EduRiskX predicts student academic risk early and explainably

Summary

EduRiskX proposes a neuro-symbolic model that aligns a temporal transformer with an F-Logic rule base (derived from Interaction Theory and the Student Integration Model). The model aims to detect at-risk online students earlier than pure deep learning baselines, reporting a 94.3% detection rate and 0.900 accuracy on the OULAD dataset by week 9.

Why it matters

Early and interpretable risk signals remain a persistent gap in learning analytics tools. If the interpretability claim holds, institutions can address why a student is flagged, not just that they are flagged; this is a prerequisite for trust and regulatory compliance. The neuro-symbolic fusion approach also offers a template for other domains where domain-theory rules must coexist with data-driven ranking models.

Highlights

• F-Logic rules are extracted from the training set and grounded in two named educational theories, offering structured, theory-linked explanations instead of opaque attention mechanism weights.
• Logistic regression fusion learns the relative weight of neural probability and symbolic confidence scores, serving as a deliberately simple and auditable combination step.
• On OULAD, the model detects risk at an average of week 9.32, outperforming PatchTST, iTransformer, LSTM, and CNN baselines in both recall and timing with a 94.3% detection rate.

Critical take

The evaluation relies on a single, well-studied dataset (OULAD) without cross-institutional or cross-platform validation, meaning the 94.3% detection rate and 'earlier than SOTA' claim are not yet definitive. Furthermore, since F-Logic rules are extracted from the same training data used to learn the neural network, the 'interpretability' is not truly independent but rather exhibits a circular structure. The logistic regression fusion also appears to be a thin layer on top of a fundamentally two-feature classifier, dressed up in neuro-symbolic terminology.