Introduction
In the field of education research, access to high-quality student data is often restricted by privacy regulations and the inherent scarcity of large-scale datasets. Synthetic data generation offers a promising solution, allowing researchers to create "digital twins" of educational environments for experimentation without compromising individual privacy. While Generative Adversarial Networks (GANs) have become the standard for synthetic generation in image and audio domains, their performance on structured tabular data—common in student records—remains a subject of debate. This paper evaluates the efficacy of traditional Bayesian Networks (BN) against modern GANs in generating realistic synthetic tabular data for educational use cases.
Methodology
The study compares two fundamentally different modeling paradigms for synthetic data generation:
- Bayesian Network (BN): A probabilistic graphical model that represents a set of variables and their conditional dependencies via a Directed Acyclic Graph (DAG). This approach explicitly models the logical hierarchies inherent in educational data, such as the relationship between aptitude scores, marks, and admission outcomes.
- Generative Adversarial Network (GAN): A deep learning framework consisting of a Generator and a Discriminator. The Generator attempts to produce synthetic samples that mimic the training data, while the Discriminator attempts to distinguish between real and synthetic samples. Through this adversarial process, the Generator theoretically learns the underlying distribution of the data.
Experimental Setup
The researchers utilized an open-source dataset containing 3,029 student records, including high school performance and three years of tertiary Grade Point Averages (GPA). Both the BN and GAN models were used to upsample this data into synthetic datasets of 10,000 samples each. To ensure compatibility with the probabilistic model, variables were preprocessed into discrete categorical formats. The quality of the synthetic data was evaluated using a k-Nearest Neighbor (kNN) classification task; a high-quality synthetic dataset should allow a model to predict student outcomes with accuracy comparable to real-world data.
Results and Analysis
The experimental results revealed a significant performance gap between the two architectures. The Bayesian Network demonstrated a clear advantage in capturing the statistical nuances of the student records.
- Classification Accuracy: The BN achieved a classification accuracy of approximately 75%, with cross-validation scores ranging between 62% and 70%. In contrast, the GAN-generated data resulted in a significantly lower accuracy of only 38%.
- Model Consistency: As shown in the performance metrics, the BN maintained a consistent lead over the GAN across various training sizes. The GAN struggled to converge on a reliable representation of the student journey, likely due to the relatively small size of the initial dataset.
Discussion
The superior performance of the Bayesian Network can be attributed to its ability to explicitly model variable interdependence. In educational contexts, variables often follow a logical hierarchy (e.g., IQ and exam difficulty directly influence marks). By encoding these relationships into the DAG, the BN respects the structural constraints of the data. Conversely, GANs attempt to learn these relationships implicitly. On smaller, structured tabular datasets, the GAN often fails to capture these dependencies, leading to synthetic data that contains significant "noise" and fails to mirror real-world student behavior.
Conclusion
While deep learning models like GANs are powerful tools for unstructured data, this study suggests that "old school" probabilistic models like Bayesian Networks remain superior for generating synthetic tabular data in education. The BN's ability to nearly double the accuracy of the GAN in downstream tasks proves that explicit interdependence modeling is crucial for small-to-medium-sized structured datasets. Future research should investigate automated structure learning for BNs and evaluate GAN performance on larger, more complex educational datasets.
Summary updated