adaptive learning systemsc21u.gatech.edu/sites/default/files/presentations... · university of new...
TRANSCRIPT
Adaptive Learning SystemsThe State of the Field
Michael Madaio
What is Adaptive Learning?
● Individualized learning:○ Content difficulty○ Content form (text, video)○ Sequence of content○ Pace
● Adapted based on:○ Student knowledge, performance○ Student learning style, preferences, goals
OverviewComponentsImplications
Brief History
● 1920’s - Personalized learning at scale (Dewey, Montessori)
● 1960’s - Mastery-based learning (Bloom)
● 1970’s / 1980’s - Computer-assisted Instruction, Cognitive Tutors
Koe
ding
er e
t al.,
199
7; M
ödrit
sche
r et a
l., 2
004
OverviewComponentsImplications
Why now?
● Proliferation of learning technologies○ Students, teachers, institution
● Learning mediated through technology○ Flipped classes, MOOC’s○ Learning apps, student collaboration on forums
OverviewComponentsImplications
Why now?
● Volume of learning data generated○ Variety, velocity, granularity
● Improvements in “big” data analytics○ Machine learning○ Hadoop, MapReduce○ Student clustering○ Learning recommendations
OverviewComponentsImplications
Potential
● Improved student engagement
● Improved student performance on learning outcomes
● Improved student retention in courses
● Improved research-based interventions
Mag
nisa
lis, 2
011;
Nat
riel
lo, 2
011;
Poe
lhub
er,
2008
OverviewComponentsImplications
Challenges
● Material:○ Bandwidth○ Laptops, tablets○ Cost to license software
● Infrastructural:○ System setup and support○ Integration with existing LMS’s○ Training and adoption curve
OverviewComponentsImplications
Challenges
● Pedagogical:○ Integration into existing teaching processes○ Course authoring○ Accuracy of recommendations○ Social dynamics of self-paced learning
OverviewComponentsImplications
Current Adaptive Landscape
● We will look at 16 major adaptive learning providers
● These 16 are currently used in several public and private universities, by a variety of public K-12 school districts, and by corporations and individuals
OverviewComponentsImplications
Adaptive Providers
● Major companies○ Knewton ○ Cerego○ CogBooks○ DreamBox
● Others○ SmartSparrow○ Open Learning Initiative○ LoudCloud○ ALEKS
OverviewComponentsImplications
Adaptive ProvidersOverviewComponentsImplications
Sites of Adaptive Learning
● Arizona State University (Knewton)○ Emporium math courses have seen an 18 percent increase in pass rates and 47 percent
drop in student withdrawals, from 2011-2013. ○ ASU leadership estimates that the institution has retained $12m in what would have
been lost tuition revenue, for 2012-2013.
● University of New South Wales (SmartSparrow)○ Developed with 6 other Australian universities to teach threshold concepts in
mechanics courses○ Saw a decline in average course drop-out rate from 31 percent to 14 percent,
even as course enrollments increased by nearly 30 percent, in 2012-2013.
OverviewComponentsImplications
Edu
catio
n G
row
th A
dvis
ors,
201
3
Other Institutions
Carnegie Mellon UniversityNew York UniversityUniversity of New HampshireSouthern New Hampshire UniversityAmerican Public UniversityWestern Governors University
OverviewComponentsImplications
Adaptive System Components
● Domain Model
● Learner Model
● Adaptation Model
OverviewComponentsImplications
Adaptive System Components
● Domain Model
● Learner Model
● Adaptation Model
OverviewComponentsImplications
Domain Model - Overview
● A conceptual map of the course or domain
● Course elements are assigned to nodes○ Could be videos, articles, assignments, quizzes
● Edges are hierarchical relationships○ Linear or nonlinear○ With prerequisites defined
Nes
bit,
2006
; Che
n, 2
008;
Gra
f & Iv
es, 2
010
OverviewComponentsImplications
Domain Model - Method
● Created and arranged either by the provider or by teachers○ Some adaptive systems use pre-authored content○ Some allow for teacher authoring and arrangement
● Multiple conventions for domain representation and navigation○ Zooming, panning, filtering
Kar
ampi
peris
& S
amps
on, 2
005;
Mag
nisa
lis,
2011
; Chu
ng &
Kim
, 201
2
OverviewComponentsImplications
Domain Model - Risks
● Trade off of teacher autonomy and effectiveness of system○ Issues of consistency and sufficiency of metadata
● Ability for students to view their own “position” in the domain○ Risks of cognitive overload and frustration
Bar
gel e
t al.,
201
2; D
iBito
nto
et a
l., 2
013
OverviewComponentsImplications
Rei
man
n, 2
013
Falm
agne
, 201
1
Domain Model - SystemsOverviewComponentsImplications
Adaptive System Components
● Domain Model
● Learner Model
● Adaptation Model
OverviewComponentsImplications
Learner Model - Overview
● Model of each learner’s current knowledge state
● Either an overlay model or a stereotype model○ Overlay - compared to the overall domain model○ Stereotype - clustered with similar learner models
● Stereotype model is based on assumptions about student similarity
OverviewComponentsImplications
Nitc
hot e
t al.,
201
0; K
nauf
et a
l., 2
010;
Kla
šnja
-M
iliće
vić
et a
l., 2
011
Learner Model - Method
● Data is collected either:○ Statically or dynamically○ Explicitly or implicitly
● Static data○ Cognitive characteristics, background knowledge○ Non-cognitive - topic preference, learning goals
● Explicit data collection methods○ Collected via pre-test, survey, feedback prompts
OverviewComponentsImplications
Kar
ampi
peris
& S
amps
on, 2
005;
Che
n, 2
008;
K
lašn
ja-M
iliće
vić
et a
l., 2
011
Learner Model - Method
● Dynamic data○ Knowledge state○ Learning style (may be static or inferred dynamically)○ Time spent on course element or LMS○ Clickstream data○ Assessment scores ○ Student feedback (prompted explicitly)
● Implicit data collection methods○ Automatically collected through interactions with the
system.
OverviewComponentsImplications
Mor
eno-
Ger
et a
l., 2
007;
Akb
ulut
& C
arda
k,
2012
● Risks for explicit data collection○ Non-completion or inaccuracy○ Interrupts learning process
● Risks for implicit data collection○ Data provenance○ Accuracy of inferences made with data○ Data privacy issues
OverviewComponentsImplicationsLearner Model - Risks
Kla
šnja
-Mili
ćevi
ć et
al.,
201
1; M
agni
salis
et a
l.,
2011
; Lo
et a
l., 2
012
Kla
šnja
-Mili
ćevi
ć, 2
011
Che
n, 2
005
Learner Model - SystemsOverviewComponentsImplications
Adaptive System Components
● Domain Model
● Learner Model
● Adaptation Model
OverviewComponentsImplications
Adaptation Model - Overview
● Unit of Adaptivity - Course element being adapted:○ Difficulty○ Content media○ Sequence (micro or macro)○ Pace
● Method of Adaptation○ Variety of algorithm types used for different purposes○ Direct or indirect presentation of adaptation
OverviewComponentsImplications
Mitr
ovic
& M
artin
, 200
4; K
umar
, 200
6; U
llric
h et
al.,
200
9; S
osno
vsky
, 201
0
● Bayesian network tracing - clustering students into groups
● Hidden Markov models - predicting likelihood of learner success
● Genetic algorithms○ Refine models and construct optimal learning paths
from pre-tests
● Neural networks○ Pattern recognition updated with input (eg: inferring
learning styles from interactions)
OverviewComponentsImplicationsAdaptation Model - Method
Bru
silo
vsky
, 200
1; C
hen,
200
8; M
agni
salis
et
al.,
2011
Adaptation Model - Method
● Direct Adaptation○ Students are given a visible recommendation for the
next course element○ Micro-level - problem feedback, explanations, links○ Macro-level - learning path presented as optional
● Indirect Adaptation○ Link hiding○ Learning path presented as only option
OverviewComponentsImplications
Hau
ger &
Köc
k, 2
007;
Mag
nisa
lis e
t al.,
201
1;
Akb
ulut
& C
arda
k, 2
012
Adaptation Model - Risks
● Risks to student agency if the system controls too much○ Feelings of paternalism
● Systems with learner choice○ Lack of student ability to choose optimal path
● Design of adaptation engine influences the learning process
OverviewComponentsImplications
Bar
gel e
t al.,
201
2; A
kbul
ut &
Car
dak,
201
2;
Kirs
chne
r and
Mie
rren
beer
, 201
3
ww
w.s
mar
tspa
rrow
.com
ww
w.k
new
ton.
com
Adaptation Model - SystemsOverviewComponentsImplications
Implications - Technical
● Must be embedded in a platform○ Either adaptive learning platform○ Integrated with LMS○ Used in a digital textbook
● Need adequate bandwidth, available hardware, technical support for set-up and maintenance
OverviewComponentsImplications
Implications - Pedagogical
● Choose an adaptive provider that fits your goals○ Consider pedagogical goals and mode of instruction○ How much teacher agency involved in authoring the course?
● Need faculty and student buy-in○ Overcoming cultural resistance to new models of learning
● Need departmental or institutional buy-in○ Freedom to experiment○ Competency-based credit
OverviewComponentsImplications
Implications - Pedagogical
● Mode of instruction○ Online course○ Blended, face-to-face
● Teacher goals○ Address differences in background knowledge○ Appeal to different learning styles, goals
OverviewComponentsImplications
Implications - Pedagogical
● How does whole-class instruction happen when students are not working on the same content at the same time?○ Challenges for peer learning, mentoring, collaborative
groups, project-based learning
○ Could be an opportunity for changing how we view those experiences
OverviewComponentsImplications
Implications - Research
● Opportunities○ Improving student learning outcomes, retention,
engagement
○ Informing curricular development
○ Connections between cognitive factors and performance (ie: self-regulation, motivation)
OverviewComponentsImplications
Implications - Research
● Challenges○ Data acquisition - greater volume, variety, and
velocity of data than many teachers or researchers may be equipped to deal with
○ Data privacy - involvement of 3rd party adaptive providers raises questions about privacy and security of data
OverviewComponentsImplications