Reproducibility Project of Enhancement of E-Learning Student’s Performance Based on Ensemble Techniques by Alsulami et al. (2023, Electronics)

Amy Tan

Project Overview

This project takes inspiration from the open science movement and consequently attempts to reproduce the results from the 2023 paper, Enhancement of E-Learning Student’s Performance Based on Ensemble Techniques by Abdulkream A. Alsulami, Abdullah S. AL-Malaise AL-Ghamdi, & Mahmoud Ragab (2023, Electronics), which can also be found in the original_paper folder.

Aluslami et al. ran 15 experiments with WEKA software, testing traditional data mining/machine learning classifiers separately, and with ensemble methods to examine which models were the best predictors of student performance. They concluded that boosting with decision trees yielded the highest accuracy out of all the experiments. I aimed to reproduce their findings in this project, using Python's skit-learn package.

Installation and Setup

Codes and Resources Used

• Editor Used: Anaconda

• Python Version: Python 3.11.5

• Full notebook and code can be found in the writeup folder

Python Packages Used

• Data Manipulation: pandas
• Data Visualization: matplotlib
• Machine Learning: sklearn

Data

Source Data

The data used in this project is xAPI-Edu-Data.csv, initially downloaded from where the authors' originally uploaded on Kaggle, saved into the variable edu in the Juypter Notebook. This csv is saved in the data folder.

Cleaned Data

All categorical variables in the dataset were cleaned and transformed into corresponding numerical values to run subsequent classifier models - this dataframe is saved in the data folder under the name edu_dummy (and it stored under the same variable name in the notebook). The dictionary of cleaned/transformed variables is as follows for reference:

Column	Category	Numerical Representation
Gender	Male	0
	Female	1
Nationality	Kuwait	0
	Lebanon	1
	Egypt	2
	Saudiarabia	3
	Usa	4
	Jordan	5
	Venzuela	6
	Iran	7
	Tunis	8
	Morocco	9
	Syria	10
	Palestine	11
	Iraq	12
	Lybia	13
Stageid	lowerlevel	0
	middleschool	1
	highschool	2
Gradeid	G-02	2
	G-04	4
	G-05	5
	G-06	6
	G-07	7
	G-08	8
	G-09	9
	G-10	10
	G-11	11
	G-12	12
Sectionid	A	0
	B	1
	C	2
Topic	IT	0
	Math	1
	Arabic	2
	Science	3
	English	4
	Quran	5
	Spanish	6
	French	7
	History	8
	Biology	9
	Chemistry	10
	Geology	11
Semester	F	0
	S	1
Relation	Father	0
	Mum	1
Parentansweringsurvey	No	0
	Yes	1
Parentschoolsatisfaction	Bad	0
	Good	1
Studentabsencedays	Under-7	0
	Above-7	1
Class	M	0
	L	1
	H	2

Results and evaluation

Ultimately, I was unable to reproduce their findings - out of all 15 experiments, I found that boosting with decision trees, naive bayes, and random forests resulted in the most accurate predictions of student performance (.7998). However, it is also worth noting that this combination of ensemble and data mining techniques was just better than random forests alone (.7973).

References

• Parts of code adapted from Umberto Mignozzetti's CSS 201/202 course - umbertomig@ucsd.edu
• Alsulami, Abdulkream & AL-Ghamdi, Abdullah & Ragab, Mahmoud. (2023). Enhancement of E-Learning Student’s Performance Based on Ensemble Techniques. Electronics. 12. 1508. 10.3390/electronics12061508.

Future work

• Rerun the Juypter Notebook with a different random state to see if/how that impacts models' performances and what model performs the best.
• Explore applying these models to different datasets to examine performance; apply these models to predict other areas in education, beyond academic success (i.e. reading comprehension, social and emotional learning, critical thinking skills, creativity, problem-solving abilities,etc.).
• Examine and consider implications and consequences of applying/using these models in education.

License

MIT License