A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 1
A	Multivariate	Elo-based	Learner	Model	for	
Adaptive	Educational	Systems
Authors:	Solmaz	Abdi,	Hassan	Khosravi,	Shazia Sadiq,	Dragan	Gasevic
Presenter:	Dr Hassan	Khosravi
Senior	Lecturer	at	The	University	of	Queensland
h.khosravi@uq.edu.au
@haskhosravi
hassan-khosravi.net
The	12th International	Conference	on	Educational	Data	Mining
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 2
Introduction
The	M-ELO Approach
Evaluation:	Predictive	Performance
Conclusion	and	Future	Work
Background
Evaluation:	Fit	for	Adaptive	Learning
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 3
Introduction
• Educators continue to face significant challenges in
providing high quality instruction in large diverse online or
on-campus classes.
• Providing tailored
learning resources.
• Monitoring
students’
achievements.
• Providing helpful
feedback and
guidance.
https://myams.org/wp-content/uploads/2015/01/lecture-22903.jpg
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 4
Adaptive Educational Systems
Adaptive	educational	systems	make	use	of	data	about	
students,	learning	process,	and	learning	products	to	adapt	
the	level	or	type	of	instruction	for	each	student.
Domain	Model Learner	Model
Content	Repository
Recommendation	Engine
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 5
Modelling Learners in Adaptive
Educational Systems
• Conventional approaches for learner modelling:
– Bayesian knowledge Tracing (BKT) and its extensions.
– Item Response Theory (IRT) and its extensions
• Neither of these approaches are well-suited for adaptive educational
systems as they generally require pre-calibration on big samples
(Planek, 2016).
• The Elo rating system has been shown to be an effective alternative
for modelling students in adaptive educational system (Wauters et
al., 2011).
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 6
The Elo Rating
• The Elo rating system was originally used to rate chess players.
This	rating	system	is	self-correcting,	meaning	that	the	ratings,	in	the	
long	run,	should	correctly	reflect	the	skill	level	of	the	player
Jack		920 Jane	1600
Outcome Jack’s	rating Jane’s	rating
Jane	wins 910 1610
Jack	wins 1000 1520
The	sum	of	the	updates	to	the	ratings	of	
the	players	is	always	zero	(zero	sum	games)
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 7
ELO Rating in Education
• A similar comparison can be conducted between a student
and a question being attempted by the student.
James	- 920
Outcome Jame’s
rating
Q38	‘s	
difficulty
Answered	
incorrectly
910 1610
Answered
correctly
1000 1520
This	rating	system	is	self-correcting,	meaning	that,	in	the	long	run,	it	should	
correctly	reflect	the	mastery	of	the	student	and	difficulty	level	of	the	question.
The	sum	of	the	updates	to	the	ratings	of	
the	players	is	always	zero	(zero	sum	games)
Question	38	- 1600
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 8
Existing Studies and Our Contribution
Existing studies
• Use repositories that
contain items that are
pure
• Are studied in the
context of adaptive
testing systems.
Our Contribution
• Introduce a new variant
that models students and
items using repositories
that contain non-pure
items.
• Investigate its fit in the
context of adaptive
learning systems.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 9
The	M-ELO Approach
Evaluation:	Predictive	Performance
Conclusion	and	Future	Work
Introduction
Evaluation:	Fit	for	Adaptive	Learning
Background
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 10
Adaptive Educational Systems
Adaptive Testing
• Conducting an exam using a
sequence of questions that are
successively administered.
• Aiming to maximise the precision
of the score within a reasonable
timeframe.
– Exam terminates upon estimating
a student’s ability with a
confidence level that exceeds a
user-specific threshold.
• Students competencies are not
expected to change while using
the system to do an exam.
Adaptive Learning
• Assisting students in their
learning by recommending
learning resources.
• Aiming to provide rich feedback
to students on their learning.
• Students can spend theoretically,
an infinite amount of time on a
learning item or on the system.
• Students competencies are
expected to improve via receiving
feedback or decline as a result of
forgetting over time.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 11
Open Learner Models
Open learner models (OLMs) are learner models that are
externalised and made accessible to students or other
stakeholders often through visualisation, as an important
means of supporting learning.
Visualisations showing	comparison	of	learning	outcome	achievements	(Law	et	al.,	2015)
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 12
RiPPLE: Adaptive Learning Meets Crowdsourcing
RiPPLE	(Recommendation	in	Personalised	Peer	Learning	Environments)	
recommends	personalised	learning	activities	to	students	based	on	
their	knowledge	state	from	a	pool	of	crowdsourced	learning	activities.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 14
Background
Evaluation:	Predictive	Performance
Conclusion	and	Future	Work
Introduction
Evaluation:	Fit	for	Adaptive	Learning
The	M-ELO Approach
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 15
Notation
Learner	model
!" A	set	of	#	students	enrolled	in	the	course,	where	%& is an	arbitrary	student
∆( A	set	of	knowledge	components	contributed	to	the	course,	where	)* is	an	arbitrary	knowledge	
component	(domain	model)
+, A	set	of	- items	contributed	to	the	course,	./ is	an	arbitrary	item	(content	model)
W,×( Matrix,	where	1/* is	2
3⁄ if	item	./ is	tagged	with 5 knowledge	components	including	knowledge
component	)*,	and	0	otherwise.	
6"×, Matrix,	where	7&/ is	1	if student	%& answers	item	./ correctly and	0	if	answered	incorrectly
Modelling learners	and	Items
8	" Vector, The		Elo-based	learner	model,	where	8& indicates	%&’s	proficiency level	on	the	entire	
domain
L"×, Matrix,	The	Elo based	learner	model	based	on	multi-concept	M-Elo (where	l&* represents	student	
%&’s	Elo rating	on	knowledge	component	)* , approximating	the	proficiency	level	of	the	student	on	
that	certain	knowledge	component
9/ Vector,	The	Elo rating	of	questions	in	the	question	repository,	where	:/ indicates	the	
approximated	difficulty	of	question	./
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 16
Standard Elo-based Learner Model
7
James	- 920
Outcome Jame’s
rating
Q38	‘s	
difficulty
Answered	
incorrectly
910 1610
Answered
correctly
1000 1520
The	sum	of	the	updates	to	the	ratings	of	
the	players	is	always	zero	(zero	sum	games)
Question	38	- 1600
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 17
Standard Elo-based Learner Model
7
;(7&/|8&, :/) =	@(8& 	−	:/)
:/ ≔	:/ + D(; 7&/|8&, :/ −	7&/)
Probability	of		%& answering	./ correctly
Logistic	function
Updating	the	estimate	of	:/
8& ≔ 8& + 	D(7&/ 	− ;(7&/|8&, :/))
Updating	%&’s	Elo rating
K	=	Sensitivity	of	the	estimations	to	the	
student’s	last	attempt.	Can	be	replaced	with:
! E =	
G
1 + 	I ∗ E
2
Number	of	prior	updates
3
1
Follows	the	principles	
of	a	zero-sum	game
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 18
Single-Concept Multivariate Elo-based Learner model
(Doebler et al., 2015; Pelanek et al., 2017)
7
James	
Outcome Jame’s rating Q38	‘s	
difficulty
Answered	
incorrectly
Answered
correctly
DBMS ER … Map-ER …
1100 1040 … 910 …
DBMS ER … Map-ER …
1170 1040 … 1000 …
1610
1520
DBMS ER … Map-ER
1100 1040 … 920
Question	40	- 1600
Question	38	– 1600
(Map-ER)
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 19
Multi-Concept Multivariate Elo-based Learner model
7
James	
Outcome Jame’s rating Q40	‘s	
difficulty
Answered	
incorrectly
Answered
correctly
DBMS ER … Map-ER …
1080 1040 … 910 …
DBMS ER … Map-ER …
1170 1130 … 970 …
1630
1510
DBMS ER … Map-ER
1100 1040 … 920
Question	40	- 1600
(DBMS,	Map-ER)
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 20
Multi-Concept Multivariate Elo-based
learner model (M-Elo)
%&’s	average	competency	on	the	knowledge	
components	associated	with	item	./
1
lK&/ =	∑ l&*×1/*
(
*M2
;(7&/|lK&/, :/) =	@(lK& 	−	:/)
:/ ≔	:/ + D(; 7&/|lK&/, :/ −	7&/)
Probability	of		%& answering	./ correctly
Logistic	function
Updating	the	estimate	of	:/
l&* 	≔ l&* + 	a N D(7&/ 	− ;(7&/|l&*, :/))
Updating	%&’s	Elo rating	on	each	knowledge	
component	)* the	question	is	tagged	with
Sensitivity	of	the	estimations	to	the	student’s	last	
attempt.	Can	be	replaced	with:
! E =	
G
1 + 	I ∗ E
a =	
|; 7&/ lK&/, :/ −	7&/|
∑ (|7&/ 	− ;(7&/|l&*, :/)×1/*|)(
&M2
2
3 4
Number	of	prior	updates
Follows	the	principles	
of	a	zero-sum	game
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 21
Background
The	M-ELO Approach
Conclusion	and	Future	Work
Introduction
Evaluation:	Fit	for	Adaptive	Learning
Evaluation:	Predictive	Performance
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 22
Evaluation: Predictive Performance
• Comparing the predictive performance of M-Elo
against Elo using
– A suite of simulated data sets
– Publicly available data sets
• Three Metrics:
– Area under the curve (AUC)
– Root Mean Squared Error (RMSE)
– Accuracy (ACC)
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 23
Synthetic Data Set
Dirichlet distribution	
(topic	level	gaps)
σ
Sparsity
U
Users
L
#	topics Discrete	uniform	
distribution
(topics)	
Q
Questions
Latent	Trait	
Models	
A
T
Normal	distribution
(difficulty,	
discrimination)
Normal	
distribution
D
Tags
Difficulties
Answers
• For	all	experiments,	100	students,	1000	learning	items	and		70000	answers	were	
sampled.	70%	of	the	created	data	set	was	used	for	training,	and	the	remaining	30%	
reserved	for	testing.
• Each	experiment	was	repeated	for	5	times	and	the	reported	values	are	the	average	
results	across	the	five	runs	(Desmarais et	al.,	2010;	Khosravi	et	al.,	2017).
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 24
Synthetic Data Sets - Results
• For	L=10,	as	σ is	increased	and	students	with	more	diversity	in	their	abilities	
across	different	concepts	are	generated,	M-Elo outperforms	Elo.
• For	L=100,	the	same	trend	is	observable;	however,	the	intersection	point	for	
where	M-Elo outperforms	Elo occurs	for	larger	value	of	σ.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 25
Public Data Sets
Data	Set Alg2005 Alg2006 BAlg2006
Students 575 1840 1146
KC 112 714 493
Items 147,914 319,151 19,954
Multi-KC 51,171 21,415 1,650
Interactions 609,979 1,825,030 1,822,697
Data	sets	from	PSLC datashop
• AlgebraI 2005-2006	(Alg2005)
• AlgebraI 2006-2007	(Alg2006)
• Bridge	to	Algebra	2006-2007	
(BAlg2006)
Data	cleaning
• Using	the	train/test	split	provided	by	KDD	Cup	2010
• Discarding	interactions	with	no	assigned	concepts
• Using	Step	Name	column	as	the	learning	item
• Each	learning	item	is	associated	with	one	or	more	concepts	(KC)	
covered	in	the	course
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 26
Public Data Sets - Results
Data	Set AUC
Elo M-Elo
Alg2005 0.726 0.750
Alg2006 0.687 0.695
BAlg2006 0.676 0.712
RMSE
Elo M-Elo
0.392 0.385
0.394 0.390
0.368 0.361
ACC
Elo M-Elo
0.787 0.790
0.784 0.797
0.827 0.828
• M-Elo outperforms	Elo on	all	three	data	sets.
• It	may	be	possible	to	hypothesise	that	students	often	
have	different	competency	levels	on	different	
concepts,	but	these	differences	are	often	not	too	
significant.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 27
Background
The	M-ELO Approach
Conclusion	and	Future	Work
Introduction
Evaluation:	Predictive	Performance
Evaluation:	Fit	for	Adaptive	Learning
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 28
Case study
• Location: The University of Queensland (UQ)
• Course: Introduction to relational databases
• General Topics: 17 concepts such mainly on
ER diagrams, relational models, functional
dependencies and SQL.
• Number of students: 521
• Number of items: 1,632
• Number of attempts: 91,340
The	learner	model	is	shared	with	students	based	on	the	principles	of	OLMs through	
a	visualisation widget
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 29
Guiding Adaptivity
• Estimating the knowledge state of students on each
topic as well as estimating the difficulty level of
each item.
• Recommending learning items based on knowledge
gaps at the right level of difficulty.
The	adaptivity at	the	concept	level	is	possible	because	of	the	
availability	of	the	additional	parameters	learned	by	M-Elo,	
which	is	not	achievable	using	standard	Elo rating
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 30
Insights from Student Survey (N=53)
• Motivation: The visualisation used by RiPPLE for showing my
knowledge state increase my motivation to study or further use the
system
• Rationality: Having the ability to visually see my knowledge state,
help to understand the rationale behind suggestions made by the
system
• Trust: Having the ability to visually see my knowledge state,
increases my trust in recommended items
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 31
Insights from the Case Study
• Challenge: Maintaining a balance between the students' proficiency
level and the learning items' difficulty level in our pilot.
• Outcome: Students lost motivation in answering questions as the
large loss of rating in answering the question incorrectly outweighed
the small rating gain received in answering the question correctly.
The	zero-sum	game	principles	of	the	Elo rating	might	not	be	ideal	for	
adaptive	learning	systems.
600
800
1000
1200
1400
Rating
Time
Avg	student	and	question	rating	over	time
Student	rating question	rating
Potential	explanation:	Students	have	
full	access	to	the	internet,	textbooks	
and	colleagues	as	well	as	an	infinite	
amount	of	time	for	answering	a	
question.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 32
Insights from the Case Study
• Challenge: reducing the sensitivity of the estimations over time as
the number of attempts is increased means that students are not able
to make significant changes to their knowledge state despite
answering questions correctly.
• Outcome: students tend to get discouraged from using the system
once they have used it for a while.
Reducing	the	sensitivity	of	the	estimations	over	time	might	not	be	
ideal	for	adaptive	learning	systems.
600
800
1000
1200
1400
Rating
Time
Avg	student	and	question	rating	over	time
Student	rating question	rating
This	seems	better	suited	for	adaptive	
testing	where	students	are	not	expected	
to	learn	in	one	sitting	and	the	uncertainty	
function	helps	stabilise the	ratings.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 33
Background
The	M-ELO Approach
Evaluation:	Fit	for	Adaptive	Learning
Introduction
Evaluation:	Predictive	Performance
Conclusion	and	Future	Work
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 34
Conclusion and Future Work
• We introduced a new multi-variate Elo-based learner
model called M-Elo where learning items can be tagged
with one or more concepts.
• M-Elo outperforms the standard Elo-based model in real-
world environments, is interpretable by students and
provides better concept-level adaptivity.
• Future work focuses on tailoring M-Elo to make a better fit
for adaptive learning than adaptive testing.
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 35
Thank you!
Presenter:	Dr Hassan	Khosravi
Senior	Lecturer	at	The	University	of	Queensland
h.khosravi@uq.edu.au
@haskhosravi
hassan-khosravi.net
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 36
The Elo Rating
• The Elo rating system was originally used to rate chess players.
• Principle: each player is assigned a rating, which is updated after
each match.
– The rating of the winner is increased and the rating of the loser is
decreased.
– If a strong player beats a weak player, the result is not surprising
and the update is small, whereas if the opposite happens, the
update is large.
– The sum of the updates to the ratings of the players is always
zero (zero sum games)
This	rating	system	is	self-correcting,	meaning	
that	the	ratings,	in	the	long	run,	should	
correctly	reflect	the	skill	level	of	the	player
A Multivariate Elo-based Learner Model for Adaptive Educational Systems Page 37
ELO Rating in Education
• A similar comparison can be conducted between a student
and a question being attempted by the student.
• Principle:	students	and	questions	are	assigned	a	rating,	which	is	
updated	after	each	attempt.
• If	the	question	is	answered	correctly,	the	student's	rating	increases	
and	the	rating	of	the	question	decreases.	
• If	the	question	is	answered	incorrectly,	the	student's	rating	decreases	
and	the	rating	of	the	question	increases.	
• The	update	to	the	ratings	is	proportional	to	the	difference	between	
the	ratings	of	the	student	and	the	question.	
• The	sum	of	the	updates	to	the	ratings	of	the	players	is	always	zero	
(zero	sum	games)

A Multivariate Elo-based Learner Model for Adaptive Educational Systems

  • 1.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 1 A Multivariate Elo-based Learner Model for Adaptive Educational Systems Authors: Solmaz Abdi, Hassan Khosravi, Shazia Sadiq, Dragan Gasevic Presenter: Dr Hassan Khosravi Senior Lecturer at The University of Queensland h.khosravi@uq.edu.au @haskhosravi hassan-khosravi.net The 12th International Conference on Educational Data Mining
  • 2.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 2 Introduction The M-ELO Approach Evaluation: Predictive Performance Conclusion and Future Work Background Evaluation: Fit for Adaptive Learning
  • 3.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 3 Introduction • Educators continue to face significant challenges in providing high quality instruction in large diverse online or on-campus classes. • Providing tailored learning resources. • Monitoring students’ achievements. • Providing helpful feedback and guidance. https://myams.org/wp-content/uploads/2015/01/lecture-22903.jpg
  • 4.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 4 Adaptive Educational Systems Adaptive educational systems make use of data about students, learning process, and learning products to adapt the level or type of instruction for each student. Domain Model Learner Model Content Repository Recommendation Engine
  • 5.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 5 Modelling Learners in Adaptive Educational Systems • Conventional approaches for learner modelling: – Bayesian knowledge Tracing (BKT) and its extensions. – Item Response Theory (IRT) and its extensions • Neither of these approaches are well-suited for adaptive educational systems as they generally require pre-calibration on big samples (Planek, 2016). • The Elo rating system has been shown to be an effective alternative for modelling students in adaptive educational system (Wauters et al., 2011).
  • 6.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 6 The Elo Rating • The Elo rating system was originally used to rate chess players. This rating system is self-correcting, meaning that the ratings, in the long run, should correctly reflect the skill level of the player Jack 920 Jane 1600 Outcome Jack’s rating Jane’s rating Jane wins 910 1610 Jack wins 1000 1520 The sum of the updates to the ratings of the players is always zero (zero sum games)
  • 7.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 7 ELO Rating in Education • A similar comparison can be conducted between a student and a question being attempted by the student. James - 920 Outcome Jame’s rating Q38 ‘s difficulty Answered incorrectly 910 1610 Answered correctly 1000 1520 This rating system is self-correcting, meaning that, in the long run, it should correctly reflect the mastery of the student and difficulty level of the question. The sum of the updates to the ratings of the players is always zero (zero sum games) Question 38 - 1600
  • 8.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 8 Existing Studies and Our Contribution Existing studies • Use repositories that contain items that are pure • Are studied in the context of adaptive testing systems. Our Contribution • Introduce a new variant that models students and items using repositories that contain non-pure items. • Investigate its fit in the context of adaptive learning systems.
  • 9.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 9 The M-ELO Approach Evaluation: Predictive Performance Conclusion and Future Work Introduction Evaluation: Fit for Adaptive Learning Background
  • 10.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 10 Adaptive Educational Systems Adaptive Testing • Conducting an exam using a sequence of questions that are successively administered. • Aiming to maximise the precision of the score within a reasonable timeframe. – Exam terminates upon estimating a student’s ability with a confidence level that exceeds a user-specific threshold. • Students competencies are not expected to change while using the system to do an exam. Adaptive Learning • Assisting students in their learning by recommending learning resources. • Aiming to provide rich feedback to students on their learning. • Students can spend theoretically, an infinite amount of time on a learning item or on the system. • Students competencies are expected to improve via receiving feedback or decline as a result of forgetting over time.
  • 11.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 11 Open Learner Models Open learner models (OLMs) are learner models that are externalised and made accessible to students or other stakeholders often through visualisation, as an important means of supporting learning. Visualisations showing comparison of learning outcome achievements (Law et al., 2015)
  • 12.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 12 RiPPLE: Adaptive Learning Meets Crowdsourcing RiPPLE (Recommendation in Personalised Peer Learning Environments) recommends personalised learning activities to students based on their knowledge state from a pool of crowdsourced learning activities.
  • 13.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 14 Background Evaluation: Predictive Performance Conclusion and Future Work Introduction Evaluation: Fit for Adaptive Learning The M-ELO Approach
  • 14.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 15 Notation Learner model !" A set of # students enrolled in the course, where %& is an arbitrary student ∆( A set of knowledge components contributed to the course, where )* is an arbitrary knowledge component (domain model) +, A set of - items contributed to the course, ./ is an arbitrary item (content model) W,×( Matrix, where 1/* is 2 3⁄ if item ./ is tagged with 5 knowledge components including knowledge component )*, and 0 otherwise. 6"×, Matrix, where 7&/ is 1 if student %& answers item ./ correctly and 0 if answered incorrectly Modelling learners and Items 8 " Vector, The Elo-based learner model, where 8& indicates %&’s proficiency level on the entire domain L"×, Matrix, The Elo based learner model based on multi-concept M-Elo (where l&* represents student %&’s Elo rating on knowledge component )* , approximating the proficiency level of the student on that certain knowledge component 9/ Vector, The Elo rating of questions in the question repository, where :/ indicates the approximated difficulty of question ./
  • 15.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 16 Standard Elo-based Learner Model 7 James - 920 Outcome Jame’s rating Q38 ‘s difficulty Answered incorrectly 910 1610 Answered correctly 1000 1520 The sum of the updates to the ratings of the players is always zero (zero sum games) Question 38 - 1600
  • 16.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 17 Standard Elo-based Learner Model 7 ;(7&/|8&, :/) = @(8& − :/) :/ ≔ :/ + D(; 7&/|8&, :/ − 7&/) Probability of %& answering ./ correctly Logistic function Updating the estimate of :/ 8& ≔ 8& + D(7&/ − ;(7&/|8&, :/)) Updating %&’s Elo rating K = Sensitivity of the estimations to the student’s last attempt. Can be replaced with: ! E = G 1 + I ∗ E 2 Number of prior updates 3 1 Follows the principles of a zero-sum game
  • 17.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 18 Single-Concept Multivariate Elo-based Learner model (Doebler et al., 2015; Pelanek et al., 2017) 7 James Outcome Jame’s rating Q38 ‘s difficulty Answered incorrectly Answered correctly DBMS ER … Map-ER … 1100 1040 … 910 … DBMS ER … Map-ER … 1170 1040 … 1000 … 1610 1520 DBMS ER … Map-ER 1100 1040 … 920 Question 40 - 1600 Question 38 – 1600 (Map-ER)
  • 18.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 19 Multi-Concept Multivariate Elo-based Learner model 7 James Outcome Jame’s rating Q40 ‘s difficulty Answered incorrectly Answered correctly DBMS ER … Map-ER … 1080 1040 … 910 … DBMS ER … Map-ER … 1170 1130 … 970 … 1630 1510 DBMS ER … Map-ER 1100 1040 … 920 Question 40 - 1600 (DBMS, Map-ER)
  • 19.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 20 Multi-Concept Multivariate Elo-based learner model (M-Elo) %&’s average competency on the knowledge components associated with item ./ 1 lK&/ = ∑ l&*×1/* ( *M2 ;(7&/|lK&/, :/) = @(lK& − :/) :/ ≔ :/ + D(; 7&/|lK&/, :/ − 7&/) Probability of %& answering ./ correctly Logistic function Updating the estimate of :/ l&* ≔ l&* + a N D(7&/ − ;(7&/|l&*, :/)) Updating %&’s Elo rating on each knowledge component )* the question is tagged with Sensitivity of the estimations to the student’s last attempt. Can be replaced with: ! E = G 1 + I ∗ E a = |; 7&/ lK&/, :/ − 7&/| ∑ (|7&/ − ;(7&/|l&*, :/)×1/*|)( &M2 2 3 4 Number of prior updates Follows the principles of a zero-sum game
  • 20.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 21 Background The M-ELO Approach Conclusion and Future Work Introduction Evaluation: Fit for Adaptive Learning Evaluation: Predictive Performance
  • 21.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 22 Evaluation: Predictive Performance • Comparing the predictive performance of M-Elo against Elo using – A suite of simulated data sets – Publicly available data sets • Three Metrics: – Area under the curve (AUC) – Root Mean Squared Error (RMSE) – Accuracy (ACC)
  • 22.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 23 Synthetic Data Set Dirichlet distribution (topic level gaps) σ Sparsity U Users L # topics Discrete uniform distribution (topics) Q Questions Latent Trait Models A T Normal distribution (difficulty, discrimination) Normal distribution D Tags Difficulties Answers • For all experiments, 100 students, 1000 learning items and 70000 answers were sampled. 70% of the created data set was used for training, and the remaining 30% reserved for testing. • Each experiment was repeated for 5 times and the reported values are the average results across the five runs (Desmarais et al., 2010; Khosravi et al., 2017).
  • 23.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 24 Synthetic Data Sets - Results • For L=10, as σ is increased and students with more diversity in their abilities across different concepts are generated, M-Elo outperforms Elo. • For L=100, the same trend is observable; however, the intersection point for where M-Elo outperforms Elo occurs for larger value of σ.
  • 24.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 25 Public Data Sets Data Set Alg2005 Alg2006 BAlg2006 Students 575 1840 1146 KC 112 714 493 Items 147,914 319,151 19,954 Multi-KC 51,171 21,415 1,650 Interactions 609,979 1,825,030 1,822,697 Data sets from PSLC datashop • AlgebraI 2005-2006 (Alg2005) • AlgebraI 2006-2007 (Alg2006) • Bridge to Algebra 2006-2007 (BAlg2006) Data cleaning • Using the train/test split provided by KDD Cup 2010 • Discarding interactions with no assigned concepts • Using Step Name column as the learning item • Each learning item is associated with one or more concepts (KC) covered in the course
  • 25.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 26 Public Data Sets - Results Data Set AUC Elo M-Elo Alg2005 0.726 0.750 Alg2006 0.687 0.695 BAlg2006 0.676 0.712 RMSE Elo M-Elo 0.392 0.385 0.394 0.390 0.368 0.361 ACC Elo M-Elo 0.787 0.790 0.784 0.797 0.827 0.828 • M-Elo outperforms Elo on all three data sets. • It may be possible to hypothesise that students often have different competency levels on different concepts, but these differences are often not too significant.
  • 26.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 27 Background The M-ELO Approach Conclusion and Future Work Introduction Evaluation: Predictive Performance Evaluation: Fit for Adaptive Learning
  • 27.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 28 Case study • Location: The University of Queensland (UQ) • Course: Introduction to relational databases • General Topics: 17 concepts such mainly on ER diagrams, relational models, functional dependencies and SQL. • Number of students: 521 • Number of items: 1,632 • Number of attempts: 91,340 The learner model is shared with students based on the principles of OLMs through a visualisation widget
  • 28.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 29 Guiding Adaptivity • Estimating the knowledge state of students on each topic as well as estimating the difficulty level of each item. • Recommending learning items based on knowledge gaps at the right level of difficulty. The adaptivity at the concept level is possible because of the availability of the additional parameters learned by M-Elo, which is not achievable using standard Elo rating
  • 29.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 30 Insights from Student Survey (N=53) • Motivation: The visualisation used by RiPPLE for showing my knowledge state increase my motivation to study or further use the system • Rationality: Having the ability to visually see my knowledge state, help to understand the rationale behind suggestions made by the system • Trust: Having the ability to visually see my knowledge state, increases my trust in recommended items
  • 30.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 31 Insights from the Case Study • Challenge: Maintaining a balance between the students' proficiency level and the learning items' difficulty level in our pilot. • Outcome: Students lost motivation in answering questions as the large loss of rating in answering the question incorrectly outweighed the small rating gain received in answering the question correctly. The zero-sum game principles of the Elo rating might not be ideal for adaptive learning systems. 600 800 1000 1200 1400 Rating Time Avg student and question rating over time Student rating question rating Potential explanation: Students have full access to the internet, textbooks and colleagues as well as an infinite amount of time for answering a question.
  • 31.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 32 Insights from the Case Study • Challenge: reducing the sensitivity of the estimations over time as the number of attempts is increased means that students are not able to make significant changes to their knowledge state despite answering questions correctly. • Outcome: students tend to get discouraged from using the system once they have used it for a while. Reducing the sensitivity of the estimations over time might not be ideal for adaptive learning systems. 600 800 1000 1200 1400 Rating Time Avg student and question rating over time Student rating question rating This seems better suited for adaptive testing where students are not expected to learn in one sitting and the uncertainty function helps stabilise the ratings.
  • 32.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 33 Background The M-ELO Approach Evaluation: Fit for Adaptive Learning Introduction Evaluation: Predictive Performance Conclusion and Future Work
  • 33.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 34 Conclusion and Future Work • We introduced a new multi-variate Elo-based learner model called M-Elo where learning items can be tagged with one or more concepts. • M-Elo outperforms the standard Elo-based model in real- world environments, is interpretable by students and provides better concept-level adaptivity. • Future work focuses on tailoring M-Elo to make a better fit for adaptive learning than adaptive testing.
  • 34.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 35 Thank you! Presenter: Dr Hassan Khosravi Senior Lecturer at The University of Queensland h.khosravi@uq.edu.au @haskhosravi hassan-khosravi.net
  • 35.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 36 The Elo Rating • The Elo rating system was originally used to rate chess players. • Principle: each player is assigned a rating, which is updated after each match. – The rating of the winner is increased and the rating of the loser is decreased. – If a strong player beats a weak player, the result is not surprising and the update is small, whereas if the opposite happens, the update is large. – The sum of the updates to the ratings of the players is always zero (zero sum games) This rating system is self-correcting, meaning that the ratings, in the long run, should correctly reflect the skill level of the player
  • 36.
    A Multivariate Elo-basedLearner Model for Adaptive Educational Systems Page 37 ELO Rating in Education • A similar comparison can be conducted between a student and a question being attempted by the student. • Principle: students and questions are assigned a rating, which is updated after each attempt. • If the question is answered correctly, the student's rating increases and the rating of the question decreases. • If the question is answered incorrectly, the student's rating decreases and the rating of the question increases. • The update to the ratings is proportional to the difference between the ratings of the student and the question. • The sum of the updates to the ratings of the players is always zero (zero sum games)