NeuroDidactica-XAI

# Explainable Deep Learning Models in Adaptive Educational Systems ## 🔍 Overview This repository hosts the implementation of our research paper titled: **"Enhancing Student Learning Outcomes Through Explainable Deep Learning Models in Adaptive Educational Systems"** by **Yang Zheng** and **Miao Yan**, School of Music and Dance, Chengdu University. We propose a novel framework combining formal curriculum representation, an explainable neural model **NeuroDidactica**, and an adaptive strategy **Pedagogical Entropy Minimization (PEM)**. This system improves learning outcomes while maintaining interpretability in adaptive educational technologies. --- ## 📌 Features - 📚 **Formal Curriculum Modeling**: Graph-theoretic and probabilistic representation of curriculum and learner trajectories.- 🧠 **NeuroDidactica**: A recurrent, attention-based architecture for simulating learner cognition.- 📉 **PEM (Pedagogical Entropy Minimization)**: Strategy to optimize instructional decisions by reducing uncertainty in student knowledge.- 🧾 **Explainability-Driven Design**: Interpretable attention mechanisms and alignment-aware forecasting. --- ## 📐 System Architecture The model includes: 1. **Cognitive-State Recurrence** Captures and updates latent learner knowledge with GRU units and contrastive regularization. 2. **Curriculum-Attentive Modulation** Aligns content delivery with learner cognitive state using curriculum-aware attention. 3. **Multi-Level Mastery Forecasting** Combines temporal convolution and curriculum alignment for predicting mastery levels. See *Figures 1 & 2* in the paper for detailed schematics. --- ## 🧪 Datasets We evaluate the proposed system using: - **ASSISTments**- **PISA**- **OULAD**- **EdNet** Each dataset provides diverse educational interactions enabling robust benchmarking. --- ## 🧾 Experimental Setup - Framework: PyTorch 2.0 - Optimizer: Adam (LR = 1e-4) - Temporal Embeddings & Multi-Scale Attention - Evaluation Metrics: MAE, RMSE, MAPE, Accuracy, F1, AUC The experiments were conducted on an NVIDIA A100 GPU with 256GB RAM. --- ## 🥇 Results Our model outperforms strong baselines including BERT, RoBERTa, ELECTRA, and DeBERTa across all datasets. Key metrics:- **Accuracy**: 91.57% (ASSISTments), 91.35% (EdNet)- **F1 Score**: 90.23% (ASSISTments), 90.05% (EdNet)- **AUC**: 92.15% (ASSISTments), 91.84% (EdNet) See *Tables 1–4* and *Figures 5–8* for detailed comparisons and ablation studies. --- ## 🚀 Getting Started ### Requirements ```bashpip install -r requirements.txt