Leveraging local network communities to predict academic performance

TL;DR

This paper introduces a novel approach to predicting academic performance by modeling success as a contagion within local learning communities, outperforming traditional centrality-based methods.

Contribution

It proposes new predictors based on learning communities that better capture peer influence, significantly improving prediction accuracy over existing centrality models.

Findings

Learning community predictors improve model accuracy

Model is 68 times more likely to be correct than centrality models

Contagion-based approach outperforms traditional methods

Abstract

For more than 20 years, social network analysis of student collaboration networks has focused on a student's centrality to predict academic performance. And even though a growing amount of sociological literature has supported that academic success is contagious, identifying central students in the network alone does not capture how peer interactions facilitate the spread of academic success throughout the network. Consequently, we propose novel predictors that treat academic success as a contagion by identifying a student's learning community, consisting of the peers that are most likely to influence a student's performance in a course. We evaluate the importance of these learning communities by predicting academic outcomes in an introductory college statistics course with 103 students. In particular, we observe that by including these learning community predictors, the resulting model is 68 times more likely to be the correct model than the current state-of-the-art centrality network models in the literature.