PaperlayerView exact revision on arXiv

Source-linked AI summary

EduRiskX: A Neuro-Symbolic Framework with F-Logic Reasoning for Early Academic Risk Prediction

Yu Fu, Yongqi Kang, Yong Zhao, Rongfang Bie

arXiv:2608.26107v1cs.AI

TL;DR

Early academic-risk prediction in online education needs systems that detect risk early while remaining interpretable. EduRiskX combines temporal Transformer prediction with F-Logic reasoning, achieving improved recall and earlier detection than tested baselines under identical conditions.

  • Problem

    Online education requires early, reliable identification of at-risk students, but current predictive systems face persistent obstacles in detection and interpretability.

  • Method

    EduRiskX combines a temporal Transformer-based predictor with pedagogically grounded F-Logic reasoning and logistic-regression fusion of their outputs.

  • Results

    EduRiskX achieves improved recall and earlier detection than state-of-the-art time-series and deep-learning baselines, with 0.900 accuracy, 0.894 F1-score, and average detection at Week 9.32.

  • Takeaways & Limitations

    Interpretable early signals linking predictions to behavioral patterns may give instructors more time for supportive measures during formative learning stages.

  • Takeaways & Limitations

    Early-stage activity traces may be sparse or variable, affecting prediction stability in purely data-driven models.

Abstract

from arXiv · show

Predicting students' academic risk in online education is crucial for enabling timely interventions that can improve retention and learning outcomes. However, existing models often suffer from limited early detection capability and insufficient interpretability, leading to a "black-box" trust crisis that hinders their adoption in real-world pedagogical settings. To address these challenges, we propose EduRiskX, a neuro-symbolic framework that integrates a temporal Transformer-based predictor with F-Logic symbolic reasoning. The neural component models longitudinal student activity sequences using temporal attention, class-weighted loss, and dynamic weekly truncation. Acting as a data-driven expert system, an F-Logic rule base -- grounded in established educational theories (Engagement Theory and Student Integration Model) to mimic the diagnostic logic of human educators -- is constructed exclusively from the training data. The neural risk probability and the symbolic confidence score are then combined through a logistic regression-based fusion mechanism that learns the relative contribution of each signal. Experiments on the Open University Learning Analytics Dataset (OULAD) using a strict 80/10/10 student-level split show that EduRiskX achieves an accuracy of 0.900 and an F1-score of 0.894 at the end of the semester (Week 38), with an average early detection week of 9.32 and a detection rate of 94.30 percent. Compared with state-of-the-art time-series models (PatchTST, iTransformer) and common deep learning baselines (LSTM, CNN), EduRiskX yields improved recall and earlier risk identification under identical conditions. Beyond predictive performance, the F-Logic module provides structured rule-based explanations linking predictions to observable behavioral patterns and educational theories.

Introduction

Online education creates a need for early, reliable academic-risk identification, but existing deep-learning models can delay detection and offer limited interpretability. EduRiskX addresses these challenges by combining a temporal Transformer with training-grounded F-Logic reasoning and learnable fusion to support predictive, interpretable, and actionable intervention.

  • Motivation: Online education faces high dropout rates and heterogeneous student performance, making early identification of students needing academic support critical.Learning analytics uses behavioral data from digital platforms to identify students who may require additional support.
  • Challenges: Existing LSTM, CNN, and Transformer-based models often require extensive sequential data, delaying early-stage risk detection and limiting intervention opportunities.Complex neural models also present an interpretability challenge that contributes to a “black-box” trust crisis.
  • Framework: EduRiskX integrates an optimized temporal Transformer with an F-Logic symbolic reasoning module to combine predictive modeling with structured, expert-like diagnostic logic.The neural predictor estimates student-level risk probabilities from longitudinal behavioral sequences, while the symbolic module addresses the black-box trust crisis.
  • Neural component: The framework’s Transformer uses temporal attention, class-weighted loss, and dynamic weekly truncation to model longitudinal behavior and address class imbalance in early prediction.These design choices are presented as an optimized temporal Transformer architecture for early prediction stages.
  • Symbolic reasoning and fusion: EduRiskX derives training-set–grounded F-Logic rules from Engagement Theory and the Student Integration Model, then learns logistic fusion to calibrate neural and symbolic evidence.The rule base is constructed through systematic rule mining that avoids data leakage, and the fusion strategy preserves interpretability while improving early detection performance.
  • Explainability: Structured outputs provide triggered rules, theory-aligned behavioral interpretations, and mapped intervention suggestions for transparent, actionable pedagogical decision support.These outputs are intended to facilitate early intervention in real-world educational settings.

Related Work

Related work spans OULAD-based academic-risk prediction, neuro-symbolic AI in education, and F-Logic reasoning for interpretable systems. Prior studies progress from static classifiers to sequential deep models, while ongoing challenges concern interpretability, early-course performance, and integrating symbolic reasoning with neural prediction.

  • Student academic-risk prediction: OULAD is a widely used benchmark, with early studies applying logistic regression and decision trees to aggregated behavioral indicators.Its scale, longitudinal structure, and realistic representation of distance higher education support its use in educational data mining.
  • Student academic-risk prediction: RNNs, LSTMs, and Transformer architectures increasingly model weekly or event-level learning sequences, with Transformers reporting state-of-the-art results through self-attention.Sequential deep learning generally improves predictive performance compared with static classifiers.
  • Open challenges: Deep models remain difficult to interpret pedagogically, and early-course performance can vary with the amount of available sequential data.These challenges motivate approaches that preserve competitive prediction while improving interpretability and early detection.
  • Neuro-symbolic AI: Neuro-symbolic AI combines neural networks with symbolic reasoning to balance predictive accuracy and interpretability in educational applications.Related approaches include rule-based systems, association rule mining, post-hoc methods such as LIME and SHAP, and tighter neural-symbolic integration.
  • F-Logic reasoning: F-Logic represents objects, attributes, and inference rules declaratively, while its integration with modern deep predictors for academic-risk prediction remains relatively underexplored.EduRiskX addresses this gap by combining a training-set-only F-Logic reasoning module with a temporal Transformer predictor.

Methods · Data Preprocessing and Temporal Feature Construction · Neural Prediction Module (Optimized Transformer)

EduRiskX uses a two-stage “Predict-then-Explain” architecture combining an optimized temporal Transformer with F-Logic reasoning. Its neural module processes weekly longitudinal behavior sequences using temporal attention, class-weighted loss, and dynamic weekly truncation for leakage-free early prediction.

  • Methods: EduRiskX comprises a neural prediction module and a symbolic reasoning module in a two-stage “Predict-then-Explain” architecture.The neural component is an optimized temporal Transformer, while the symbolic component uses F-Logic.
  • Data Preprocessing and Temporal Feature Construction: OULAD event logs are aggregated into fixed-length weekly sequences to align student activity with course structure.The resulting representation is a multivariate time series for each student.
  • Data Preprocessing and Temporal Feature Construction: 38 weeks of OULAD data include content access, assessment activity, forum participation, and engagement-intensity features.Each weekly vector is d-dimensional, and missing weeks are explicitly encoded to preserve temporal continuity for early prediction.
  • Data Preprocessing and Temporal Feature Construction: The weekly temporal representation serves as the common input format for all downstream models, supporting fair comparison across baselines and advanced architectures.This passage describes the unified preprocessing format used for comparative modeling.
  • Neural Prediction Module (Optimized Transformer): The optimized Transformer augments standard self-attention with a learnable temporal weight matrix that emphasizes pedagogically critical weeks and downweights distant history.The temporal weights are jointly trained and enable explicit modeling of long-range temporal dependencies, such as around assessment deadlines.
  • Neural Prediction Module (Optimized Transformer): At-risk students comprise only ∼12% of the population, so the model uses class-weighted cross-entropy to penalize false negatives more heavily and improve recall.Recall is identified as the most critical metric for early warning systems.
  • Neural Prediction Module (Optimized Transformer): Separate model instances are trained for prediction weeks t ∈{5,10,15,20,25,30,35,38}, with each input truncated at its corresponding cutoff.This design simulates realistic early prediction and prevents future-data leakage by restricting learning to information available by week t.
  • Neural Prediction Module (Optimized Transformer): The neural module embeds truncated sequences, adds learnable positional encodings, applies L Transformer encoder layers, and projects the final representation to a sigmoid risk probability.The entire parameterized module is trained end-to-end with the class-weighted loss.

Symbolic Reasoning Module (F-Logic)

EduRiskX’s F-Logic module mines pedagogically grounded academic-risk rules from training data and combines rule-based evidence with frozen neural predictions through learnable logistic fusion. This design supports interpretable, adaptive, and calibrated risk estimation while preventing information leakage.

  • Module functions: The module performs F-Logic rule mining and hybrid inference to improve interpretability and early academic-risk detection.Its two core functions are rule mining and hybrid inference.
  • Rule mining: 120 pedagogically grounded rules remain after confidence-based redundancy removal, which discards rules whose feature sets are subsets of higher-confidence rules.Candidate rules have lengths k = 1 to 4 and are evaluated using support and confidence.
  • Rule mining: All rule discovery uses only the 80% training subset, and the frozen 120-rule base is applied unchanged to validation and test data.This includes support computation, confidence estimation, and threshold selection, preventing validation and test information from entering rule discovery.
  • Hybrid inference: Learnable logistic fusion combines frozen neural and rule-based probabilities with adaptive weighting in logit space, replacing a fixed probabilistic OR operator.The two-stage procedure freezes neural parameters before fusion learning, while learned coefficients reflect the relative reliability of each evidence source and support calibrated estimates.
  • Rule selection: Rule selection combines statistical reliability with pedagogical coverage: primary rules require Support(r) > 0.1 and Conf(r) ≥0.7, while auxiliary rules require theory alignment score ≥0.2.Auxiliary rules additionally require 0.3 ≤Conf(r) < 0.7; triggered-rule increments are bounded between 0.1 and 0.3.

F-Logic Rule Base for Academic Risk Assessment · Behavior-Rule Mapping Framework

EduRiskX uses a 120-rule F-Logic base grounded in Engagement Theory and the Student Integration Model to translate online-learning behaviors into interpretable academic-risk signals. Static, temporal, and time-window rules cover multiple behavioral and temporal scales, supporting early detection when neural predictions are less reliable.

  • F-Logic Rule Base for Academic Risk Assessment: 120 pedagogically grounded rules identify students at risk of poor academic performance, including 16 strongly aligned rules with theory alignment score ≥0.3.The rules are grounded in Engagement Theory and the Student Integration Model.
  • F-Logic Rule Base for Academic Risk Assessment: Three rule categories—Static Pattern, Temporal Sequence, and Time Window—serve distinct diagnostic purposes across behavioral and temporal patterns.The categories cover behaviors such as forum participation, content access, and assessment performance.
  • F-Logic Rule Base for Academic Risk Assessment: Automated theory alignment combines theoretical-keyword coverage with SBERT semantic similarity using the pre-trained all-MiniLM-L6-v2 model to reduce subjectivity in expert scoring.Skeyword measures keyword overlap, while Ssemantic measures cosine similarity between rule-feature and theory-construct embeddings.
  • Behavior-Rule Mapping Framework: The comprehensive rule base spans multiple temporal scales and behavioral dimensions, providing interpretable evidence while maintaining high sensitivity during early course stages when neural predictions are less reliable.Behavior-rule mapping helps educators understand how observable online-learning behaviors translate into risk signals.
  • Behavior-Rule Mapping Framework: Static Pattern Rules screen persistent low engagement using single-feature thresholds, including STAT001 for total_clicks_cat = ’VL’, covering 331,923 students (support = 0.611) with confidence 0.312.Additional static rules target zero forum participation, zero resource access, and zero content engagement.
  • Behavior-Rule Mapping Framework: Temporal Sequence Rules detect declining engagement trends and represent the most reliable risk signals, with confidence ≥0.833 across four core rules.TEMP001 identifies a negative 3-week activity trend with confidence 0.85 and support 0.907 (23,661 students), aligned with Engagement Theory.
  • Behavior-Rule Mapping Framework: Time Window Rules support early-term detection across Weeks 0-4, mid-term detection across Weeks 5-12, and late-term detection across Weeks 13-20.Examples include low early-term forum or content clicks, mid-term declining or inconsistent participation, and late-term persistent engagement, assessment, or social-integration risks.

Risk-Gated Invocation of the Reasoning Engine · Explanation and Intervention Generation

EduRiskX selectively invokes symbolic reasoning for students with elevated or uncertain neural risk and translates final risk probabilities into intervention severity levels. For at-risk students, it produces structured, theory-linked explanations and adaptable intervention recommendations that support educational decision-making.

  • Risk-Gated Invocation of the Reasoning Engine: The F-Logic reasoning engine activates only when predicted neural risk exceeds threshold τ or falls within a borderline uncertainty region.This risk-gated strategy limits detailed reasoning to elevated-risk or uncertain cases while monitoring low-risk students without unnecessary explanations.
  • Risk-Gated Invocation of the Reasoning Engine: EduRiskX maps final risk probability Pfinal into four severity levels, each linked to a recommended intervention intensity.The levels are Critical (Pfinal ≥0.9), High (0.7 ≤Pfinal < 0.9), Medium (0.5 ≤Pfinal < 0.7), and Low (Pfinal < 0.5).
  • Risk-Gated Invocation of the Reasoning Engine: Critical cases receive immediate, high-intensity intervention, whereas High cases receive prompt structured support.Examples include personal outreach or academic advising for Critical cases and weekly check-ins or scaffolded tasks for High cases.
  • Risk-Gated Invocation of the Reasoning Engine: Medium cases receive preventive monitoring and low-touch guidance, while Low cases receive no immediate action beyond regular tracking.Examples include automated reminders and study tips for Medium cases.
  • Explanation and Intervention Generation: For each at-risk student, EduRiskX generates a machine-readable package containing the final risk probability, severity level, triggered rules, associated theories, and intervention suggestions.The package extends beyond a single risk score to support practical educational use.
  • Explanation and Intervention Generation: This structured representation tells educators why a student is flagged and what actions can be taken without requiring interpretation of opaque model internals.It bridges statistical prediction and pedagogical practice through structured, theory-linked outputs.
  • Explanation and Intervention Generation: The explanation package lists triggered F-Logic rules, associated theoretical constructs, a synthesized natural-language reasoning chain, and mapped intervention suggestions.Examples of constructs include low self-efficacy and weak academic integration.
  • Explanation and Intervention Generation: Intervention mappings are maintained separately from predictive logic, allowing domain experts to update recommendations without retraining models.This separation supports adaptation to different institutional contexts and pedagogical strategies.

Algorithmic Workflow of EduRiskX · Risk Assessment Rule Framework

EduRiskX uses a sequential predict–then–reason workflow that combines neural risk probabilities with F-Logic rules to produce predictions, explanations, and interventions. Its three-tier rule framework captures static, temporal, and early-term time-window signals grounded in Engagement Theory and the Student Integration Model.

  • Algorithmic Workflow of EduRiskX: EduRiskX sequentially integrates predictive modeling and symbolic reasoning to generate final risk predictions, activated rules, explanations, and intervention recommendations.The workflow processes weekly student activity sequences with a trained Transformer and F-Logic knowledge base.
  • Risk Assessment Rule Framework: The rule framework has three tiers—Static Pattern Rules, Temporal Sequence Rules, and Time Window Rules—grounded in Engagement Theory and the Student Integration Model.The theories represent behavioral persistence and social-academic integration, respectively.
  • Risk Assessment Rule Framework: 331,923 students are covered by STAT001, which flags total_clicks_cat = ’VL’ with support = 0.611 and confidence = 0.312 for broad initial screening.The rule targets persistent low engagement and serves as a baseline despite its relatively low confidence.
  • Risk Assessment Rule Framework: Students below the low-risk threshold bypass symbolic reasoning, while borderline or higher neural probabilities trigger rule matching and subsequent risk fusion.The pipeline computes Pneural, checks τ and its borderline interval, then initializes rule evidence for applicable cases.
  • Risk Assessment Rule Framework: Activated rules accumulate bounded risk increments, after which logistic fusion combines rule-based and neural signals and links explanations to educational theories.Rule-based risk is capped at 1.0 before final risk computation and intervention mapping.
  • Risk Assessment Rule Framework: 0.85 confidence is reported for TEMP001, identifying negative three-week activity trends with support = 0.907 among 23,661 students.Temporal sequence rules are described as the most reliable risk signals, with confidence ≥0.833, and TEMP001 reflects declining behavioral persistence.
  • Risk Assessment Rule Framework: 20,310 students are covered by TEMP002, which flags decreasing forum clicks as an indicator of poor social-academic integration.This rule operationalizes the Student Integration Model through a declining forum-engagement pattern.
  • Risk Assessment Rule Framework: 100% of students are covered by TW007, which flags low clicks_content with support = 1.0 as a critical Weeks 0–4 engagement-risk indicator.Time Window Rules evaluate low or moderately fluctuating engagement features early in the term; TW001 addresses forum engagement, while TW019 has theory alignment score 0.363.

Results

EduRiskX shows its strongest advantage in early academic-risk detection, maintaining superior recall, balanced accuracy, and precision–recall performance while preserving gains through Week 38. F-Logic contributes to this advantage by improving sensitivity and accelerating identification relative to neural and time-series baselines.

  • Training convergence: 14 epochs mark the early-stopping point after monotonically decreasing training and validation losses initially converged, with later validation loss plateauing.Early stopping was triggered when validation loss showed no further improvement, limiting overfitting.
  • Early-week performance: 0.80+ accuracy by Week 10, versus Week 15–20 for most baselines, and leadership in 5 of 6 Week 5 metrics demonstrate strong early performance.The Week 5 advantage is attributed to F-Logic rules identifying pedagogical risk patterns without requiring extensive statistical correlations.
  • Multi-metric performance: 0.706 balanced accuracy at Week 5 exceeds PatchTST’s 0.635 and iTransformer’s 0.668, while PR-AUC reaches 0.895 at Week 10 and 0.822 at Week 5.EduRiskX leads in F1-Score at Week 5 with 0.694 and maintains a Week 38 F1 advantage of 0.894 versus Ablation’s 0.891.
  • Recall: 0.681 recall at Week 5 exceeds PatchTST’s 0.355 and iTransformer’s 0.390, while Week 10 recall reaches 0.753 against LSTM’s 0.555 and CNN’s 0.593.EduRiskX maintains the highest recall through Week 35, reaching 0.870, and records 0.864 versus LSTM’s 0.819 at Week 38.
  • Early detection timeline: 4.38 weeks earlier than PatchTST and 3.95 weeks earlier than iTransformer, EduRiskX identifies at-risk students on average, with a 1.24-week advantage over Ablation.The average timelines are 9.32 weeks for EduRiskX and 10.56 weeks for the ablation variant.

Case Study

The case studies show that EduRiskX combines neural predictions with F-Logic evidence to identify risk earlier and explain it through theory-aligned behavioral patterns. This supports differentiated interventions while avoiding unnecessary alerts for stable, low-risk students.

  • High-risk case: 0.98 (Pneural) marked Student #27891 as high risk, while eight F-Logic rules identified declining activity, irregular participation, delayed submission, low quiz performance, and absent forum participation.The resulting explanation linked the behaviors to low self-efficacy and poor academic integration, recommending scaffolded assessments, progress check-ins, and peer interaction.
  • Borderline case: Week 8 vs. Week 11 for the neural-only model was the detection timing for Student #32930, whose persistent content-access decline and low quiz participation confirmed risk.The neural predictor assigned 0.65 (Pneural), while the rule-based risk score was 0.55; interventions emphasized interaction frequency and learning routines.
  • Low-risk case: 0.07 was both the neural and final risk probability for Student #26192, with Prule = 0 and no intervention suggested.The case illustrates that the framework preserves trust by avoiding unnecessary complexity or alerts for stable, low-risk students.
  • Overall implication: The cases demonstrate that theory-aligned explanations enable differentiated and timely interventions, while the neuro-symbolic double safety net converts ambiguous early signals into actionable alerts.EduRiskX also mimics human educational diagnostic reasoning, allowing educators to review, approve, or adjust automated intervention suggestions.

Discussion

EduRiskX combines temporal neural prediction with F-Logic reasoning to identify at-risk students earlier while providing structured, pedagogically grounded explanations. The discussion emphasizes temporal detection, interpretability, and hybrid design as complementary priorities for educational early warning systems.

  • Early detection: 9.32 weeks average detection occurred earlier than PatchTST at 15.70 weeks and iTransformer at 13.92 weeks, extending the potential intervention window.The difference may provide instructors additional time for supportive measures during formative learning stages, although intervention effectiveness depends on institutional context and instructional design.
  • Limitations: Sparse or variable early activity traces can reduce prediction stability in purely data-driven models, including advanced Transformer architectures.These models generally benefit from increased behavioral observations later in the course.
  • Hybrid mechanism: Training-set-grounded F-Logic rules and logistic regression fusion combine symbolic confidence with neural probabilities to improve early-week consistency while maintaining competitive overall performance.The symbolic module is grounded in established educational constructs and complements neural predictions with structured rule-based signals.
  • Interpretability: EduRiskX links risk predictions to specific F-Logic rules and pedagogical constructs, providing interpretable evidence about students’ behavioral risk basis.These structured explanations may help educators understand risk alerts.
  • Evaluation criteria: End-of-semester accuracy alone is insufficient because detection timing and temporal stability distinguish models with comparable final metrics.The discussion identifies early detection week and temporal stability as complementary evaluation criteria, while noting that recall gains over neural-only models are modest in absolute percentage terms.
  • Broader implications: Neuro-symbolic modular designs can jointly address predictive performance and interpretability in sequential, potentially sparse educational early warning settings.The discussion presents symbolic guidance and neural representation learning as complementary rather than strictly competing objectives.

Conclusion

EduRiskX combines temporal Transformer prediction with F-Logic reasoning through adaptive fusion to support early academic risk prediction. On OULAD, it delivers strong performance and earlier detection while providing theory-grounded explanations and a path toward broader validation.

  • Framework: EduRiskX integrates a temporal Transformer predictor with F-Logic reasoning, combining neural and symbolic outputs through logistic regression-based adaptive fusion.The framework explicitly separates neural prediction from symbolic explanation while learning each component’s contribution.
  • Results: 0.900 accuracy, 0.894 F1-score, and 0.864 recall demonstrate competitive end-of-semester performance on OULAD under a strict 80/10/10 student-level split.The framework also achieves an average detection week of 9.32.
  • Interpretability and comparison: EduRiskX improves recall and earlier detection over state-of-the-art time-series models and widely used deep learning baselines under identical experimental conditions.Its expert-system design grounds symbolic rules in educational constructs and connects predictions with observable behavioral patterns.
  • Future work: Future work will investigate semi-automatic rule refinement, alternative calibration strategies, and validation on additional educational datasets to assess generalizability.These directions target refinement of the rule base, calibration, and broader empirical validation.

Author contributions statement

Y.F. conceived and developed EduRiskX, conducted the OULAD experiments, and prepared the manuscript materials, while the other authors supervised the research, provided pedagogical grounding, contributed preprocessing and feature construction, and approved the final manuscript.

  • Author contributions: Y.F. conceived the study, developed EduRiskX, and performed all experimental evaluations on the OULAD dataset.Y.F. also constructed the F-Logic rule base and implemented the optimized temporal Transformer and logistic fusion mechanism.
  • Author contributions: Y.Z. and R.B. supervised the research and provided pedagogical grounding for the symbolic rules.
  • Author contributions: Y.K. contributed to data preprocessing and temporal feature construction.
  • Author contributions: Y.F. wrote the main manuscript text and prepared all figures and tables.
  • Author contributions: All authors reviewed and approved the final manuscript.
Loading 2608.26107v1…