ASTRA: An AI Model for Analyzing Math Strategies

Model/Method

AI model demo that analyzes math learning strategies using digital learning data from Carnegie Learning’s MATHia platform.

Visit Resource
This link will take you to an external website.
Purpose/Abstract

ASTRA is a collaborative research project between researchers at the University of Memphis and Carnegie Learning that uses AI to improve understanding of math learning strategies.

The demo uses a pre-trained model based on an architecture similar to BERT. The model learns math strategies from data collected across hundreds of U.S. schools using Carnegie Learning’s MATHia, formerly known as Cognitive Tutor.

This demo focuses on a 7th grade math domain related to ratio and proportions. The fine-tuned model learns to predict which strategies lead to correct or incorrect solutions.

In the demo domain, students worked on word problems involving ratios and proportions and could choose optional tasks to demonstrate their thinking. These optional tasks were based on solving problems using Equivalent Ratios and Means and Extremes, or cross-multiplication.

The demo allows users to select a percentage of schools to analyze and choose among several fine-tuned models, including models based on high-graduation-rate school data, prior skills encoded using Bayesian Knowledge Tracing, temporal features measuring student engagement with MATHia, or a unified model combining prior skills and temporal features.

Results are displayed in a dashboard showing model accuracy using ROC-AUC, outcomes for high-graduation-rate and low-graduation-rate schools, and patterns across different problem types.

The tool also supports generation and visualization of strategies from the AI model, helping researchers analyze how students use optional tasks and how different strategies relate to math learning.

Citation
A S T R A: An AI model for analyzing math strategies.

Areas researched: Student Learning, Platform/Program, Professional Learning, AI

Previous
Previous

Finding Belonging: The Search for Comprehensive Measurements of Student Belonging in Math

Next
Next

Advancing Math Motivation, Engagement, and Persistence: Insights from Digital Interventions and Research-Practice Partnerships