SLIDE 1
Exercise-Enhanced Sequential Modeling for Student Performance Prediction
The 38st Association for the Advancement of Artificial Intelligence (AAAI'17) 2018/02/02 - 02/07, New Orleans, Louisiana
- Outline
Exercise-Enhanced Sequential Modeling for Student Performance - - PowerPoint PPT Presentation
Exercise-Enhanced Sequential Modeling for Student Performance - - PowerPoint PPT Presentation
The 38 st Association for the Advancement of Artificial Intelligence (AAAI'17) 2018/02/02 - 02/07, New Orleans, Louisiana Exercise-Enhanced Sequential Modeling for Student Performance Prediction
SLIDE 2
SLIDE 3
!
Background- Online Education System
Online education systems provide students with open access for self- learning (e.g., learning remedy suggestion and personalized exercise recommendation).
SLIDE 4
)
Background- Predict Student Performance
How can we say an education system understand a student?
- (
)
SLIDE 5
- Research Problem
l Urgent issue: Predict Student Performance(PSP)
l How to automatically predict student performance without manual
intervention?
l This work regards score as performance.
l Opportunity
l Exercises records of students
l Text materials of exercise
SLIDE 6
2
Challenge 1 for PSP
l Diverse expressions of exercises l Need a unified way to understand and represent them automatically.
211
SLIDE 7
Challenge 2 for PSP
l Long-term historical exercising
!" !# !$ !%
- !%&"
SLIDE 8
- Challenge 3 for PSP
l Cold start problem
SLIDE 9
Related Work for PSP
l Education Psychology
l
IRT (Item Response Theory)
l
models student exercising records by a logistic-like function
l
BKT (Bayesian Knowledge Tracing)
l
traces them with a kind of hidden Markov model
l Machine Learning and Data Mining
l
PMF
l
projects students and exercises into latent factors
l Deep Learning
l
DKT
l
deep learning method uses RNN to model student exercising process for prediction .
SLIDE 10
- Outline
SLIDE 11
- Problem Definition
- Given: the exercising records of each student and the text
descriptions of each exercise from 1 to T
- Goal: train a unified model M, predict the scores on the next exercise
!"#$ of each specific student.
SLIDE 12
- Outline
SLIDE 13
- Exercise-Enhanced Recurrent Neural Network(EERNN)
SLIDE 14
Step 1: Word Embedding
l
Main procedures:
l Word split l Latex to feature l Word to vector
l
Goal: learn word representations from semantic perspective in math exercise.
SLIDE 15
- Step 2: Exercise Embedding
- Goal: Exercise Embedding learns the semantic representation of each exercise
!" from its text input #" automatically.
SLIDE 16
- Step 3: Student Embedding
l
Goal: Student Embedding aims at modeling the whole student exercising process and learning the hidden representations of students.
!" 0000 … 000000 !" 0000 … 000000
Combine the score and embedding %&
SLIDE 17
- Step 4: Prediction (Two Strategies)
- Goal: Predicting her performance on exercise !"#$ at step T +1.
SLIDE 18
- Step 4: Prediction (Two Strategies)
- ……
- Predicting her score on exercise "#$% is &̃#$%
SLIDE 19
- Step 4: Prediction (Two Strategies)
- Goal: predicting her performance on exercise !"#$ at step T +1.
SLIDE 20
- Outline
SLIDE 21
Experiments
l Experiments dataset
l Supplied by Zhixue, IFLYTEK l a widely-used online learning system, which provides senior high school
students with a large exercise resources for exercising.
SLIDE 22
Experiments
l Baseline methods
l Variants of EERNN: LSTMM, LSTMA
l just utilize knowledge-specific representations l To validate the importance to incorporate exercise texts for the prediction in
EERNN
l Education Psychology: IRT, BKT l Machine Learning and Data Mining: PMF l deep learning method: DKT
l The most similar method to ours
l Evaluation metrics
l Regression perspective: RMSE l classification perspective: ACC, AUC
SLIDE 23
Experiments
SLIDE 24
Experiments
Information can be used by EERNNM More Information can be used by EERNNM Old Information forget by EERNNM Focus useful information by EERNNA
SLIDE 25
- Experiments
SLIDE 26
- Outline
SLIDE 27
Conclusion
l Proposed a novel EERNN framework to predict student future
performance.
l EERNN integrated two critical components, BiLSTM to extract
exercise semantic representations from texts, LSTM architecture to trace student states.
l Proposed two strategies for prediction : EERNNM with Markov
property and EERNNA with Attention mechanism.
l Experiments on real-world dataset demonstrated the effectiveness
(specially cold start problem) of EERNN .
SLIDE 28
Future Work
l Different exercise types (e.g., the subjective exercises
with continuous scores)
l Incorporate more information
l Knowledge concepts l The time cost on exercises
l Integrate some educational theories
l learning and forgetting curves l Guess and slip
SLIDE 29