Towards Responsible and Trustworthy Educational Data Mining: Comparing Symbolic, Sub-Symbolic, and Neural-Symbolic AI Methods

Different AI methods prioritize different factors when making predictions about student performance: Symbolic AI focused on cognitive and motivational factors, while sub-symbolic (neural network) methods emphasized cognitive aspects and demographics like gender, but crucially, both largely overlooked metacognitive factors—the ability to think about one's own thinking—which may be important for understanding student learning.

Neural-symbolic AI offers a more balanced and comprehensive approach: By combining symbolic and neural methods, this hybrid approach considered a broader range of factors including motivation, metacognition, and learned knowledge together, providing a more theoretically grounded explanation for its predictions compared to using either method alone.

Standard performance metrics can be misleading in educational AI: While symbolic and sub-symbolic methods performed similarly on balanced datasets, they failed to reliably identify low-performing students when data was imbalanced (when there were fewer examples of struggling students). The neural-symbolic method performed better in this realistic scenario, suggesting that predictive accuracy alone doesn't tell the full story about an AI system's trustworthiness.

Responsible AI in education requires examining how predictions are made, not just whether they're accurate: The study found that different AI methods reached similar predictions through different reasoning processes, highlighting that educators and policymakers need to understand the factors driving AI recommendations before deploying them in schools.