Exploring the structure of the school curriculum with graph neural networks

Area of Science:

• Educational Technology
• Machine Learning
• Graph Theory

Background:

• School curricula organize knowledge for optimal student learning, often reflecting implicit educational theories.
• Educational assessment data, like student responses, can provide insights into curriculum design and effectiveness.
• Traditional analysis of student-item response matrices often overlooks complex relational patterns.

Purpose of the Study:

• To model student-item response data using a graph neural network (GNN) to uncover underlying curriculum structures.
• To investigate the ability of GNNs to predict student responses and extract meaningful features from educational assessment data.
• To compare the performance of GNNs against classical models in identifying group patterns in student performance.

Main Methods:

• Utilized a sparse student-item response matrix from a Computer-Based Formative Assessment system.
• Represented the data as a bipartite graph with students and items as nodes and responses as edges.
• Applied a graph neural network (GNN) to learn node and edge embeddings for predicting response labels.

Main Results:

• Learned item embeddings captured curriculum properties like item difficulty and subject domain structures.
• The GNN model demonstrated advantages over classical models, especially when group patterns existed in student-item interactions.
• Simulations confirmed the GNN's effectiveness in retrieving structural aspects of the school curriculum from response patterns.

Conclusions:

• Student response patterns in educational assessments contain significant information about curriculum structure.
• Graph-based neural models, specifically GNNs, can effectively retrieve and represent these curriculum structures.
• This approach offers a novel way to analyze and potentially refine educational curricula based on empirical learning data.