Usingmachinelearningtopredictstudentdifficultiesfromlearningsessiondata资源-CSDN文库

研究论文

174 浏览量 2021-02-12 00:10:15 上传评论收藏 1.24MB PDF 举报

资源推荐

资源详情

资源评论

Artif Intell Rev

https://doi.org/10.1007/s10462-018-9620-8

Using machine learning to predict student difﬁculties

from learning session data

Mushtaq Hussain

· Wenhao Zhu

· Wu Zhang

Syed Muhammad Raza Abidi

· Sadaqat Ali

Abstract The student’s performance prediction is an important research topic because it can

help teachers prevent students from dropping out before ﬁnal exams and identify students that

need additional assistance. The objective of this study is to predict the difﬁculties that students

will encounter in a subsequent digital design course session. We analyzed the data logged

by a technology-enhanced learning (TEL) system called digital electronics education and

design suite (DEEDS) using machine learning algorithms. The machine learning algorithms

included an artiﬁcial neural networks (ANNs), support vector machines (SVMs), logistic

regression, Naïve bayes classiﬁers and decision trees. T he DEEDS system allows students to

solve digital design exercises with different levels of difﬁculty while logging input data. The

input variables of the current study were average time, total number of activities, average

idle time, average number of keystrokes and total related activity for each exercise during

individual sessions in the digital design course; the output variables were the student(s) grades

for each session. We then trained machine learning algorithms on the data from the previous

session and tested the algorithms on the data from the upcoming session. We performed

k-fold cross-validation and computed the receiver operating characteristic and root mean

square error metrics to evaluate the models’ performances. The results show that ANNs

and SVMs achieve higher accuracy than do other algorithms. ANNs and SVMs can easily

be integrated into the TEL system; thus, we would expect instructors to report improved

student’s performance during the subsequent session.

Keywords Machine learning · Educational data mining (EDM) · Decision support tools ·

E-learning · Neural networks (NN) · Support vector machine (SVM)

Wenhao Zhu

whzhu@shu.edu.cn

School of Computer Engineering and Science, Shanghai University, Shanghai, China

123

M. Hussain et al.

1 Introduction

The educational advantages of e-learning include online teaching and course delivery, which

do not require physical classrooms for students. Compared to traditional modes of learning,

e-learning is less expensive, and a larger number of students can register for online c ourses.

However, in e-learning, there is no direct communication between students and teachers.

Therefore, e-learning poses some challenges. First, it is difﬁcult for instructors to assess the

effectiveness of a course. Second, the dropout rate of students in e-learning courses is much

higher than that in traditional modes of learning. Third, assessing student’s performance is

difﬁcult. Fourth predicting at-risk students in new courses is also difﬁcult. Finally, teachers

are interested in predicting students’ expected results on upcoming assessments (Lykourent-

zou et al. 2009; Pahl and Donnellan 2002; Smith-Gratto 1999; Kuzilek et al. 2015; Bakki

et al. 2015).

Web-based learning environments such as massive open online courses (MOOC), digital

electronics education and design suite (DEEDS) and learning management systems (LMSs)

allow teachers to study student performances using logged student data, but teachers may

have difﬁculty analyzing the student logs. MOOCs and LMSs are popular types of web-based

learning platforms; they provide free higher education to the entire world and offer courses

from different universities. Furthermore, they provide administration, documentation, content

assembly, student management and self-services (He et al. 2015). LMSs are online portals

for both students and teachers that facilitate teacher-student interactions and allow them to

perform their educational tasks and activities. More-over LMSs deliver courses to students,

and the students can select their own courses through a course selection process (Imran et al.

2014). MOOCs are free web-based learning platforms that supply all their courses online.

Students can register and attend these courses from any location (Kloft et al. 2014). These

web-based learning environments affect how teachers and students think during class, and

they can be used to predict a student’s performance during the next class or a student’s

behavior at different times. In addition, these environments can be used to improve course-

related content (Chen et al. 2000).

Predicting a student’s progress in a class or session through, for example, quizzes, assign-

ments, exams, and session activities can provide instructors with in-depth information on the

progress of students throughout the course. To achieve this goal, researchers have applied

various machine learning and statistical techniques to data acquired from both traditional and

online universities.

In traditional universities, researchers mostly use a student’s educational history (e.g., quizzes,

midterm exams, degrees, and attended schools) and demographic information (country, sex,

race and zip code) to predict student’s performance.

Acharya and Sinha (2014) forecast students’ performances using machine learning tech-

niques (e.g., C4.5, sequential minimal optimization (SMO), Naïve bayes, 1-NN (1-Nearest

Neighborhood), and MLP (multi-layer perceptron) with input features (e.g., gender, income,

board marks and attendance). They applied correlation-based feature selection (CBFS) tech-

niques to improve the model performances and determined that SMO achieves a higher

effective average testing accuracy (66%) than do other methods.

De Albuquerque et al. (2015) employed artiﬁcial neural networks (ANNs) to predict

student’s performance. These models achieved high accuracy (85%) using input features

such as grades, periods of study and school scores.

123

Using machine learning to predict student difﬁculties…

Marbouti et al. (2016) used logistic regression, support vector machines (SVMs), decision

trees (DTs), ANNs and a Naïve bayes classiﬁer (NBC) to identify at-risk students in advance

of the next course. This study used input features, such as grades, attendance, quizzes, weekly

homework, team participation, project milestones, mathematical modeling activity tasks, and

exams from an ofﬂine course. Analysis of the results found that the NBC algorithm provided

satisfactory accuracy (85%).

Huang and Fang (2013) performed a study that used machine learning techniques to

predict student academic performance in engineering courses. In this study, the input fea-

tures included course grades from all semesters and the output variable was exam scores.

The researchers observed that SVMs are suitable for predicting an individual student’s per-

formance and that multilinear regression is suitable for forecasting the performance of all

students in a course.

Abu Saa (2016) performed a study to ﬁnd the best classiﬁer to predict student’s perfor-

mance in higher education using social and personal input features.

Some probabilistic models (i.e., Bayesian knowledge tracing) have been used to predict

student’s performance by analyzing logs compiled during student computer gameplay (Käser

et al. 2017). However, these models do not predict hidden patterns of students.

Furthermore, in traditional universities, some statistical methods have been used to predict

student’s performance; these include linear mixed-effect models (LMEM) and survival analy-

sis techniques that use variable multimodal data (heart rate, step count, weather condition and

learning activity) as input along with cumulative student pre-enrollment and semester-wise

information (Di Mitir et al. 2017; Ameri et al. 2016).

Currently, most universities provide courses using e-learning systems accessible from

any location. Scientists use input features common to these e-learning systems (e.g., time,

activity, assessment and online discussion forums) to forecast student performance.

Kotsiantis et al. (2003) predicted student’s performance on ﬁnal exams using machine

learning techniques (e.g., Naïve bayes, 3-NN, RIPPER C4.5 and WINNOW). They used

demographic features as inputs (e.g., sex, age, marital status, and number of children) along

with performance-related input features (e.g., meetings and assignment grades) from an e-

learning system and found that the Naïve bayes approach achieved a higher average accuracy

(73%) than did the alternatives.

Hu et al. (2014) developed a student warning system using e-learning system features such

as course login time, average login time and delay in reading the assignment. They found

that C4.5 and CART achieved satisfactory accuracy (93 and 94%, respectively).

Kaur and Kaur (2015) examined student difﬁculties in a course on mathematics, system

analysis, and design using data mining techniques. They used test grades as input features and

determined that AdaBoost was the best classiﬁer for predicting the difﬁculties that students

would experience in subjects.

Vahdatetal.(2015) used process mining (PM) and complexity matrix (CM) methods to

analyze the relationship between grades and students’ learning processes using DEEDS data.

They concluded that the average student grades are positively correlated with the CM and

that difﬁculty is negatively correlated with the CM. In addition, they determined that process

discovery using PM and CM models provides valuable information regarding student learning

processes.

Chen et al. (2000) used database systems and decision tree techniques on e-learning system

logs to check the performance of students using an approach helpful for teachers.

Hlosta et al. (2017) introduced a self-learning system using machine learning algorithm

to ﬁnd at-risk students in a new course without any previous history data. This study demon-

strated that XGBoost achieved the best performance.

123

M. Hussain et al.

He et al. (2015) provided early predictions of at-risk students in a MOOC course using an

LR technique by analyzing assignment and lecture features.

Arnold and Pistilli (2012) developed a learner analytics system that allowed teachers

to give real-time support to students and solved student retention problems. Moreover, this

system depends on student demographic characteristics, past academic history, student efforts

and student grades. The results showed that students using the analytics system achieved

higher grades compared to those who did not use the system.

Liu and d’Aquin (2017) used a supervised learning algorithm to p redict student’s perfor-

mance. They investigated how demographic variables and online learning activities affect

student’s performance. Furthermore, they used the k-prototypes clustering algorithm to ﬁnd

the group of weak students who needed additional help from the teacher. They concluded

that the successful groups of students mostly came from privilege and most of these students

complete their higher education.

The authors Kai et al. (2017) used the J-48 and J-Rip classiﬁers to identify students who

do not continue past the course orientation stage and found that these models provide good

information to teachers that can aid in student retention.

Another study Elbadrawy et al. (2015) predicted student grades using a collaborative

multi-regression model based on students’ performances, activities and Moodle interactions

as features. The results revealed that the performance of a collaborative multi-regression

model using Moodle interaction features is comparable to that of a linear regression model.

Studies have also been conducted that use ANNs with only slight modiﬁcations to classify

students based on to their ﬁnal grades using web-based education system features (VOD-

watching times, courseware download times and BBS posting times) (Zheng et al. 2013).

Some commonly used machine learning techniques have been investigated to predict student’s

performance and identify at-risk students in e-learning course (Kuzilek et al. 2015).

Recently, an early predictive model was developed using student demographic, LMS

data, and aptitude-related features. The authors developed a learning analytic system with an

applied LR model that sent emails to high-risk students (Jayaprakash et al. 2014).

The majority of early studies that focused on predicting student grades and learning

behaviors in upcoming course sessions used datasets from traditional universities and e-

learning systems. However, their input features did not reﬂect the students’ performances

during in-session problem-solving exercises or projects. None of the existing studies predicted

students’ performances in a technology-enhanced learning (TEL) domain. Most studies used

academic input features (e.g., GPA (grade point average) and grade and semester marks)

and non-academic input features (e.g., age and gender), which are less effective for making

timely predictions concerning student’s performance in a TEL system. Moreover, it may be

costly to collect these data. Some of the early studies that used training and test data from

the same course suffered from the same difﬁculty; such methods do not help the teacher

correctly evaluate the model accuracy in the succeeding session because student course

outlines and activities change from one class to another. Performance prediction regarding

future coursework sessions based on log data is not a straightforward task because every

session has its own difﬁculties and unique problems. Prediction also depends on course

features and teaching techniques; thus, it is important to build an intelligent TEL data system

that forecasts the difﬁculty of the upcoming session.

The ﬁrst step in improving student learning is being able to predict the difﬁculty that

students will have with the subsequent class session. Predicting student difﬁculty for the next

session using DEEDS logs is important to both the instructor and to the students in MOOCs

and TEL systems. However, because teachers of MOOCs and TEL systems are not machine

learning experts, they cannot easily interpret the DEEDS log data. A data interpretation

123

Using machine learning to predict student difﬁculties…

feature can be easily integrated into a TEL system or a MOOC to identify the students’

difﬁculties, improve their learning performances, and prevent performance degradation in

the subsequent session. In addition, such predictions allow the instructor to use the DEEDS

logs in the learning model to determine the probability that a student will encounter difﬁculties

in the next session and to provide feedback to the student in real time. Overall, by using this

method, the instructor is better able to prepare students who experience difﬁculties before

they start their subsequent sessions. Thus, this approach is expected to increase retention and

provide advance information about the challenges that individual students experience.

This study used machine learning algorithms to predict individual students’ difﬁculties

in the subsequent session of a TEL system when the students performed different activities

(problem-solving exercises, laboratory assignments, reading course-related materials, etc.)

during the course session. These data can be easily integrated into DEEDS and MOOCs to

assist teachers in identifying potential student difﬁculties in upcoming sessions. There does

not appear to be any related prior research on using TEL and machine learning techniques to

predict student difﬁculties in a subsequent session of a digital design course.

In the current study, we used log data obtained from the DEEDS (https://www.

digitalelectronicsdeeds.com), a TEL tool and virtual digital electronic laboratory used by

instructors at the University of Genoa, Italy, both in and out of the classroom to improve

student learning. Students remember concepts better when reading course-related materials

using the TEL system than when reading course-related materials without the TEL system

(Vahdatetal.2015). Additionally, the DEEDS is an e-learning environment used by stu-

dents to complete various laboratory assignments in electronic and information engineering

classes at the University of Genoa (Donzellini and Ponta 2007). By applying the DEEDS to

massive open online courses (MOOCs) and learning management systems (LMSs), teachers

can easily track students activities and provide students with news, guidelines and feedback

(Donzellini and Ponta 2007).

Our main goals were as follows:

• To identify the most appropriate machine learning algorithms for predicting the difﬁculty

an individual student would have in the next session of a digital design course based on

prior session activities and the current session.

• To investigate which machine learning algorithms used in the current study are appro-

priate for predicting student difﬁculty in the next session of digital design course while

using the fewest features.

The results of the current study show that SVMs and ANNs are appropriate machine learning

models to predict a student’s performance as well as the difﬁculty a student will experience

over the entire next session in a digital design course. The remainder of this paper is organized

as follows: Section 2 includes problem formulation. Section 3 describes the materials and

methods, Sect. 4 presents the experimental results, and Sect. 5 presents conclusions and

describes future work.

2 Problem formulations

The DEEDs is a technology-enhanced learning and virtual digital electronic laboratory used to

improve student learning. The problem of predicting student difﬁculty in the DEEDs involved

investigating the most appropriate machine learning algorithm to predict student difﬁculties in

terms of the grades they would earn in the subsequent session of digital design course exercises

123

剩余26页未读，继续阅读

评论收藏

内容反馈

weixin_38624914

粉丝: 7
资源: 950

Using machine learning to predict student difficulties from lear...

最新资源

Using machine learning to predict student difficulties from lear...

IntentNet- Learning to Predict Intention from Raw Sensor Data.pdf

Python Machine Learning

IntentNet- Learning to Predict Intention from Raw Sensor Data

Hands-On Data Science and Python Machine Learning

Python Machine Learning Blueprints 2nd Edition

Python Machine Learning By Example [2017].azw3电子书下载

Introduction to Machine Learning第二版

Python Machine Learning.pdf 无水印书签修正完美版 2015

Machine Learning by Tutorials：Beginning machine learning for Apple and iOS.pdf

Machine Learning by Tutorials: Beginning machine learning for Apple and_code.zip

Machine Learning by Tutorials: Beginning machine learning for Apple and_code.z01

Kevin P. Murphy. Machine learning A Probabilistic Perspective

Introduction to Machine Learning 英文书

基于CORDIC的反正弦和反余弦计算的FPGA实现

使用3DCNN和卷积LSTM进行手势识别学习时空特征

BA无标度网络中的SIR模型

基于三次贝塞尔曲线的类汽车曲率连续路径平滑

基于机器学习的设备剩余寿命预测方法综述

基于FPGA的奇异值和特征值分解的快速实现。

基于BP神经网络的人口预测

磁悬浮系统自适应模糊PID控制器的设计

最新资源