TLDR: CTGraph is a self-supervised graph-level representation learning approach that profiles student behaviors and performance in curriculum-based online learning systems. It provides a holistic view of student learning journeys, identifies struggling students, and enables comparative analysis of diverse student groups to pinpoint when and where students are facing difficulties, empowering educators with rich insights for targeted interventions.
Intelligent Tutoring Systems (ITSs) are becoming increasingly common in education, from K-12 to higher education and professional training. While these systems offer personalized learning paths, there’s a concern that without proper monitoring, they could widen performance gaps among students. The rise of generative AI also presents challenges, as students might use these tools to complete assignments without genuine understanding, making traditional assessments less reliable. This highlights a critical need for advanced methods to track student progress, identify those who are struggling, and understand their learning behaviors comprehensively.
To address these challenges, researchers have introduced CTGraph, a novel approach that uses graph-level representation learning to profile student behaviors and performance in curriculum-based online learning systems. CTGraph offers a holistic view of a student’s learning journey, considering various aspects like content coverage, learning intensity, and proficiency in different concepts, all while accounting for variations in their learning paths as aligned with the curriculum structure.
How CTGraph Works
CTGraph operates by transforming student learning data into a unique graph representation for each student. Here’s a simplified breakdown:
Data Preprocessing: The system collects data from ITSs, including student learning records and curriculum structures. For each concept a student interacts with, it extracts tracing attributes such as average accuracy, the total number of attempts, and the median week number when the student engaged with the concept.
Graph Construction (Node Absorption): Since students often only attempt a subset of problems, their learning records might have gaps. CTGraph uses a process called “node absorption” to create a personalized “student curriculum-based learning graph.” This involves removing concepts the student didn’t interact with and reconnecting the remaining concepts based on the curriculum’s prerequisite structure. This ensures each student’s graph accurately reflects their unique learning path and engagement.
Graph-Level Encoding with InfoGraph: The core of CTGraph uses InfoGraph, a graph neural network technique, to convert each student’s learning graph into a fixed-length vector. This vector encapsulates rich information about the student’s learning behaviors and performance. The process is self-supervised, meaning it learns patterns without needing manual labeling by educators, making it highly scalable.
Latent-Space Representation and Visualization: These vectorized representations can then be compressed and visualized in a 3D “latent space.” In this space, students with similar learning paths, behaviors, and performance tend to cluster together. This visualization allows educators to quickly identify groups of students, including those who might be outliers or struggling.
Key Findings and Applications
Experiments using a real-world dataset from Adaptemy, an ITS for mathematical learning, demonstrated CTGraph’s effectiveness across various topics like Algebra I, Algebra II, Functions I, and Fractions.
The research showed that CTGraph’s latent representations effectively encode a student’s learning path, behaviors, and performance. By observing the distribution of students in the latent space, educators can identify struggling students—often appearing as outliers with lower average accuracy. The system also allows for identifying groups of similar learners, revealing commonalities in the behaviors of students who are falling behind.
Furthermore, CTGraph enables a fine-grained comparative analysis. Educators can identify “cohort groups” of students who exhibit largely similar behaviors but have subtle differences in specific aspects, such as their engagement with advanced concepts or mastery of foundational knowledge. For example, some students might achieve similar overall performance but show greater effort in tackling more advanced topics, even if their scores on those topics are lower. This nuanced insight helps educators understand not just *who* is struggling, but *when* and *where* in the curriculum they began to face difficulties.
Also Read:
- New AI Model Learns Hidden Concepts to Improve Student Learning and Exercise Recommendations
- New Embedding Techniques Enhance Student Performance Prediction in Learning Systems
Impact on Education
CTGraph represents a significant step forward in student profiling for ITSs. By providing a holistic, evidence-based framework, it empowers educators with rich insights into individual student learning journeys. This understanding can lead to more targeted interventions, personalized feedback, and the development of more effective pedagogical strategies. It emphasizes the importance of evaluating students not just on overall performance, but also on their efforts, learning paths, and progress within the curriculum structure. For more details, you can refer to the full research paper here.
