Modeling students' Chinese language learning pathways by introducing a behavior-driven semantic graph construction mechanism

BDSGM-Learning Pathways Behavior-Driven Semantic Graph Modeling for Chinese Language Learning Pathways Overview This repository provides a lightweight research implementation inspired by the Behavior-Driven Semantic Graph Model (BDSGM) and its associated strategy for modeling students’ Chinese language learning pathways.The central idea is to integrate behavioral signals, such as time on task, revisit frequency, and assessment performance, with semantic representations of linguistic concepts, constructing a dynamic semantic graph that evolves as learners interact with content. The framework combines behavior-driven graph construction, multimodal encoding, sequence modeling, and graph-based propagation to offer interpretable pathway analytics and personalized learning recommendations. Method Behavior-Driven Semantic Graph Construction A directed concept graph G=(V,E)G = (V, E)G=(V,E) is built where each node is a learning concept, and edges encode transitions inferred from behavior logs. This combines behavioral importance with conceptual movement patterns. Multimodal Encoder and Graphical Propagation A multimodal encoder maps concept and behavior features into embeddings.A graph propagation layer performs message passing with behavior-aware attention coefficients αij\alpha_{ij}αij, reflecting concept influence adjusted by actual learner behavior. Figures in the referenced paper illustrate the modules: Fig. 1: Overall BDSGM structure with encoder and propagation (p. 6) Fig. 2: Multimodal attention fusion blocks (p. 8) Fig. 3: Strategy with progressive graph blocks and mixture-of-experts (p. 9) Fig. 4: Spatial–channel attention details (p. 10) Sequence Modeling and Prediction Temporal dynamics are modeled with a BiLSTM applied to the concept transition sequence.The temporal representation is fused with graph-level embeddings to predict learning pathway stage, mastery likelihood, or related educational outcomes.An optional adaptive threshold converts probabilities into alert signals for early intervention. Architecture at a Glance Behavior-driven edge weights derived from time, frequency, and score signals Semantic concept embeddings for similarity modeling BiLSTM encoder for longitudinal learning sequences Behavior-aware graph propagation for structural refinement Dual-head outputs for pathway classification and auxiliary tasks Datasets and Format The referenced study evaluates the model on four learning datasets: SLAP, SGLE, BALL, and PMLP, demonstrating consistent improvements from behavior-driven propagation. Training Configuration Component Default Optimizer AdamW LR 1e-3 Weight Decay 1e-4 Batch Size 64 Epochs 50 Loss BCE or CE Metrics Accuracy, Macro F1 Efficiency and ROC analyses in the paper (Fig. 6–8, Table 7) show that behavior-aware propagation yields strong accuracy with moderate computational cost. Repository Structure train.py — training entry point model.py — encoder, propagation layers, and prediction heads dataset.py — JSONL loader for concept and behavior sequences utils.py — metrics, checkpointing, random seeds, device helpers inference.py — batch inference and alert generation Results Summary Across conceptual learning datasets, the behavior-driven semantic graph consistently improves F1 scores by several points compared to sequence-only or static-graph baselines.Performance also aligns well with semantic consistency and adaptivity metrics.See Tables 1–4 and Fig. 6 in the paper for detailed comparisons and ROC curves. License Recommended: MIT or Apache-2.0.Include a LICENSE file in the project root. Acknowledgments This implementation is inspired by the behavior-driven semantic graph modeling and associated strategy described in the referenced paper.Figures and formulas correspond to pages 6–10 and 13–21 of the publication.