Evelance - Evaluating Neural Network Predictions of Student Academic Performance Using Synthetic Data

There is a link between the academic performance of students, their study habits, attendance, sleep, and computer usage; however, forecasting examinations based on these traits is difficult. While machine learning algorithms are capable of finding statistical correlations between student characteristics, it is not enough for the prediction to be reliable for other student profiles or incomplete inputs. In this experiment, the author has studied if a neural network can predict the results of the exam better than a regular regression algorithm and if there are any changes in the performance of the model in controlled experimental conditions. An experiment with a neural network written in PyTorch, consisting of three hidden layers (32, 64, and 128 neurons) has been carried out using 10,000 artificial student records that contained six academic and behavioral characteristics. The experiment included ten tests related to comparing model performances, repeated validation, contributions of different features, errors in subgroups, characteristics of the dataset used, shortcuts in predictions, training-set size, sensitivity to missing data, and controlled changes to model inputs. The neural network reached the result of mean absolute error of 5.7784 and R² of 0.7358, which outperformed the results of the linear regression (mean absolute error – 6.9173) and random forest regression (mean absolute error – 5.8657). In five repeated tests, it gave the mean MAE of 5.7206 ± 0.1821. However, it has proven to be much worse at predicting the results of low-scoring records (mean absolute error 19.7964) than those of the high-scoring records (mean absolute error 4.9707). At 40% randomly missing input values, mean MAE increased by approximately 48.4%.

Authors

Publication Details

Journal
International Journal of Modern Science and Research Technology
Published
2026-11-15
DOI
https://doi.org/10.67317/ijmsrt26nov054
Primary Topic
Online Learning and Analytics
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Evelance - Evaluating Neural Network Predictions of Student Academic Performance Using Synthetic Data

Atharv Narang
International Journal of Modern Science and Research Technology
Online Learning and Analytics
article

Evelance - Evaluating Neural Network Predictions of Student Academic Performance Using Synthetic Data

Atharv Narang
article en

Abstract

There is a link between the academic performance of students, their study habits, attendance, sleep, and computer usage; however, forecasting examinations based on these traits is difficult. While machine learning algorithms are capable of finding statistical correlations between student characteristics, it is not enough for the prediction to be reliable for other student profiles or incomplete inputs. In this experiment, the author has studied if a neural network can predict the results of the exam better than a regular regression algorithm and if there are any changes in the performance of the model in controlled experimental conditions. An experiment with a neural network written in PyTorch, consisting of three hidden layers (32, 64, and 128 neurons) has been carried out using 10,000 artificial student records that contained six academic and behavioral characteristics. The experiment included ten tests related to comparing model performances, repeated validation, contributions of different features, errors in subgroups, characteristics of the dataset used, shortcuts in predictions, training-set size, sensitivity to missing data, and controlled changes to model inputs. The neural network reached the result of mean absolute error of 5.7784 and R² of 0.7358, which outperformed the results of the linear regression (mean absolute error – 6.9173) and random forest regression (mean absolute error – 5.8657). In five repeated tests, it gave the mean MAE of 5.7206 ± 0.1821. However, it has proven to be much worse at predicting the results of low-scoring records (mean absolute error 19.7964) than those of the high-scoring records (mean absolute error 4.9707). At 40% randomly missing input values, mean MAE increased by approximately 48.4%.

International Journal of Modern Science and Research TechnologyVol. 4(11)
Online Learning and Analytics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon