EduRiskX: A Neuro-Symbolic Framework with F-Logic Reasoning for Early Academic Risk Prediction
EduRiskX is a neuro-symbolic framework combining a temporal Transformer with F-Logic symbolic reasoning for early academic risk prediction in online education The neural component uses temporal attention, class-weighted loss, and dynamic weekly truncation to model longitudinal student activity sequences The F-Logic rule base is constructed exclusively from training data and grounded in Engagement Theory and Student Integration Model to mimic human educator diagnostic logic EduRiskX achieves 0.90
Analysis
TL;DR
- EduRiskX is a neuro-symbolic framework combining a temporal Transformer with F-Logic symbolic reasoning for early academic risk prediction in online education
- The neural component uses temporal attention, class-weighted loss, and dynamic weekly truncation to model longitudinal student activity sequences
- The F-Logic rule base is constructed exclusively from training data and grounded in Engagement Theory and Student Integration Model to mimic human educator diagnostic logic
- EduRiskX achieves 0.900 accuracy and 0.894 F1-score on OULAD, with an average early detection week of 9.32 and a 94.30% detection rate
- The framework outperforms state-of-the-art time-series models (PatchTST, iTransformer) and deep learning baselines (LSTM, CNN) in recall and early risk identification
Why It Matters
This work directly addresses the critical "black-box" trust crisis in educational AI by providing interpretable, theory-grounded predictions that educators can understand and act upon. For AI practitioners, it demonstrates a practical neuro-symbolic fusion architecture that balances predictive performance with explainability — a combination increasingly demanded in high-stakes deployment domains. The approach offers a blueprint for integrating domain knowledge into neural systems without sacrificing data-driven accuracy.
Technical Details
- Neural Component: A temporal Transformer that models longitudinal student activity sequences using temporal attention mechanisms, class-weighted loss to handle imbalance, and dynamic weekly truncation to adapt to varying course structures
- Symbolic Component: An F-Logic rule base constructed exclusively from training data, grounded in established educational theories (Engagement Theory and Student Integration Model), functioning as a data-driven expert system that mimics human educator diagnostic logic
- Fusion Mechanism: Logistic regression-based combination of neural risk probability and symbolic confidence score, learning the relative contribution of each signal rather than using fixed weighting
- Evaluation: Tested on the Open University Learning Analytics Dataset (OULAD) with a strict 80/10/10 student-level split; achieves 0.900 accuracy, 0.894 F1-score at Week 38, average early detection at Week 9.32, and 94.30% detection rate
- Baselines: Outperforms PatchTST, iTransformer, LSTM, and CNN under identical conditions, with particular gains in recall and earlier risk identification
Industry Insight
- Neuro-symbolic architectures represent a viable path toward deploying AI in trust-sensitive domains like education and healthcare, where interpretability is not optional but a prerequisite for adoption
- The fusion strategy using logistic regression for signal combination is elegantly simple and could be generalized to other domains where neural and symbolic outputs need to be reconciled
- Constructing symbolic rule bases exclusively from training data — rather than hand-engineering them — offers a scalable approach to domain-knowledge integration that could reduce the expert bottleneck in neuro-symbolic system development
Disclaimer: The above content is generated by AI and is for reference only.