Using artificial neural networks to predict first-year traditional students second year retention rates

Abstract

This research investigates the use of Artificial Neural Networks (ANNs) to predict first year student retention rates. Based on a significant body of previous research, this work expands on previous attempts to predict student outcomes using machinelearning techniques. Using a large data set provided by Columbus State University's Information Technology department, ANNs were used to analyze incoming first-year traditional freshmen students' data over a period from 2005-2011. Using several different network designs, the students' data was analyzed, and a basic predictive network was devised. While the overall accuracy was high when including the first and second semesters worth of data, once the data set was reduced to a single semester, the overall accuracy dropped significantly. Using different network designs, more complex learning algorithms, and better training strategies, the prediction accuracy rate for a student's return to the second year approached 75% overall. Since the rate is still low, there is room for improvements, and several techniques that might increase the reliability of these networks are discussed.

Interpretable Deep Learning for University Dropout Prediction

“…The performance of our models is in line with the ones reported in related works; for example, Plagge achieved an accuracy of 75.7% with their best neural network model for student retention [28], while Alkhasawneh and Hobson could classify 70.1% of the students correctly [3].…”

“…The first neural network model is a simple fully connected deep neural network (FCNN), which is able to achieve relatively high performance on tabular datasets as the cited works present [3,28,39]. A recent paper [4] proposes a novel interpretable canonical deep tabular data learning architecture, called TabNet, and the authors claim that it can even outperform decision tree based machine learning models on a wide range of tabular data, thus our second deep neural network model utilizes TabNet architecture.…”

“…Various other network architectures (e.g. denoising autoencoders) have been used to predict student retention and identify at-risk students [3,20,28], nevertheless a common conclusion can be drawn: deep learning models frequently outperform traditional machine learning models for student dropout prediction.…”

Estudio descriptivo de variables sociodemográficas y motivacionales asociadas a la deserción: la perspectiva de personas universitarias de primer ingreso

“…Tales estudios se han ejecutado analizando el fenómeno en diferentes periodos, desde solo el primer año universitario hasta 5 años y más (Mesarić y Šebalj, 2016;Miranda y Guzmán, 2017), como en el caso de la Universidad Nacional, en donde se determinó un máximo de 18 ciclos lectivos, o sea 9 años como periodo de observación para que ocurra la deserción (Rodríguez y Zamora, 2014). Igualmente, los criterios para definir la deserción han sido variantes, desde dejar de matricular en el primer o segundo ciclo lectivo del primer año (Rodríguez y Zamora, 2014;Plagge, 2013;Aguiar, Ambrose, Chawla, Goodrich y Brockman, 2014y Oñate, 2016. En esta investigación se asumió el criterio de la Comisión Nacional de Rectores de Costa Rica (CONARE), de contar con dos años consecutivos de no matrícula en la universidad para ser clasificado como estudiante desertor, ello según el esfuerzo conjunto entre la Cátedra UNESCO y el CONARE, desde el año 2008.…”

A Comparative Analysis of Selected Studies in Student Performance Prediction

“…A generic approach is to predict an outcome that is binary in nature. Some of the studies reviewed follow this approach to predict overall outcome such as passing or failing a course or even forecasting successful completion of college as marked by graduation ([19], [20], [21], [22], [23]). Others have taken a further step in predicting the classification of the degree or achievement ( [24], [25]).…”