Novel Considerations in the ML/AI Modeling of Large-Scale Learning Loss

Abstract

This study is a path forward for the large-scale, data-driven quantitative analysis of noisy open-source data resources. The goal is to support qualitative findings of smaller studies with extensive open-source data-driven analytics in a new way. The study presented in this research focuses on learning interventions. It uses nine publicly accessible datasets to understand and mitigate factors contributing to learning loss and the practical learning recovery measures in Texas public school districts after the recent school closures. The data came from the Census Bureau 2010, USAFACTS, Texas Department of State Health Services (DSHS), the National Center for Education Statistics (CCD), the US Bureau of Labor Statistics (LAUS), and three sources from the Texas Education Agency (STAAR, TEA, ADA, ESSER). We demonstrate a novel data-driven approach to discover insights from an extensive collection of heterogeneous public data sources. For the pandemic school closure period, the mode of instruction and prior score emerged as the primary resilience factors in the learning recovery intervention method. Grade level and census community income level are the most influential factors in predicting learning loss for both Math and Reading. We demonstrate that data-driven unbiased data analysis at a larger scale can offer policymakers an actionable understanding of how to identify learning-loss tendencies and prevent them in public schools.

Keywords

• Noisy tabular data
• data in the wild
• gradient boosting
• feature selection
• dimensionality reduction