In this paper, we study the effect of using domain knowledge
structure on predicting student performance with parameterized Java
programming exercises. Domain knowledge structure defines connections
between elementary knowledge items. While known to be beneficial in
general, it has not been used to predict performance.We compare five different
approaches for this purpose: Bayesian Knowledge Tracing (BKT),
Performance Factor Analysis (PFA), and three dimensional Bayesian
Probabilistic Tensor Factorization (3D-BPTF), that are not able to take
into account knowledge structure; and four-dimensional Bayesian Probabilistic
Tensor Factorization (4D-BPTF) and Feature-Aware Student
Knowledge Tracing (FAST), that can take into account knowledge structure.
We approach the problem using both topic-level and question-level
Knowledge Components (KCs) and test the methods on a dataset of
parameterized questions. Our work is the first in the field that models
students’ behavior in a four dimensional tensor. Our experiments show
that, when having only the knowledge-item-level information, all of the
models work similarly in predicting student performance, but adding
the topic-level information that integrates knowledge items changes the
performance of these models in different directions.
Parameterized Exercises in Java Programming: using Knowledge Structure for Performance Prediction
1.
2. 2Shaghayegh Sahebi (Sherry)
• Programming questions
– Java problems
• Can be designed with parameterized exercises
– One question with multiple parameter sets
– Can be repeated multiple times by one student
• Authoring tool for Java questions
– Create and modify questions
– Indexing service to define concepts inside the
question
3. 3Shaghayegh Sahebi (Sherry)
Each question is generated from a template, and
students can try multiple attempts.
Students give values for specified variable, or give
the output of the code.
A question for practicing skill
nested loops
8. 8Shaghayegh Sahebi (Sherry)
• Students can choose what question to solve
– Using social navigation support
• Adding guidance to the question
– Use the whole set of data to develop personalized
guidance
– Predict how likely the problem will be solved
– Avoid too simple and too complex problems
9. 9Shaghayegh Sahebi (Sherry)
• Predicting the student’s capability to perform
an educational task
• Assumption: the student can learn by
practicing over time by repeating
– Time sequence modeling effect on PSP
• Will present at the Problem Solving & Strategies session
on Monday
– Knowledge structure effect on PSP
• Today’s talk
10. 10Shaghayegh Sahebi (Sherry)
• Questions related to topics, concepts, or skills
– many dimensions in the data
– Structure in the data (knowledge structure)
• Traditional methods: mostly consider student’s
past performance
– Only consider correct/incorrect attempts of students
(ignoring the multidimensionality of the data)
– Bayesian Knowledge Tracing (BKT)
– Performance Factor Analysis (PFA)
12. 12Shaghayegh Sahebi (Sherry)
• Study the effect of knowledge structure
modeling in PSP for parameterized questions
• Compare five approaches:
– Bayesian Knowledge Tracing (BKT)
– Performance Factor Analysis (PFA)
– Feature-Aware Knowledge Tracing (FAST)
– 3D and 4D Tensor Factorization (3D-BPTF, 4D-
BPTF)
13. 13Shaghayegh Sahebi (Sherry)
• Markov Model with two states
• No knowledge structure: Only one type of
knowledge component
• Guess, slip, learning, and initial knowledge
parameters
Knowledge
Structure
14. 14Shaghayegh Sahebi (Sherry)
• Regression model
• No knowledge structure
m(i, j Î KCs
,k Î items,s,f) = bk
+ (gj
si,j
+ rj
fi,j
)
jÎKCs
å
Knowledge
Structure
15. 15Shaghayegh Sahebi (Sherry)
• Extension of BKT
• Can include knowledge structure as regression
variables
fq,l,t
features L
yq,t
kq,t
timesteps T
# of skills Q
fq,e,t
features E
Knowledge
Structure
16. 16Shaghayegh Sahebi (Sherry)
• Tensors: n-dimensional arrays
• Used in collaborative-filtering recommender
systems
– Estimates each tensor as the sum of multiple rank-1
tensors
• Can be extended to as many dimensions
– Can include the data structure
– Each dimension of the data ≈ one dimension of the
tensor
17. 17Shaghayegh Sahebi (Sherry)
• Used successfully in PSP for traditional PSP
• No knowledge structure
• We use Bayesian Probabilistic Tensor
Factorization Model (3D-BPTF) [Xiong et al., 2010]
Students
Questions/ topics
…
Knowledge
Structure
18. 18Shaghayegh Sahebi (Sherry)
• Used for the first time in PSP
• Adds knowledge structure modeling
• Can be extended to more dimensions if
needed
Students
Questions Students Questions
Students
Questions
Attempt 1 Attempt 2 Attempt 3
Knowledge
Structure
19. 19Shaghayegh Sahebi (Sherry)
• From QuizJET system
• Java programming questions
• Six semesters
• 166 students
• 103 questions
• 69.04% majority class (successes)
20. 20Shaghayegh Sahebi (Sherry)
• Here, each topic can have multiple questions and
each question is related to one topic
– Two dimensions: questions and topics
• Study 1: traditional approach
– Question as knowledge unit
• Study 2: considering knowledge structure
– Topic added as knowledge unit
• 5-Fold cross validation
– 80% of students in train data, rest in test data
– User-stratified
33. 33Shaghayegh Sahebi (Sherry)
• Accuracy in predicting students performance
depends on the input of the method
– When ignoring the topic of questions as KCs, all
models perform similarly
– When including topic information, in addition to
the question information, the methods that can
leverage it perform better
34. 34Shaghayegh Sahebi (Sherry)
• Adding the extra topic data in the methods
that cannot model this information decreases
the method’s accuracy
• Knowledge structure can add to the accuracy
of PSP in parameterized questions
• Tensor factorization methods are as good, or
better than the pioneers PSP methods
35. 35Shaghayegh Sahebi (Sherry)
• Include additional structure into tensor
factorization using more dimensions
• Use of other collaborative filtering methods
for PSP
• Test on other programming courses
37. 37Shaghayegh Sahebi (Sherry)
• EM algorithm for BKT and set the initial
parameters as follows: p(L0) = 0:5 , p(G) = 0:2 ,
p(S) = 0:1 , p(T) = 0:3 . For running PFA, we use
• the implementation of logistic regression in
WEKA [3]. For BPTF and BPMF,
• we utilize the Matlab code prepared by Xiong et.
al. We experimented with different latent space
dimensions for BPTF and BPMF (5, 10, 20 and 30)
and chose the best one, which has the latent
space dimension of 10
38. 38Shaghayegh Sahebi (Sherry)
• From collaborative filtering
• Used successfully in PSP for static questions
• No attempt sequence modeling
• We use Bayesian Probabilistic Matrix
Factorization (BPMF)
1 0 0 0
1 1 0 1
0 0 1 1
0 0 0 1
Students
Questions/ topics
0.9 0
1.5 0.4
0 1.4
0 0.9
Students
KCs
0.8 0.5 0 0.3
0 0 0.5 0.8
KCs
Questions/ topics
43. 43Shaghayegh Sahebi (Sherry)
If FAST predicts a success for a student and if BKT predicts a failure
for students, their prediction is more likely to be true compared to
the other methods.
45. 45Shaghayegh Sahebi (Sherry)
BKT and PFA perform similarly to their results in Study 1 and 3D-BPTF
on topics is slightly weaker than 3D-BPTF on questions in terms of
accuracy.
46. 46Shaghayegh Sahebi (Sherry)
• Visual, interactive, adaptive E-learning
platform
– Multi-facet social comparison
– Multi-type learning materials support
– Social navigation
– Personalized guidance
• Integration with other systems with little set
up and modification
Editor's Notes
#attempts per seq = seq length
Visual, interactive, adaptive E-learning platform
Multi-facet social comparison
Multi-type learning materials support
Social navigation
Personalized guidance
Integration with other systems with little set up and modification
Configuration bar
Problem Solving & Strategies session on Monday
Add to motivation: multi dimensional data
Old methods only use correct/incorrect
But we have new approach
Say that we can include any many dimensions in tensors
Compare in a table: time: yes/no,
Add figures here and add time to make it clear
Give a message in your title instead of repeating
Say PSP instead of predicting ….
Change the tables to charts to make them visible
Say one step between TN and conclusion
Minority and majority proportion?
Change the row colors in tables
Seq length instead of attempt per sequence , explain session (or remove the stats table)
Say user stratified or student stratified
Say why we choose question as knowledge component or say knowledge unit
Highlight with different colors the baselines and other methods