Enterprise AI Analysis

MAGE-KT employs a two-stage process: Heterogeneous Graph Construction and Multi-view Fusion Prediction. It constructs a Multi-relational KC Graph using a multi-agent pipeline and a Student-Question Interaction Graph. These are then used to retrieve subgraphs and fuse signals for accurate prediction, avoiding irrelevant attention diffusion.

A core innovation is the multi-agent pipeline for KC relation extraction. Specialized agents (Semantic, Scoring, Arbitration) generate, score, and adjudicate five types of inter-KC relations (Association, Containment, Equivalence, Sibling, Predecessor-Successor). This ensures semantic validity and structural consistency, enhancing the KC graph.

The Student-Question (S-Q) Interaction Graph captures personalized student-question dynamics. It incorporates IRT-derived student abilities and question difficulties as node attributes, connecting students and questions with correctness labels, and questions/students with similarity weights. This graph provides behavioral signals complementary to the KC graph's semantic signals.

This stage involves student-conditioned subgraph retrieval and Asymmetric Cross-attention Fusion. Relevant subgraphs are retrieved from both the KC and S-Q graphs. The fusion module, with its K-Q and K-S pathways, fully captures interdependencies among students, questions, and KCs, leading to robust prediction.

MAGE-KT was evaluated on ASSIST09, Junyi, and Statics2011 datasets, showing substantial improvements in KC-relation accuracy and next-question prediction (AUC and ACC) over existing DL, Transformer-based, and Graph-based methods. This validates its effectiveness and robust generalization.

Ablation studies confirmed the necessity of MAGE-KT's components. Removing Asymmetric Cross-attention, KC Graph, S-Q Graph, or Subgraph Retrieval consistently lowered accuracy, demonstrating their synergistic contribution. The multi-agent pipeline significantly improved Prerequisite-Successor Extraction accuracy on the Junyi dataset.