Triangular Learner Model

Triangular Learner Model (TLM)
The 1st International Conference of TESOL
& Education (ICTE) and VLTESOL2022
Innovation in E-learning and Emerging Issues in
Teaching Foreign Languages in Post-Covid Era
Date: 22nd January 2022 on Microsoft Teams
Place: Van Lang University, Ho Chi Minh city, Vietnam
Presenter: Dr. Loc Nguyen, PhD
Email: ng_phloc@yahoo.com
Homepage: www.locnguyen.net
22/01/2022 TLM - Core of PhD research 1

Triangular Leaner Model
User model is description of users’ information and characteristics in
abstract level. User model is very important to adaptive software
which aims to support user as much as possible. The process to
construct user model is called user modeling. Within learning context
where users are learners, the research proposes a so-called
Triangular Learner Model (TLM) which is composed of three
essential learners’ properties such as knowledge, learning style, and
learning history. TLM is the user model that supports built-in
inference mechanism. So, the strong point of TLM is to reason out
new information from users, based on mathematical tools. This paper
focuses on fundamental algorithms and mathematical tools to
construct three basic components of TLM such as knowledge sub-
model, learning style sub-model, and learning history sub-model. In
general, the paper is a summary of results from research on TLM.
Algorithms and formulas are described by the succinct way.

Triangular Learner Model
I. Triangular Learner Model (TLM)
II. Zebra: a user modeling system for TLM
III. Knowledge sub-model in TLM
IV. Learning style sub-model in TLM
V. Learning history sub-model in TLM
Original PhD. Report: Triangular Learner Model
(November 17 2009)
Advisor: Prof. Dr. Đồng Thị Bích Thủy
Researcher: Nguyễn Phước Lộc
Affiliation: Department of IS, Faculty of IT, University of
Science

I. Triangular Leaner Model
Adaptive System
Selection Rules
User Modeling System
User Model
TARGET: Adaptive System
changes its action to provide
learning materials for every
student in accordance with her/his
model
Learning Materials

• Too much information about individuals to
model all users’ characteristics → it is
necessary to choose essential
characteristics from which a stable
architecture of user model is built.
• Some user modeling systems (UMS) lack
of powerful inference mechanism → need
a solid inference UMS
Hazards of User Modeling

I. Triangular Leaner Model (TLM)
Triangular
Learner
Model
(TLM)
• Knowledge (K) sub-model is the combination of overlay model and Bayesian network
• Learning style (LS) sub-model is defined as the composite of characteristic cognitive,
affective and psychological factors
• Learning history (LH) is defined as a transcript of all learners’ actions such as learning
materials access, duration of computer use, doing exercise, taking an examination, doing
test, communicating with teachers or classmates, etc

• Knowledge, learning styles and learning history are
prerequisite for modeling learner
• While learning history changes themselves frequently,
learning styles and knowledge are relatively stable.
The combination of them ensures the integrity of
information about learner
• User knowledge is domain specific information and
learning styles are personal traits. The combination of
them supports user modeling system to take full
advantages of both domain specific information and
domain independent information
Why TLM?

extended Triangular Leaner Model

• How to build up TLM?
• How to manipulate (manage) TLM?
• How to infer new information from TLM?
→ Zebra: the user modeling system for TLM

• Mining Engine (ME) manages
learning history sub-model of
TLM
• Belief Network Engine (BNE)
manages knowledge sub-
model and learning style sub-
model of TLM
• Communication Interfaces
(CI) allows users and adaptive
systems to see or modify
restrictedly TLM

• Collecting learners’ data, monitoring their
actions, structuring and updating TLM.
• Providing important information to belief
network engine
• Supporting learning concept recommendation
• Discovering some other characteristics
(beyond knowledge and learning styles) such
as interests, goals, etc
• Supporting collaborative learning through
constructing learner groups (communities)
Mining Engine

• Inferring new personal traits from TLM by
using deduction mechanism available in
belief network
• This engine applies Bayesian network
and hidden Markov model into inference
mechanism.
• Two sub-models: knowledge & learning
style are managed by this engine
Belief Network Engine

The extended
architecture of
Zebra when
interacting with AES

III. Knowledge sub-model
Knowledge sub-model = overlay model + Bayesian network (BN)





n
i
i
i
n
h
w
Y
Y
Y
X
1
*
)
,...,
,
|
1
Pr( 2
1
Determining CPT (s) is based on weights of arcs


 

otherwise
X
if Y
h i
i
0
1

T1
C O I p(J = 1)
P(J = 0)
1- p(J = 1)
1 1 1 1.0 (0.1*1 + 0.5*1 + 0.4*1) 0.0
1 1 0 0.6 (0.1*1 + 0.5*1 + 0.4*0) 0.4
1 0 1 0.5 (0.1*1 + 0.5*0 + 0.4*1) 0.5
1 0 0 0.1 (0.1*1 + 0.5*0 + 0.4*0) 0.9
0 1 1 0.9 (0.1*0 + 0.5*1 + 0.4*1) 0.1
0 1 0 0.5 (0.1*0 + 0.5*1 + 0.4*0) 0.5
0 0 1 0.4 (0.1*0 + 0.5*0 + 0.4*1) 0.4
0 0 0 0.0 (0.1*0 + 0.5*0 + 0.4*0) 1.0
T2
E Pr(E = 1)
Pr(E = 0)
1- Pr(E = 1)
1 0.8 (0.8*1) 0.2
0 0.0 (0.8*0) 1.0




n
i
i
i
n
h
w
Y
Y
Y
X
1
*
)
,...,
,
|
1
Pr( 2
1


 

otherwise
X
if Y
h i
i
0
1
T3
Q Pr(Q = 1)
Pr(Q = 0)
1- Pr(Q = 1)
1 0.2 (0.2*1) 0.8
0 0.0 (0.2*0) 1.0
Determining CPT (s) is based on weights of arcs

• Parameter Learning: using Expectation
Maximization (EM) algorithm or Maximum
Likelihood Estimation (MLE) algorithm.
Both of them are used for beta
distributions
• Structure Learning and monitoring:
using Dynamic Bayesian Network (DBN)
Improving knowledge sub-model

III. Knowledge sub-model (EM)
Beta density function

EM technique

III. Knowledge sub-model (MLE)
• The essence of maximizing the likelihood
function is to find the peak of the curve of
LnL(θ).
• This can be done by setting the first-order partial
derivative of LnL(θ) with respect to each
parameter θi to 0 and solving this equation to
find parameter θi
  

  










n
i
n
i
n
i
b
i
a
i
n
b
i
a
i
n
i
i x
x
b
a
B
x
x
b
a
B
b
a
x
f
L
1 1 1
1
1
1
1
1
)
1
(
)
,
(
1
)
1
(
)
,
(
1
)
,
,
(
)
(
MLE technique



























2
2
1
1
1
1
1
1
)
,
(
)
,
(
)
1
ln(
1
)
1
(
ln
1
)
1
(
L
b
a
F
L
b
a
F
x
n
C
k
e
x
n
C
k
e
n
i
i
a
k
k
a
k
b
n
i
i
b
k
k
b
k
a
The equations whose solutions are
parameter estimators

Iterative Algorithm for MLE

III. Knowledge sub-model (DBN)
• An initial BN G0 = {X[0], Pr(X[0]} at first time t = 0
• A transition BN is a template consisting of a transition DAG G→
containing variables in X[t], X[t+1] and a transition probability
distribution Pr→ (X[t+1] | X[t])
A DBN is BN containing variables that comprise T variable vectors X[t]

• DBN can model the temporal relationships among
variables. It can capture the dynamic aspect
• So DBN allows monitoring user’s process of
gaining knowledge and evaluating her/his
knowledge
• The size of DBN becomes numerous when the
process continues for a long time
• The number of transition dependencies among
points in time is too large to compute posterior
marginal probabilities
Strong points of DBN
Drawbacks of DBN

• To overcome these drawbacks, the new algorithm
that both the size of DBN and the number of
Conditional Probability Tables (CPT) in DBN are
kept intact when the process continues for a long
time
• To solves the problem of temporary slip and lucky
guess: “learner does (doesn’t) know a particular
subject but there is solid evidence convincing that
she/he doesn’t (does) understand it; this evidence
just reflects a temporary slip (or lucky guess)”.
Purposes of suggested algorithm to improve DBN

1. Initializing DBN
2. Specifying transition weights
3. Re-constructing DBN
4. Normalizing weights of dependencies
5. Re-defining CPT (s)
6. Probabilistic inference
The algorithm for DBN includes 6 steps that
repeated whenever evidences occur

1. Initializing DBN 2. Specifying transition weights 3. Re-constructing
4. Normalizing weights
5. Re-defining CPT(s)
6. Probabilistic inference

IV. Learning style sub-model
• S={s1, s2,…, sn} is the finite set of states
• Ө={θ1, θ2,…, θm} is the set of observations
• A is the transition probability matrix in which aij is
the probability that, the process change the
current state si to next state sj
• B is the observation probability matrix. Let bi(k) be
the probability of observation θk when the second
stochastic process is in state si
• ∏ is the initial state distribution where πi
represents the probability that the stochastic
process begins in state si
Hidden Markov Model (HMM) is the 5-tuple Δ=<S,θ,A,B,Π>

Weather forecast example

• Given HMM and a sequence of
observations O = {o1 → o2 →…→ ok}, how
to find the sequence of states U = {sk → sk+1
→…→ sk+m} so that U is most likely to have
produced the observation sequence O
• This is the uncovering problem: which
sequence of state transitions is most likely
to have led to this sequence of observations
→ Viterbi algorithm
Uncovering problem

• Each learning style is now considered as
a state
• Users’ learning actions are considered
as observations
• After monitoring users’ learning process,
we collect observations about them and
then discover their styles by using
inference mechanism in HMM, namely
Viterbi algorithm
Basic idea

• Suppose we choose Honey-Mumford model
and Felder-Silverman model as principal
models which are presented by HMM
• We have three dimensions: Verbal/Visual,
Activist/ Reflector, Theorist/ Pragmatist which
are modeled as three HMM(s): ∆1, ∆2, ∆3
respectively
∆1 = 〈 S1, Ө1, A1, B1, ∏ 1〉
∆2= 〈 S2, Ө2, A2, B2, ∏ 2〉.
∆3 = 〈 S3, Ө3, A3, B3, ∏ 3〉.
Basic idea

1. Defining states (S1, S2, S3)
2. Defining initial state distributions
(∏ 1, ∏ 2, ∏ 3 )
1. Defining transition probability matrices
(A1, A2, A3)
2. Defining observations (Ө1, Ө2, Ө1)
3. Defining observation probability matrices
(B1, B2, B3)
Technique includes 5 steps

Learning
objects
selected
Sequence of state transitions → this student is a verbal, reflective and theoretical person.
Sequence of student observations
An example for inferring student’s learning styles

V. Learning history sub-model
1. Providing necessary information for two
remaining sub-models: learning style sub-
model and knowledge sub-model
2. Supporting learning concept recommendation
3. Mining learners’ educational data in order to
discover other learners’ characteristics such
as interests, background, goals…
4. Supporting collaborative learning through
constructing learner groups.
Learning history managed by Mining Engine has
four responsibilities

• Rule-based filtering: manually or automatically
generated decision rules that are used to
recommend items to users
• Content-based filtering: recommends items
that are considered appropriate to user
information in his profile
• Collaborative filtering: considered as social
filtering when it matches the rating of a current
user for items with those of similar users in
order to produce recommendations for new
items
Recommendation methods
V. Learning history sub-model (recommendation)

• Sequential pattern mining belongs to
collaborative filtering family
• User does not rate explicitly items but his
series of chosen items are recorded as
sequences to construct the sequence
database which mined to find frequently
repeated patterns he can choose in future
• In learning context, items can be domain
concepts / learning objects which students
access or learn
Sequential pattern

• Suppose concepts in Java course: data type, package, class & OOP,
selection structure, virtual machine, loop structure, control structure, and
interface which in turn denoted as d, p, o, s, v, l, c, f
• At our e-learning website, students access learning material relating such
concepts in sessions, each session contains only one itemset and is
ordered by time. The student's learning sequence is constituted of
itemsets accessed in all his sessions
Given problem

1. Applying techniques of mining user
learning data to find learning sequential
patterns (not discussed here)
2. Breaking such patterns into concepts
which are recommended to users
Students accessed learning material in their past sessions, how
system recommends appropriate concepts to student for next
visits → mining sequential patterns → solution:

• Suppose the sequential pattern 〈osc(sc)〉
discovered which means:
“class & OOP” → “selection structure” → ”control structure” → ”selection
structure, control structure”
• Pattern is considered as the learning "route"
that student preferred or learned often in past
• In the next time if a student chooses one
concept, the adaptive learning system should
recommend which next concepts?
→ the patterns should be broken into
association rules with their confidence
Problem

1. Breaking entire 〈osc(sc)〉 into litemsets such as o, s, c, (sc)
and determining all possible large 2-sequences whose
order must comply with the order of sequential pattern.
There are six large 2-sequences: 〈os〉, 〈oc〉, 〈o(sc)〉, 〈sc〉,
〈s(sc)〉, 〈c(sc)〉.
2. Thus, we have six rules derived from these large 2-
sequences in form: “left-hand litemset → right-hand
litemset”, for example, rule “s→c” derived from 2-sequence
〈sc〉
3. Computing the confidences of rules and sorting them,
confidence(x → y) = support(〈xy〉) / support((x)). The rules
whose confidence is less than threshold min_conf is
removed
Breaking technique

• If student choose the concept (itemset) x,
system will find whole rules broken from all
sequential patterns and the left-hand litemset
of each rule must contain x
• Then, these rules are sorted by their
confidences in descending order
• Final outcome is an ordered list of right-hand
litemsets (concepts), which are recommended
to students
Recommended list
if user choose concept “class & oop”
Recommended List

• The series of user access in his/her
history are modeled as documents. So
user is referred indirectly to as
“document”.
• User interests are classes such
documents are belong to
There are two new points of view
V. Learning history sub-model (user interest)

1. Documents in training corpus are represented according to
vector model. Each element of vector is product of term
frequency and inverse document frequency. However the
inverse document frequency can be removed from each
element for convenience
2. Classifying training corpus by applying decision tree or
support vector machine or neural network.
3. Mining user’s access history to find maximum frequent
itemsets. Each itemset is considered a interesting
document and its member items are considered as terms.
Such interesting documents are modeled as vectors
4. Applying classifiers (see step 3) into these interesting
documents in order to choose which classes are most
suitable to these interesting documents. Such classes are
user interests
Our approach includes four following steps

• Suppose in some library or website, user U do his search
for his interesting books, documents
• There is demand of discovering his interests so that such
library or website can provide adaptive documents to him
whenever he visits in the next time
• Given there is a set of key words or terms {computer,
programming language, algorithm, derivative} that user U
often looking for, the his searching history is showed in
following table:
User searching history

Using SVM, ANN or Decision Tree to classify this vector
This vector belongs to class compute science → user interest is computer science

V. Learning history sub-model (clustering)
• Individual adaptation regards to each
user
• Community (or group) adaptation
focuses on a community (or group) of
users
There are two kinds of adaptations

• Common features in a group are
relatively stable, so it is easy for
adaptive systems to perform
accurately adaptive tasks
• If a new user logins system,
she/he will be classified into a
group and initial information of his
model is assigned to common
features in his group
• It is very useful if the collaborative
learning is restricted in a group of
similar users
The problem that needs to be solved now is
to cluster user models because a group is a
cluster of similar user models.

Clustering in case that user model is represented
as a vector: Ui = {ui1, ui2,…, uij,…, uin}
The dissimilarity of two user models is defined as Euclidean distance between them
K-means algorithm
2
2
1
2
22
12
2
21
11
2
1
2
1 )
(
...
)
(
)
(
)
,
(
tan
)
,
( n
n u
u
u
u
u
u
U
U
ce
dis
U
U
dissim 








Clustering in case that user model is a overlay model (graph)
The dissimilarity of two graph models 




n
j
j
j
j
v
depth
v
v
G
G
ce
dis
G
G
dissim
1
1
2
1
2
1
2
1
)
(
)
,
(
tan
)
,
(

Clustering in case that user model is a weighted graph
The dissimilarity of two graphs 




n
j
j
j
j
j
v
weight
v
depth
v
v
G
G
ce
dis
G
G
dissim
1
1
1
2
1
2
1
2
1 )
(
*
)
(
)
,
(
tan
)
,
(

Clustering in case that user model is Bayesian network
The dissimilarity of two graph models 




n
j
j
j
j
v
depth
v
v
G
G
ce
dis
G
G
dissim
1
1
2
1
2
1
2
1
)
(
)
Pr(
)
Pr(
)
,
(
tan
)
,
(

• Cosine similarity measure
• Correlation coefficient










n
k
jk
n
k
ik
n
k
jk
ik
j
i
j
i
j
i
j
i
u
u
u
u
U
U
U
U
U
U
U
U
sim
1
2
1
2
1
*
|
|
.
|
|
)
,
cos(
)
,
(












n
j
j
jk
n
k
i
ik
n
k
j
jk
i
ik
j
i
j
i
U
u
U
u
U
u
U
u
U
U
correl
U
U
sim
1
2
1
2
1
)
(
)
(
)
)(
(
)
,
(
)
,
(
K-medoids algorithm and two similarity measures

TLM - Core of PhD research 56
THANK FOR CONSIDERATION
22/01/2022

Triangular Learner Model

More Related Content

What's hot (18)

Similar to Triangular Learner Model (20)

More from Loc Nguyen (20)

Recently uploaded (20)

Triangular Learner Model