Mathematik  |  Informatik


Joanne Azariah, 2003 | Riehen, BS


This paper explores the use of neural networks in enabling predictive analysis in the education system, allowing for proactive intervention. To provide a holistic view of their potential, a literature review was conducted and networks were trained using real-world data. Finally, a data strategy and risk assessment were provided. While results suggest neural networks have significant potential, hurdles such as data quality remain. However, it is hoped this paper provides schools with a starting point for taking advantage of this technology.


The goal of the work was to answer the question “To what extent can predictive analytics, enabled by artificial neural networks, provide early indicators for teacher support and intervention in the education system?”


The hypothesis was that neural networks have significant potential for predictive analytics in the education system. The work is divided into three sections: literature review, experiment, and data strategy and risk assessment. The review explored specific use cases, limitations, and ways to overcome them by examining research published between 2010 to 2022. The experiment tested their potential using a dataset sourced from a local school which contained the grade averages, class and mother tongue of 1,926 students. The goal was to predict the final grade average per subject. After data processing, network architectures were explored, specifically multi-layered perceptrons (MLP), long-short-term memory (LSTM) networks, convolutional neural networks (CNN) and hybrids (LSTM-CNN). The architecture with the lowest mean average error (MAE) and root mean squared error (RMSE) was selected and learning rate, momentum, batch size and number of layers were optimized. Then, the best-performing network was compared to random forest, linear regression, support vector machine and ZeroR models. Shapley additive explanations (SHAP values) were calculated to determine feature importance, and the impact of dataset size was analyzed. All models were developed using Python libraries such as Keras and Tensorflow. Finally, a data strategy framework was proposed as a foundation to implement neural-network-enabled predictive analytics in schools and risk factors were assessed.


The literature review suggests neural networks are skilled at identifying patterns, spotting temporal dependencies and feature extraction, outperforming other machine learning (ML) models in predicting student performance, drop-out rates and even satisfaction. Yet, they are limited by the quantity and quality of data. To overcome this, digitization and a paradigm shift in the approach to data is key. The experiment found that, among neural networks, LSTMs were most effective. After hyperparameter tuning, LSTMs performed similarly to other ML models. For example, the LSTM, linear regression and random forest all predicted German grades with an average RMSE of 0.35. The SHAP values revealed that non-academic factors such as mother tongue contributed significantly to model results, and holding back ten percent of training data decreased performance. To effectively exploit data, a data strategy with a clear vision that accounts for architecture, security, governance, and maintenance are essential, as are risk mitigations to address policy compliance, data breaches, and change management issues.


The results confirm the hypothesis that neural networks have potential for predictive analytics in education. The research method provided a holistic view of the technology, from use cases to real-life performance to strategies for implementation. Although the research finds data is a limitation, the experiment suggests that even with minimal data, neural networks were able to predict student performance at par with other models. Further work is required to determine which features and how much data is needed to fully exploit their power. Once schools implement a data strategy and address the risks involved, neural networks can begin to be unlocked.


By investing the necessary resources into implementing neural networks, schools can provide the data teachers need to offer students targeted support. Although the initial transition may be challenging, schools should focus on building a comprehensive data strategy to fully unlock their potential. In the future, the key will be linking AI-enabled predictive analysis to prescriptive analytics to suggest learning paths, diagnostic analysis to troubleshoot student progress, and automation of tasks through AI, building a digital learning experience that personalizes education for every student.



Würdigung durch den Experten

Claudiu Musat

This paper explores the fundamentals of the applicability of neural networks in the education system for predictive analysis, allowing for
proactive intervention. The work has undergone many rounds of improvements and developed to something that we can show to schools and justify investments in a data strategy. I commend the student for the sustained and effective work.


sehr gut

Sonderpreis «Swiss Innovation Forum (SIF)» gestiftet vom Eidgenössischen Institut für Geistiges Eigentum




Gymnasium am Münsterplatz, Basel
Lehrer: Michael Hartmann