
Reputation systems & Di
se
ing
GS
ng
Social features ag
ing
Sa
mp
lin
g
0.1 0.51.0 2.0 4.0 7.0 10.0
Average Forum Activity
51. Geographical Distribution
Trend confirmed by top four participating countries
United States, India, Russia, United Kingdom
52. Gender
Au
d it
in g
Co
mp
let
ing
Di
se
HS
ng
ag
Female Completing learners
ing
Sa
underrepresented in
mp
lin
g
advanced courses Au
d it
in g
Co
mp
let
ing
Di
se
UG
ng
Stereotype threat?
ag
ing
Sa
Spencer et al., 1999 mp
lin
g
Au
d it
in g
Co
mp
let
ing
Frame assessments to Di
se
GS
ng
ag
ing
minimize stereotype threat Sa
mp
lin
g
2 4 6 8 10 12 14 16
Odds Ratio (Male/Female)
54. Future Directions
Experiments
Collaboration and Peer Effects
Interface Customization
Targeted Interventions
Nuanced Analytics
Auditing: MOOC-as-a-resource vs. MOOC-as-a-class
Disengaging: Early prediction for intervention
Reasons to enroll and trajectories
Engagement trajectories for real-time analytics in
MOOCs
Dashboard visualizations
55. Thank you!
Stanford Lytics Lab lytics.stanford.edu
Office of the Vice Provost for Online Learning
Roy Pea, Clifford Nass, Daphne Koller
Our LAK reviewers
Reference
S. Spencer, C. Steele, and D. Quinn. Stereotype threat and women’s math
performance. Journal of Experimental Social Psychology, 35(1):4–28, 1999.
56. More info?
René Kizilcec kizilcec@stanford.edu
Chris Piech piech@cs.stanford.edu
Emily Schneider elfs@cs.stanford.edu
Stanford’s Learning Analytics Group:
Lytics Lab lytics.stanford.edu
Paper: http://goo.gl/OSX72
Editor's Notes
Better exp. Than DisengagingIdentification via self-report or algorithmice.g. reducing emphasis on assessments for Auditors—but don’t deprive them of the option
Positive feedback loopInvestigate causal relationshipStudy effect of RS&SF on participation
GS-level for illustration
Anyone who prepared for the test can perform well on it.
Experiments based on learning scienceNot certificate vs. no certificate